diff --git a/.claude/CLAUDE.md b/.claude/CLAUDE.md
new file mode 100644
index 000000000..661eea3f8
--- /dev/null
+++ b/.claude/CLAUDE.md
@@ -0,0 +1,127 @@
+# Truss Examples — Project Rules
+
+This repo contains production-ready inference examples for Truss. Follow these rules when creating, editing, or reviewing examples.
+
+## Repository structure
+
+Top-level directories:
+
+- `tutorials/` — Getting-started guides (BERT, LLMs, streaming, image generation, caching, batching)
+- `llm/` — Text generation models, organized by family (`llm/llama/`, `llm/qwen/`, `llm/mistral/`, etc.)
+- `embeddings/` — Embedding and reranker models (`bei/`, `tei/`, `clip/`)
+- `image/` — Image generation, editing, segmentation (`stable-diffusion/`, `flux/`, `segment-anything/`)
+- `audio/` — Speech-to-text, TTS, music generation (`whisper/`, `kokoro/`, `musicgen-large/`)
+- `optimized/` — Autogenerated TRT-LLM configs (`briton/`, `bisv2/`)
+- `infrastructure/` — Patterns and techniques (custom servers, gRPC, model caching, chains)
+- `_internal/` — Build tooling, templates, test scripts (not customer-facing)
+- `_archive/` — Deprecated examples
+
+## Placement rules
+
+| Model type | Directory | Example |
+|---|---|---|
+| Text generation / chat / LLM | `llm/<family>/` | `llm/llama/tinyllama-1.1B-chat-v1.0` |
+| Embedding or reranker | `embeddings/bei/`, `embeddings/tei/`, or `embeddings/clip/` | `embeddings/bei/baai-bge-large-en-v1.5-embedding` |
+| Image generation or processing | `image/` or `image/<family>/` | `image/stable-diffusion/stable-diffusion-xl-1.0` |
+| Audio (STT, TTS, music) | `audio/` | `audio/whisper/faster-whisper-v3` |
+| Infrastructure pattern | `infrastructure/` | `infrastructure/model-cache` |
+| Getting-started tutorial | `tutorials/` | `tutorials/getting-started-bert` |
+
+## Naming conventions
+
+- Use hyphens, not underscores: `falcon-7b` not `falcon_7b`
+- Do not include "truss" in folder names: `falcon-7b` not `falcon-7b-truss`
+- Include parameter count when multiple variants exist: `Falcon 7B` not `Falcon`
+- All directory names should be lowercase
+
+## config.yaml requirements
+
+Always include:
+- `model_name`
+- `description` (one-sentence summary)
+- `model_metadata.example_model_input`
+
+Recommended:
+- `model_metadata.repo_id` (Hugging Face model ID)
+- `model_metadata.avatar_url` (128x128 PNG)
+- `model_metadata.cover_image_url` (452x423 PNG)
+- `model_metadata.tags`
+
+### Requirements pinning
+
+Pin versions for ALL Python requirements:
+
+```yaml
+requirements:
+- accelerate==0.20.3
+- torch==2.0.1
+- transformers==4.30.2
+```
+
+Never leave a requirement unpinned. For git+ dependencies, pin to a specific commit hash, not `@master` or `@main`.
+
+### Secrets
+
+If the model requires a HuggingFace token, always name the secret `hf_access_token`:
+
+```yaml
+secrets:
+  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
+```
+
+### Hardware
+
+Configure with the least expensive hardware that runs at reasonable speed and quality. Note tradeoffs in the README when applicable.
+
+## README template
+
+Every example must include a `README.md` following this structure:
+
+```markdown
+# <Model Name>
+
+<One or two sentence description.>
+
+## Deploying <Model Name>
+
+<Deploy instructions with `truss push <path>`.>
+
+## Invoking <Model Name>
+
+<Curl or Python snippet with example input and expected output.>
+```
+
+- Deploy path must match the actual directory path relative to repo root
+- OpenAI-compatible models must show `/v1/chat/completions` endpoint
+- Never use `--trusted` or `--publish` flags
+- If the config requires `hf_access_token`, the README must mention setting up the secret
+
+Reference example: `image/stable-diffusion/stable-diffusion-xl-1.0`
+
+## Model I/O conventions
+
+- Models that support streaming should accept a `stream` kwarg defaulting to false
+- Models that take text input should call the parameter `prompt`
+
+## CI
+
+CI auto-discovers examples by finding directories with `config.yaml`, skipping `_archive/` and `_internal/`. To exclude a specific example from CI, add its path to `ci_excludes.yaml` at the repo root. The test suite (`_internal/bin/test_all.py`) validates all configs, READMEs, naming, links, and pinning.
+
+Run tests locally:
+```bash
+python _internal/bin/test_all.py           # default
+python _internal/bin/test_all.py --verbose  # show every check
+python _internal/bin/test_all.py --category llm  # filter by category
+```
+
+## Automatic documentation
+
+To include an example in auto-generated docs on https://truss.baseten.co/, add a `doc.yaml`:
+
+```yaml
+title: "Text-to-image"
+description: "Building a text-to-image model with SDXL"
+files:
+  - model/model.py
+  - config.yaml
+```
diff --git a/.droid.yaml b/.droid.yaml
deleted file mode 100644
index ab24dfd50..000000000
--- a/.droid.yaml
+++ /dev/null
@@ -1,62 +0,0 @@
-code:
-  plan_guidelines:
-    - These models are meant to be quick start examples. We are always generating a new example. Always include all necessary files.
-    - Never modify the README at the root of the repo.
-    - Always create the model first.
-    - Always create the requirements.txt file before the config. Point to the requirements.txt file in the config.
-    - Instead of putting the python requirements in the `config.yaml` file, we recommend putting them in a `requirements.txt` file. Then in the config.yaml file, you can use the `requirements_file` field to point to the `requirements.txt` file. This will make it easier to manage the dependencies and make it easier for the users to install the dependencies.
-    - The config creation step will have full access to the documentation for the config file. We will determine which specific keys to use at that point
-    - Model caching should be done in the config file. You do not need to handle this in the model.
-    - Always create the README for the model last. The README should always be markdown.
-  coding_tips:
-    - path : "**/config.yaml"
-      instructions: |
-        - The reference docs for the config file are at: https://truss.baseten.co/reference/config - DO NOT use values that are not in the reference docs.
-        - Add a model_metadata section to the config. Inside the model_metadata include a value for example_model_input
-          - example_model_input should be a JSON string that can be passed to the model's predict function to test its functionality.
-          - If a base64 image is required simply leave the string as 'image'. For example you might do "{'image': 'image'}" as an example input to a model that takes a single image as input.
-          - We will be using this example input to test the model. Ensure that the JSON is well formmatted and conforms to the expected inputs of the predict() function in model.py.
-        - make sure to include ALL dependencies.
-        - Always use py311 python_version.
-        - Default to using an A10G for the GPU when relevant. This should go under the resources section.
-        - Look at the model.py file closely and ensure that all the necessary information is included. For example if the model requires a GPU specify that a GPU is required in the config file.
-        - Any caching of models that's required should happen here in the config.
-    - path: "**/README.md"
-      instructions: |
-        - The readme should always credit the creators of the model and the dataset. It should also include a link to the original source of the model and the dataset.
-        - The explanation for how to run the model should use Baseten's platform as an example. Look at the existing READMEs in the repo.
-        - The README should be markdown formatted.
-        - The structure of the READMEs should be as follows. Do not deviate from this format:
-          - <model_name> Truss
-            - A short introduction to the model and its capabilities. Include any acknowledgements here.
-          - Deploying <model_name> on Baseten
-            - Instructions for cloning the repo (git clone https://github.com/basetenlabs/truss-examples/).
-            - And deploying the model on Baseten using `truss push`.
-          - Input
-            - The format of the input. All inputs are json. Specify if the model expects a base64 encoded image or a string, or a dictionary with multiple value.
-            - Explain what each input does.
-          - Output
-            - The format of the output. All outputs are json. Specify if the model returns a base64 encoded image or a string, or a dictionary with multiple value.
-          - Example usage
-            - Instructions for using the model on Baseten using `truss predict`.
-            - Be specific about input parameters and the expected output.
-            - e.g. `truss predict --input '{"prompt": "A photo of a cat"}'`
-            - Make it relevant to the model.
-            - The truss will never be executed locally. The README should not include instructions for running the model locally.
-    - path: "**/requirements.txt"
-      instructions: |
-        - Always pin the dependencies in the requirements using >=.
-        - Some dependencies may require you to install directly from a git repository. In that case, you can specify it like 'git+<repo_url>.git'
-    - path: "**/model.py"
-      instructions: |
-         - These models are being created as examples for the Truss platform. They should be runnable without additional configuration. That means you should select reasonable default values and not leave decisions up to the user.
-          - For image models that use a pretrained base model always use SG161222/RealVisXL_V3.0
-          - If there is an example of how to use the model in an .ipnyb file, use it. Use as much of it as possible. For the load() method, et up the pipelines as they have done so in the notebooks. Implement predict() to run prediction as they have done in the notebooks.
-          - All models must implement a load() function and a predict() function. The load() function should load the model and the predict() function should take in a dict of the input parameters and return the output. For image inputs and outputs use base64 strings.
-          - Images passed as inputs will be base64 encoded strings. Make sure to properly process these inputs into a format that's usable by the model.
-          - The output of the predict function will be converted to JSON and sent over the internet. Make sure that the output is JSON serializable. For images convert them to base64 strings.
-          - The model class must be called Model.
-          - Do not make up functions. Only use functions that you know exist or you can see being used in example code.
-          - Ensure all your imports are correct. Use the examples provided to ensure you're importing things correctly.
-          - When creating prompt templates, ensure there is enough detail filled in for a user to understand the model's capabilities and how to use it. If there are any specific requirements for the prompt, ensure they are included in the prompt.
-          - Try to always use a GPU if the model supports it.
diff --git a/.github/actions/setup-python/action.yml b/.github/actions/setup-python/action.yml
deleted file mode 100644
index e02b0b6ca..000000000
--- a/.github/actions/setup-python/action.yml
+++ /dev/null
@@ -1,34 +0,0 @@
-name: "Setup Python"
-description: "setups python, poetry and associated cache"
-
-
-runs:
-  using: "composite"
-  steps:
-    - uses: actions/setup-python@v5
-      with:
-        python-version: '3.11.4'
-
-    - name: Get full Python version
-      id: full-python-version
-      shell: bash
-      run: echo ::set-output name=version::$(python -c "import sys; print('-'.join(str(v) for v in sys.version_info))")
-
-    - run: curl -sSL https://install.python-poetry.org | python3 -
-      shell: bash
-
-    - name: Configure poetry
-      run: poetry config virtualenvs.in-project true
-      shell: bash
-
-    - name: Set up cache
-      uses: actions/cache@v4
-      id: cache
-      with:
-        path: .venv
-        key: venv-${{ runner.os }}-${{ steps.full-python-version.outputs.version }}-${{ hashFiles('**/poetry.lock') }}
-
-    - name: Ensure cache is healthy
-      if: steps.cache.outputs.cache-hit == 'true'
-      run: timeout 10s poetry run pip --version || rm -rf .venv
-      shell: bash
diff --git a/.github/workflows/pr.yml b/.github/workflows/pr.yml
index 85e832de1..a7189709b 100644
--- a/.github/workflows/pr.yml
+++ b/.github/workflows/pr.yml
@@ -12,6 +12,8 @@ jobs:
     runs-on: ubuntu-latest
     steps:
       - uses: actions/checkout@v4
-      - uses: ./.github/actions/setup-python/
-      - run: poetry install
-      - run: poetry run pre-commit run --all-files
+      - uses: actions/setup-python@v5
+        with:
+          python-version: '3.11.4'
+      - run: pip install truss pre-commit
+      - run: pre-commit run --all-files
diff --git a/.github/workflows/test-examples.yml b/.github/workflows/test-examples.yml
index a277d58af..e223057ee 100644
--- a/.github/workflows/test-examples.yml
+++ b/.github/workflows/test-examples.yml
@@ -6,25 +6,51 @@ on:
     - cron: '0 6 * * *'
   workflow_dispatch:
 
+permissions:
+  contents: read
+
 jobs:
+  validate:
+    runs-on: ubuntu-latest
+    timeout-minutes: 15
+    steps:
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+      - name: Set up Python environment
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
+        with:
+          python-version: '3.11.4'
+          cache: 'pip'
+      - name: Install dependencies
+        run: |
+          python -m pip install --upgrade pip
+          pip install git+https://github.com/basetenlabs/truss.git pyyaml --upgrade
+      - name: Run local validation (test_all.py)
+        run: |
+          python _internal/bin/test_all.py
+
   generate_tests:
     runs-on: ubuntu-latest
+    timeout-minutes: 5
     outputs:
       tests: ${{ steps.generate-matrix.outputs.tests }}
 
     steps:
-      - uses: actions/checkout@v4
-      - name: yq - portable yaml processor
-        uses: mikefarah/yq@v4.35.2
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
+      - uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
+        with:
+          python-version: '3.11.4'
+      - run: pip install pyyaml
       - name: Generate all tests that need to be run
         id: generate-matrix
         run: |
-          TESTS=$(cat ci.yaml | yq e ".tests" | yq eval -o=json | jq -c .)
+          TESTS=$(python _internal/bin/discover_examples.py)
           echo "tests=$TESTS" >> $GITHUB_OUTPUT
 
   ci:
     runs-on: ubuntu-latest
+    timeout-minutes: 30
     needs:
+      - validate
       - generate_tests
     strategy:
       fail-fast: false
@@ -32,28 +58,29 @@ jobs:
       matrix:
         test: ${{ fromJSON(needs.generate_tests.outputs.tests) }}
     steps:
-      - uses: actions/checkout@v4
+      - uses: actions/checkout@34e114876b0b11c390a56381ad16ebd13914f8d5 # v4
       - name: Set up Python environment
-        uses: actions/setup-python@v5
+        uses: actions/setup-python@a26af69be951a213d495a4c3e4e4022e16d87065 # v5
         with:
           python-version: '3.11.4'
-      - name: Install dependencies (if any)
+          cache: 'pip'
+      - name: Install dependencies
         run: |
           python -m pip install --upgrade pip
           pip install git+https://github.com/basetenlabs/truss.git requests tenacity --upgrade
-      - name: Test => (${{ matrix.test }}
+      - name: Test ${{ matrix.test }}
         run: |
-          python ./bin/test_example.py ${{secrets.BASETEN_API_KEY}} ${{matrix.test}}
+          python _internal/bin/test_example.py ${{ secrets.BASETEN_API_KEY }} ${{ matrix.test }}
+
   report_to_slack:
     runs-on: ubuntu-latest
+    timeout-minutes: 5
     if: always() && github.ref == 'refs/heads/main'
     needs:
       - ci
     steps:
-      - name: get-branch
-        run: echo ${{ github.ref }}
-      - name: show-slack-status
-        uses: 8398a7/action-slack@v3
+      - name: Report CI result
+        uses: 8398a7/action-slack@77eaa4f1c608a7d68b38af4e3f739dcd8cba273e # v3.19.0
         with:
           status: custom
           fields: author, job, commit, repo
diff --git a/.github/workflows/truss_deploy.yml b/.github/workflows/truss_deploy.yml
index 188871205..0556bb3b9 100644
--- a/.github/workflows/truss_deploy.yml
+++ b/.github/workflows/truss_deploy.yml
@@ -31,4 +31,4 @@ jobs:
             - name: Run tests
               env:
                 BASETEN_API_KEY: ${{ secrets.BASETEN_API_KEY }}
-              run: python bin/test_truss_deploy.py
+              run: python _internal/bin/test_truss_deploy.py
diff --git a/.github/workflows/warm-chains.yml b/.github/workflows/warm-chains.yml
index 4333850ec..ebe09a62f 100644
--- a/.github/workflows/warm-chains.yml
+++ b/.github/workflows/warm-chains.yml
@@ -12,7 +12,7 @@ jobs:
     steps:
       - uses: actions/checkout@v4
       - name: Set up Python environment
-        uses: actions/setup-python@v2
+        uses: actions/setup-python@v5
         with:
           python-version: '3.11.4'
       - name: Install dependencies (if any)
@@ -29,5 +29,5 @@ jobs:
           EOF
       - name: Warm up chains
         run: |
-          truss chains deploy chains-examples/docs/poems/poems.py
-          truss chains deploy chains-examples/docs/audio-transcription/whisper_chainlet.py
+          truss chains deploy infrastructure/chains-examples/docs/poems/poems.py
+          truss chains deploy infrastructure/chains-examples/docs/audio-transcription/whisper_chainlet.py
diff --git a/.gitignore b/.gitignore
index 3babb0120..bd7d47898 100644
--- a/.gitignore
+++ b/.gitignore
@@ -1,163 +1,12 @@
 .DS_Store
-**/.DS_Store
-
-# Byte-compiled / optimized / DLL files
 __pycache__/
 *.py[cod]
-*$py.class
-
-# C extensions
 *.so
-
-# Distribution / packaging
-.Python
-build/
-develop-eggs/
-dist/
-downloads/
-eggs/
-.eggs/
-lib/
-lib64/
-parts/
-sdist/
-var/
-wheels/
-share/python-wheels/
 *.egg-info/
-.installed.cfg
 *.egg
-MANIFEST
-
-# PyInstaller
-#  Usually these files are written by a python script from a template
-#  before PyInstaller builds the exe, so as to inject date/other infos into it.
-*.manifest
-*.spec
-
-# Installer logs
-pip-log.txt
-pip-delete-this-directory.txt
-
-# Unit test / coverage reports
-htmlcov/
-.tox/
-.nox/
-.coverage
-.coverage.*
-.cache
-nosetests.xml
-coverage.xml
-*.cover
-*.py,cover
-.hypothesis/
-.pytest_cache/
-cover/
-
-# Translations
-*.mo
-*.pot
-
-# Django stuff:
-*.log
-local_settings.py
-db.sqlite3
-db.sqlite3-journal
-
-# Flask stuff:
-instance/
-.webassets-cache
-
-# Scrapy stuff:
-.scrapy
-
-# Sphinx documentation
-docs/_build/
-
-# PyBuilder
-.pybuilder/
-target/
-
-# Jupyter Notebook
-.ipynb_checkpoints
-
-# IPython
-profile_default/
-ipython_config.py
-
-# pyenv
-#   For a library or package, you might want to ignore these files since the code is
-#   intended to run in multiple environments; otherwise, check them in:
-# .python-version
-
-# pipenv
-#   According to pypa/pipenv#598, it is recommended to include Pipfile.lock in version control.
-#   However, in case of collaboration, if having platform-specific dependencies or dependencies
-#   having no cross-platform support, pipenv may install dependencies that don't work, or not
-#   install all needed dependencies.
-#Pipfile.lock
-
-# poetry
-#   Similar to Pipfile.lock, it is generally recommended to include poetry.lock in version control.
-#   This is especially recommended for binary packages to ensure reproducibility, and is more
-#   commonly ignored for libraries.
-#   https://python-poetry.org/docs/basic-usage/#commit-your-poetrylock-file-to-version-control
-#poetry.lock
-
-# pdm
-#   Similar to Pipfile.lock, it is generally recommended to include pdm.lock in version control.
-#pdm.lock
-#   pdm stores project-wide configurations in .pdm.toml, but it is recommended to not include it
-#   in version control.
-#   https://pdm.fming.dev/#use-with-ide
-.pdm.toml
-
-# PEP 582; used by e.g. github.com/David-OConnor/pyflow and github.com/pdm-project/pdm
-__pypackages__/
-
-# Celery stuff
-celerybeat-schedule
-celerybeat.pid
-
-# SageMath parsed files
-*.sage.py
-
-# Environments
 .env
 .venv
-env/
-venv/
-ENV/
-env.bak/
-venv.bak/
-
-# Spyder project settings
-.spyderproject
-.spyproject
-
-# Rope project settings
-.ropeproject
-
-# mkdocs documentation
-/site
-
-# mypy
 .mypy_cache/
-.dmypy.json
-dmypy.json
-
-# Pyre type checker
-.pyre/
-
-# pytype static type analyzer
-.pytype/
-
-# Cython debug symbols
-cython_debug/
-
-# PyCharm
-#  JetBrains specific template is maintained in a separate JetBrains.gitignore that can
-#  be found at https://github.com/github/gitignore/blob/main/Global/JetBrains.gitignore
-#  and can be added to the global gitignore or merged into this file.  For a more nuclear
-#  option (not recommended) you can uncomment the following to ignore the entire idea folder.
-#.idea/
+.pytest_cache/
+.ruff_cache/
+validation_report.json
diff --git a/.isort.cfg b/.isort.cfg
deleted file mode 100644
index b9fb3f3e8..000000000
--- a/.isort.cfg
+++ /dev/null
@@ -1,2 +0,0 @@
-[settings]
-profile=black
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index fa6cd8b17..709e1f468 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -13,24 +13,21 @@ repos:
       - id: check-merge-conflict
       - id: check-symlinks
       - id: debug-statements
-        exclude: ^sana/
+        exclude: ^image/sana/
 
   - repo: https://github.com/astral-sh/ruff-pre-commit
-    # Ruff version.
     rev: v0.11.6
     hooks:
-      # Run the linter.
       - id: ruff
         args: [--fix, --fix-only]
-      # Run the formatter.
       - id: ruff-format
         types_or: [ python, pyi ]
 
   - repo: local
     hooks:
       - id: validate-ci
-        name: validate-ci-local
-        entry: poetry run python ./bin/validate_ci.py
+        name: validate-ci
+        entry: python _internal/bin/test_all.py
         language: python
-        additional_dependencies: ['truss', 'pyyaml']
+        additional_dependencies: [truss, pyyaml]
         pass_filenames: false
diff --git a/.tool-versions b/.tool-versions
deleted file mode 100644
index 0fb6db853..000000000
--- a/.tool-versions
+++ /dev/null
@@ -1,2 +0,0 @@
-python 3.11.9
-poetry 2.1.3
diff --git a/01-getting-started-bert/config.yaml b/01-getting-started-bert/config.yaml
deleted file mode 100644
index e833ba318..000000000
--- a/01-getting-started-bert/config.yaml
+++ /dev/null
@@ -1,60 +0,0 @@
-# # Step 2: Writing the config.yaml
-#
-# Each Truss has a config.yaml file where we can configure
-# options related to the deployment. It's in this file where
-# we can define requirements, resources, and runtime options like
-# secrets and environment variables
-#
-# ### Basic Options
-#
-# In this section, we can define basic metadata about the model,
-# such as the name, and the Python version to build with.
-model_name: bert
-python_version: py310
-model_metadata:
-  example_model_input: { "text": "Hello my name is {MASK}" }
-
-
-# ### Set up python requirements
-#
-# In this section, we define any pip requirements that
-# we need to run the model. To run this, we need PyTorch
-# and Tranformers.
-requirements:
-  - torch==2.0.1
-  - transformers==4.33.2
-  - numpy==1.26.4
-
-# ### Configure the resources needed
-#
-# In this section, we can configure resources
-# needed to deploy this model. Here, we have no need for a GPU
-# so we leave the accelerator section blank.
-resources:
-  accelerator: null
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: false
-
-# ### Other config options
-#
-# Truss also has provisions for adding other runtime options
-# packages. In this example, we don't need these, so we leave
-# this empty for now.
-secrets: {}
-system_packages: []
-environment_variables: {}
-external_package_dirs: []
-
-# # Step 3: Deploying & running inference
-#
-# Deploy the model with the following command:
-#
-# ```bash
-# $ truss push
-# ```
-#
-# And then you can performance inference with:
-# ```
-# $ truss predict -d '"Truss is awesome!"'
-# ```
diff --git a/02-llm/config.yaml b/02-llm/config.yaml
deleted file mode 100644
index 3e03fb42c..000000000
--- a/02-llm/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-# # Setting up the config.yaml
-#
-# Running Mistral 7B requires a few libraries, such as
-# `torch`, `transformers` and a couple others.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input: {"prompt": "What is the meaning of life?"}
-model_name: Mistral 7B
-python_version: py311
-requirements:
-- transformers==4.42.3
-- sentencepiece==0.1.99
-- accelerate==0.23.0
-- torch==2.0.1
-- numpy==1.26.4
-# ## Configure resources for Mistral
-#
-# Note that we need an A10G to run this model.
-resources:
-  accelerator: A10G
-  use_gpu: true
-secrets:
-  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
-system_packages: []
-# # Deploy the model
-#
-# Deploy the model like you would other Trusses, with:
-# ```bash
-# $ truss push
-# ```
-# You can then invoke the model with:
-# ```bash
-# $ truss predict -d '{"inputs": "What is a large language model?"}'
-# ```
diff --git a/03-llm-with-streaming/config.yaml b/03-llm-with-streaming/config.yaml
deleted file mode 100644
index a6ca9504b..000000000
--- a/03-llm-with-streaming/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# # Setting up the config.yaml
-#
-# Running Falcon 7B requires torch, transformers,
-# and a few other related libraries.
-model_name: "LLM with Streaming"
-model_metadata:
-    example_model_input: {"prompt": "what is the meaning of life"}
-requirements:
-- torch==2.0.1
-- peft==0.4.0
-- scipy==1.11.1
-- sentencepiece==0.1.99
-- accelerate==0.21.0
-- bitsandbytes==0.41.1
-- einops==0.6.1
-- transformers==4.31.0
-- numpy==1.26.4
-# ## Configure resources for Falcon
-#
-# Note that we need an A10G to run this model.
-resources:
-  cpu: "3"
-  memory: 14Gi
-  use_gpu: true
-  accelerator: A10G
diff --git a/04-image-generation/config.yaml b/04-image-generation/config.yaml
deleted file mode 100644
index 3472543ce..000000000
--- a/04-image-generation/config.yaml
+++ /dev/null
@@ -1,79 +0,0 @@
-# # Setting up the config yaml
-#
-# Running SDXL requires a handful of Python libraries, including
-# diffusers, transformers, and others.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    { "prompt": "A tree in a field under the night sky", "use_refiner": true }
-model_name: Stable Diffusion XL
-python_version: py39
-requirements:
-  - transformers==4.34.0
-  - accelerate==0.23.0
-  - safetensors==0.4.0
-  - git+https://github.com/basetenlabs/diffusers.git@9a353290b1497023d4745a719ec02c50f680499a
-  - invisible-watermark>=0.2.0
-  - xformers==0.0.22
-  - numpy==1.26.4
-# ## Configuring resources for SDXL 1.0
-#
-# Note that we need an A10G to run this model.
-resources:
-  accelerator: A10G
-  cpu: 3500m
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-# ## System Packages
-#
-# Running diffusers requires `ffmpeg` and a couple other system
-# packages.
-system_packages:
-  - ffmpeg
-  - libsm6
-  - libxext6
-# ## Enabling Caching
-#
-# SDXL is a very large model, and downloading it could take up to 10 minutes. This means
-# that the cold start time for this model is long. We can solve that by using our build
-# caching feature. This moves the model download to the build stage of your model--
-# caching the model will take about 10 minutes initially but you will get ~9s cold starts
-# subsequently.
-#
-# To enable caching, add the following to the config:
-# ```yaml
-# model_cache:
-#   - repo_id: madebyollin/sdxl-vae-fp16-fix
-#     revision: main
-#     allow_patterns:
-#       - config.json
-#       - diffusion_pytorch_model.safetensors
-#     use_volume: true
-#     volumne_folder: "sdxl-vae-fp16-fix"
-#   - repo_id: stabilityai/stable-diffusion-xl-base-1.0
-#     allow_patterns:
-#       - "*.json"
-#       - "*.fp16.safetensors"
-#       - sd_xl_base_1.0.safetensors
-#     use_volume: true
-#     volumne_folder: "sdxl-base"
-#   - repo_id: stabilityai/stable-diffusion-xl-refiner-1.0
-#     allow_patterns:
-#       - "*.json"
-#       - "*.fp16.safetensors"
-#       - sd_xl_refiner_1.0.safetensors
-#     use_volume: true
-#     volumne_folder: "sdxl-refiner"
-# ```
-# # Deploy the model
-#
-# Deploy the model like you would other Trusses, with:
-# ```bash
-# $ truss push
-# ```
-# You can then invoke the model with:
-# ```bash
-# $ truss predict -d '{"prompt": "A tree in a field under the night sky", "use_refiner": true}'
-# ```
diff --git a/05-speech-to-text/config.yaml b/05-speech-to-text/config.yaml
deleted file mode 100644
index 56ab01839..000000000
--- a/05-speech-to-text/config.yaml
+++ /dev/null
@@ -1,20 +0,0 @@
-environment_variables: {}
-model_metadata:
-  example_model_input: {"url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}
-model_name: Whisper
-python_version: py39
-requirements:
-- openai-whisper==20250625
-- torch==2.0.1
-- numpy==1.26.4
-resources:
-  cpu: "4"
-  memory: 16Gi
-  use_gpu: true
-  accelerator: A10G
-secrets: {}
-system_packages:
-- ffmpeg
-external_data:
-  - url: https://baseten-public.s3.us-west-2.amazonaws.com/models/whisper/small.pt
-    local_data_path: models/small.pt
diff --git a/06-high-performance-cached-weights/config.yaml b/06-high-performance-cached-weights/config.yaml
deleted file mode 100644
index 4748b9b55..000000000
--- a/06-high-performance-cached-weights/config.yaml
+++ /dev/null
@@ -1,58 +0,0 @@
-# # Setting up the config.yaml
-#
-# The `config.yaml` file is where you need to include the changes to
-# actually cache the weights at build time.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input: { "prompt": "What is the meaning of life?" }
-model_name: Llama with Cached Weights
-python_version: py39
-requirements:
-  - accelerate==0.21.0
-  - safetensors==0.3.2
-  - torch==2.0.1
-  - transformers==4.34.0
-  - sentencepiece==0.1.99
-  - numpy==1.26.4
-# # Configuring the model_cache
-#
-# To cache model weights, set the `model_cache` key.
-# The `repo_id` field allows you to specify a Huggingface
-# repo to pull down and cache at build-time, and the `ignore_patterns`
-# field allows you to specify files to ignore. If this is specified, then
-# this repo won't have to be pulled during runtime.
-#
-# Check out the [guide](https://truss.baseten.co/guides/model-cache) for more info.
-model_cache:
-  - repo_id: "NousResearch/Llama-2-7b-chat-hf"
-    revision: main
-    ignore_patterns:
-      - "*.bin"
-    use_volume: true
-    volume_folder: "llama-2-7b-chat-hf"
-
-# The remaining config options are again, similar to what you would
-# configure for the model without the weight caching.
-resources:
-  cpu: "4"
-  memory: 30Gi
-  use_gpu: True
-  accelerator: A10G
-secrets: {}
-# # Deploy the model
-#
-# Deploy the model like you would other Trusses, with:
-# ```bash
-# $ truss push
-# ```
-#  <Note>
-#  The build step will take  longer than with the normal
-#  Llama Truss, since bundling the model weights is now happening during the build.
-#  The deploy step & scale-ups will happen much faster with this approach.
-#  </Note>
-#
-# You can then invoke the model with:
-# ```bash
-# $ truss predict -d '{"inputs": "What is a large language model?"}'
-# ```
diff --git a/07-high-performance-dynamic-batching/README.md b/07-high-performance-dynamic-batching/README.md
deleted file mode 100644
index 88b45a4e8..000000000
--- a/07-high-performance-dynamic-batching/README.md
+++ /dev/null
@@ -1,29 +0,0 @@
-# Dynamic Batching in Truss
-
-This repository contains an implementation designed to enable dynamic batching for machine learning models within the Truss framework. The core of this implementation lies in the `model/model.py` file, which introduces a `MlBatcher` class extending `AsyncBatcher`. This class is responsible for collecting individual prediction requests and processing them in batches, thereby improving throughput and efficiency.
-
-## Key Features
-
-- **Dynamic Batching:** The `MlBatcher` class dynamically batches incoming prediction requests, allowing for more efficient use of resources and faster response times.
-- **Asynchronous Processing:** Utilizes asynchronous programming to handle concurrent prediction requests without blocking, ensuring high throughput.
-- **Easy Integration:** Designed to be deployed as a normal Truss, making integration into existing projects straightforward.
-
-## Deployment
-
-To deploy this as a normal Truss, ensure you have the Truss CLI installed and configured. Then, follow these steps:
-
-1. Clone this repository to your local machine.
-2. Navigate to the repository directory and build the Truss using the command `truss push --publish`.
-3. Once the build completes, deploy the Truss to your desired environment.
-
-## Configuration
-
-The `config.yaml` file contains configuration options for the model, including the Python version, required packages, and runtime settings such as `predict_concurrency`. Adjust these settings as needed to optimize performance for your specific use case.
-
-## Testing
-
-The `test.py` file provides an example of how to send concurrent requests to the deployed model for testing purposes. Modify the URL and data as needed to match your deployment.
-
-## Conclusion
-
-This implementation showcases how dynamic batching can be seamlessly integrated into the Truss framework, providing a scalable and efficient solution for handling machine learning inference at scale.
diff --git a/07-high-performance-dynamic-batching/config.yaml b/07-high-performance-dynamic-batching/config.yaml
deleted file mode 100644
index fae83526a..000000000
--- a/07-high-performance-dynamic-batching/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-base_image:
-  image: baseten/trtllm-server:r23.12_baseten_v0.9.0.dev2024022000
-  python_executable_path: /usr/bin/python3
-model_name: TRT Whisper - Dynamic Batching
-python_version: py311
-requirements:
-  - async-batcher==0.2.0
-  - mpi4py==3.1.5
-  - pynvml==11.5.0
-  - huggingface_hub==0.20.3
-  - tiktoken==0.6.0
-  - datasets==2.17.1
-  - kaldialign==0.9
-  - openai-whisper==20250625
-  - soundfile==0.12.1
-model_cache:
-  - repo_id: baseten/trtllm-whisper-a10g-large-v2-1
-    revision: main
-    use_volume: true
-    volume_folder: trtllm-whisper-a10g-large-v2-1
-system_packages:
-  - python3.10-venv
-  - ffmpeg
-resources:
-  accelerator: A10G
-runtime:
-  predict_concurrency: 256
-external_data:
-  - local_data_path: assets/multilingual.tiktoken
-    url: https://raw.githubusercontent.com/openai/whisper/main/whisper/assets/multilingual.tiktoken
-  - local_data_path: assets/mel_filters.npz
-    url: https://raw.githubusercontent.com/openai/whisper/main/whisper/assets/mel_filters.npz
diff --git a/09-private-huggingface/config.yaml b/09-private-huggingface/config.yaml
deleted file mode 100644
index 8581f513e..000000000
--- a/09-private-huggingface/config.yaml
+++ /dev/null
@@ -1,36 +0,0 @@
-# # Setting up the config.yaml
-#
-# The main things that need to be set up in the config are
-# `requirements`, which need to include Hugging Face transformers,
-# and the secrets.
-environment_variables: {}
-model_name: private-model
-python_version: py39
-requirements:
-- torch==2.0.1
-- transformers==4.30.2
-resources:
-  cpu: "1"
-  memory: 2Gi
-  use_gpu: false
-  accelerator: null
-# To make the `hf_access_token` available in the Truss, we need to include
-# it in the config. Setting the value to `null` here means that the value
-# will be set by the Baseten secrets manager.
-secrets:
-  hf_access_token: null
-system_packages: []
-# # Deploying the model
-#
-# An important note for deploying models with secrets is that
-# you must use the `--trusted` flag to give the model access to
-# secrets stored on the remote secrets manager.
-#
-# ```bash
-# $ truss push --trusted
-# ```
-#
-# After the model finishes deploying, you can invoke it with:
-# ```bash
-# $ truss predict -d '"It is a [MASK] world"'
-# ```
diff --git a/10-using-system-packages/config.yaml b/10-using-system-packages/config.yaml
deleted file mode 100644
index a7bc9cfb4..000000000
--- a/10-using-system-packages/config.yaml
+++ /dev/null
@@ -1,36 +0,0 @@
-# # Setting up the config.yaml file
-#
-# The main items that need to be configured in the config.yaml file are requirements
-# and `system_packages` sections.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-    example_model_input: {"url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png", "prompt": "What is the invoice number?"}
-model_name: LayoutLM Document QA
-python_version: py39
-# Specify the versions of the Python requirements that are needed.
-# <Note>Always pin exact versions for your Python dependencies. The ML/AI space moves fast, so you want to have an up-to-date version of each package while also being protected from breaking changes.</Note>
-requirements:
-- Pillow==10.0.0
-- pytesseract==0.3.10
-- torch==2.0.1
-- transformers==4.30.2
-- numpy==1.26.4
-resources:
-  cpu: "4"
-  memory: 16Gi
-  use_gpu: false
-  accelerator: null
-secrets: {}
-# The system_packages section is the other important bit here, you can
-# add any package that's available via `apt` on Debian.
-system_packages:
-- tesseract-ocr
-# # Deploy the model
-# ```bash
-# $ truss push
-# ```
-# You can then invoke the model with:
-# ```
-# $ truss predict -d '{"url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png", "prompt": "What is the invoice number?"}'
-# ```
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/README.md
deleted file mode 100644
index 49d402ba2..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-modernbert-base-embedding
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-modernbert-base-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Alibaba-NLP/gte-modernbert-base  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-modernbert-base
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/config.yaml
deleted file mode 100644
index 08fc103c6..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-modernbert-base
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
deleted file mode 100644
index 00fd6923d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-Qwen2-1.5B-instruct-embedding
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-Qwen2-1.5B-instruct-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-Qwen2-1.5B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Alibaba-NLP/gte-Qwen2-1.5B-instruct  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-Qwen2-1.5B-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
deleted file mode 100644
index 06d001826..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-Qwen2-1.5B-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
deleted file mode 100644
index 9d5e46b80..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-Qwen2-7B-instruct-embedding
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-Qwen2-7B-instruct-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Alibaba-NLP/gte-Qwen2-7B-instruct  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-Qwen2-7B-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
deleted file mode 100644
index f95eb18ba..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-Qwen2-7B-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/README.md
deleted file mode 100644
index f4b4585e4..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/README.md
+++ /dev/null
@@ -1,185 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-reranker-modernbert-base
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with Alibaba-NLP/gte-reranker-modernbert-base. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-reranker-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Reranker models may have at most one label, which contains the score of the reranking.
-
-Alibaba-NLP/gte-reranker-modernbert-base  is a reranker model, used to re-rank a list of items, given a query. \nIt is frequently used in search engines, recommendation systems, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank`:
-```json
-{
-    "query": "What is Baseten?",
-    "raw_scores": true,
-    "return_text": false,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": true,
-    "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-response = client.rerank(
-    query="What is Baseten?",
-    texts=["Deep Learning is ...", "Baseten is a fast inference provider"],
-    raw_scores=True,
-    return_text=False,
-    truncate=True,
-)
-print(response.data)
-```
-
-Sometimes, you may want to apply a custom template to the texts before reranking them and call the predict endpoint instead:
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    # Custom template function to apply to the texts
-    # a popular template might be "{query}\n{document}"
-    # or also chat-style templates like "User: {query}\nDocument: {document}"
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["What is baseten? A: Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-
-### Requests python library
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank",
-    json={
-    "query": "What is Baseten?",
-    "raw_scores": True,
-    "return_text": False,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": True,
-    "truncation_direction": "Right"
-}
-```
-Returns:
-```json
-[
-  {
-    "index": 0,
-    "score": 1,
-    "text": "Deep Learning is ..."
-  }
-]
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/rerank`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI.com does not have a rerank endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-reranker-modernbert-base
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/config.yaml
deleted file mode 100644
index c0afa0b0d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: Alibaba-NLP/gte-reranker-modernbert-base
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/README.md
deleted file mode 100644
index 8fe64d72a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/README.md
+++ /dev/null
@@ -1,185 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with BAAI/bge-reranker-large
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with BAAI/bge-reranker-large. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Reranker models may have at most one label, which contains the score of the reranking.
-
-BAAI/bge-reranker-large  is a reranker model, used to re-rank a list of items, given a query. \nIt is frequently used in search engines, recommendation systems, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-baai-bge-reranker-large-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank`:
-```json
-{
-    "query": "What is Baseten?",
-    "raw_scores": true,
-    "return_text": false,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": true,
-    "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-response = client.rerank(
-    query="What is Baseten?",
-    texts=["Deep Learning is ...", "Baseten is a fast inference provider"],
-    raw_scores=True,
-    return_text=False,
-    truncate=True,
-)
-print(response.data)
-```
-
-Sometimes, you may want to apply a custom template to the texts before reranking them and call the predict endpoint instead:
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    # Custom template function to apply to the texts
-    # a popular template might be "{query}\n{document}"
-    # or also chat-style templates like "User: {query}\nDocument: {document}"
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["What is baseten? A: Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-
-### Requests python library
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank",
-    json={
-    "query": "What is Baseten?",
-    "raw_scores": True,
-    "return_text": False,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": True,
-    "truncation_direction": "Right"
-}
-```
-Returns:
-```json
-[
-  {
-    "index": 0,
-    "score": 1,
-    "text": "Deep Learning is ..."
-  }
-]
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/rerank`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI.com does not have a rerank endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-baai-bge-reranker-large-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: BAAI/bge-reranker-large
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/config.yaml
deleted file mode 100644
index 07cee6c55..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-baai-bge-reranker-large/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-baai-bge-reranker-large-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: BAAI/bge-reranker-large
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/README.md
deleted file mode 100644
index 349a6af91..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with google/embeddinggemma-300m
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with google/embeddinggemma-300m. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-google/embeddinggemma-300m  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-google-embeddinggemma-300m-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-google-embeddinggemma-300m-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: google/embeddinggemma-300m
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/config.yaml
deleted file mode 100644
index f995b6e69..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-google-embeddinggemma-300m/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-google-embeddinggemma-300m-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: google/embeddinggemma-300m
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/README.md
deleted file mode 100644
index 1a067245a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with intfloat/multilingual-e5-large-instruct
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with intfloat/multilingual-e5-large-instruct. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-intfloat/multilingual-e5-large-instruct  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-intfloat-multilingual-e5-large-instruct-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-intfloat-multilingual-e5-large-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: intfloat/multilingual-e5-large-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/config.yaml
deleted file mode 100644
index 5526ab547..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-intfloat-multilingual-e5-large-instruct/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-intfloat-multilingual-e5-large-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: intfloat/multilingual-e5-large-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/README.md
deleted file mode 100644
index 91770d5cc..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with jina-ai/jina-embeddings-v2-base-en
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with jina-ai/jina-embeddings-v2-base-en. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [jinaai/jina-embeddings-v2-base-en](https://huggingface.co/jinaai/jina-embeddings-v2-base-en).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-jinaai/jina-embeddings-v2-base-en  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-jina-ai-jina-embeddings-v2-base-en-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-jina-ai-jina-embeddings-v2-base-en-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: jinaai/jina-embeddings-v2-base-en
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/config.yaml
deleted file mode 100644
index 5d201f46c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jina-ai-jina-embeddings-v2-base-en/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-jina-ai-jina-embeddings-v2-base-en-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: jinaai/jina-embeddings-v2-base-en
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/README.md
deleted file mode 100644
index adb9d9228..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with jinaai/jina-embeddings-v2-base-code
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with jinaai/jina-embeddings-v2-base-code. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [jinaai/jina-embeddings-v2-base-code](https://huggingface.co/jinaai/jina-embeddings-v2-base-code).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-jinaai/jina-embeddings-v2-base-code  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-jinaai-jina-embeddings-v2-base-code-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-jinaai-jina-embeddings-v2-base-code-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: jinaai/jina-embeddings-v2-base-code
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/config.yaml
deleted file mode 100644
index c27ee0692..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-jinaai-jina-embeddings-v2-base-code/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-jinaai-jina-embeddings-v2-base-code-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: jinaai/jina-embeddings-v2-base-code
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
deleted file mode 100644
index 7671ee6a2..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with mixedbread-ai/mxbai-embed-large-v1-embedding
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with mixedbread-ai/mxbai-embed-large-v1-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-mixedbread-ai/mxbai-embed-large-v1  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: mixedbread-ai/mxbai-embed-large-v1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
deleted file mode 100644
index a1ac58726..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: mixedbread-ai/mxbai-embed-large-v1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased/README.md
deleted file mode 100644
index 748318cc2..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased/README.md
+++ /dev/null
@@ -1,187 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with NER/bert-base-ner-uncased
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with NER/bert-base-ner-uncased. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [baseten-admin/bert-base-ner-uncased](https://huggingface.co/baseten-admin/bert-base-ner-uncased).
-Suitable models can be identified by the `ForTokenClassification` suffix in the model name. NER models classify each token in the input text into entity categories (e.g., PER, ORG, LOC) or 'O' (outside any entity).
-
-baseten-admin/bert-base-ner-uncased  is a Named Entity Recognition (NER) model, used to identify and classify named entities in text. \nIt is frequently used for information extraction, entity linking, and document analysis. Common entities include persons, organizations, locations, dates, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-ner-bert-base-ner-uncased-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict_tokens`
-```json
-{
-  "inputs": ["Apple is looking at buying U.K. startup for $1 billion"],
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-
-response = client.batch_post(
-    route="/predict_tokens",
-    payloads=[{
-        "inputs": [["Apple is looking at buying U.K. startup for $1 billion"]],
-        "raw_scores": False,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }]
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-response = requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict_tokens",
-    json={
-        "inputs": [["Apple is looking at buying U.K. startup for $1 billion"]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-print(response.json())
-```
-Returns:
-```json
-[
-  [
-    {
-      "token": "[CLS]",
-      "token_id": 101,
-      "start": 0,
-      "end": 0,
-      "results": {
-        "O": 9.4140625,
-        "B-MISC": -1.15625,
-        "I-MISC": -0.859375,
-        "B-PER": -1.2744141,
-        "I-PER": -1.6552734,
-        "B-ORG": -0.88378906,
-        "I-ORG": -0.9345703,
-        "B-LOC": -1.2275391,
-        "I-LOC": -1.4042969
-      }
-    },
-    {
-      "token": "Apple",
-      "token_id": 6207,
-      "start": 0,
-      "end": 5,
-      "results": {
-        "B-ORG": 6.7578125,
-        "O": -1.7929688,
-        "B-LOC": 0.6015625,
-        "B-MISC": 0.2467041,
-        "B-PER": 0.17675781,
-        "I-ORG": -0.6484375,
-        "I-MISC": -1.9873047,
-        "I-LOC": -1.3808594,
-        "I-PER": -2.21875
-      }
-    }
-  ]
-]
-```
-Important, this uses the `predict_tokens` endpoint for token-level classification. The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict_tokens`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a NER endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Apple is looking at buying U.K. startup for $1 billion
-    - - John works at Google in Mountain View, California
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-ner-bert-base-ner-uncased-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: baseten-admin/bert-base-ner-uncased
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased/config.yaml
deleted file mode 100644
index af374e453..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-ner-bert-base-ner-uncased/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-      - - Apple is looking at buying U.K. startup for $1 billion
-      - - John works at Google in Mountain View, California
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-ner-bert-base-ner-uncased-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: baseten-admin/bert-base-ner-uncased
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-  version_overrides:
-    engine_builder_version: null
-    bei_bert_version: 1.8.6.ner
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/README.md
deleted file mode 100644
index 51ccd44c8..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with nomic-ai/nomic-embed-text-v1.5
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with nomic-ai/nomic-embed-text-v1.5. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nomic-ai/nomic-embed-text-v1.5](https://huggingface.co/nomic-ai/nomic-embed-text-v1.5).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-nomic-ai/nomic-embed-text-v1.5  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-nomic-ai-nomic-embed-text-v1.5-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nomic-ai-nomic-embed-text-v1.5-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nomic-ai/nomic-embed-text-v1.5
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/config.yaml
deleted file mode 100644
index de83f65ed..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v1.5/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nomic-ai-nomic-embed-text-v1.5-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nomic-ai/nomic-embed-text-v1.5
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/README.md
deleted file mode 100644
index 17f648545..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with nomic-ai/nomic-embed-text-v2-moe
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with nomic-ai/nomic-embed-text-v2-moe. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nomic-ai/nomic-embed-text-v2-moe](https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-nomic-ai/nomic-embed-text-v2-moe  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-nomic-ai-nomic-embed-text-v2-moe-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nomic-ai-nomic-embed-text-v2-moe-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nomic-ai/nomic-embed-text-v2-moe
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/config.yaml
deleted file mode 100644
index 481f63362..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nomic-ai-nomic-embed-text-v2-moe/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nomic-ai-nomic-embed-text-v2-moe-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nomic-ai/nomic-embed-text-v2-moe
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/README.md
deleted file mode 100644
index 751e5c09c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with nvidia/llama-embed-nemotron-8b
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with nvidia/llama-embed-nemotron-8b. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nvidia/llama-embed-nemotron-8b](https://huggingface.co/nvidia/llama-embed-nemotron-8b).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-nvidia/llama-embed-nemotron-8b  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-nvidia-llama-embed-nemotron-8b-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nvidia-llama-embed-nemotron-8b-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nvidia/llama-embed-nemotron-8b
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/config.yaml
deleted file mode 100644
index e54df7d78..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-embed-nemotron-8b/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nvidia-llama-embed-nemotron-8b-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nvidia/llama-embed-nemotron-8b
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/README.md
deleted file mode 100644
index de54a8b4d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with nvidia/llama-nemotron-embed-1b-v2
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with nvidia/llama-nemotron-embed-1b-v2. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nvidia/llama-nemotron-embed-1b-v2](https://huggingface.co/nvidia/llama-nemotron-embed-1b-v2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-nvidia/llama-nemotron-embed-1b-v2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-nvidia-llama-nemotron-embed-1b-v2-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nvidia-llama-nemotron-embed-1b-v2-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nvidia/llama-nemotron-embed-1b-v2
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/config.yaml
deleted file mode 100644
index 6741abc99..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-nvidia-llama-nemotron-embed-1b-v2/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-nvidia-llama-nemotron-embed-1b-v2-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: nvidia/llama-nemotron-embed-1b-v2
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/README.md
deleted file mode 100644
index fcea68db2..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with redis/langcache-embed-v2
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with redis/langcache-embed-v2. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [redis/langcache-embed-v2](https://huggingface.co/redis/langcache-embed-v2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-redis/langcache-embed-v2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-redis-langcache-embed-v2-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-redis-langcache-embed-v2-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: redis/langcache-embed-v2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/config.yaml
deleted file mode 100644
index 5055c160f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-redis-langcache-embed-v2/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-redis-langcache-embed-v2-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: redis/langcache-embed-v2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/README.md
deleted file mode 100644
index b2b9f1f5f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with sentence-transformers/all-MiniLM-L6-v2-embedding
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with sentence-transformers/all-MiniLM-L6-v2-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [sentence-transformers/all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-sentence-transformers/all-MiniLM-L6-v2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: sentence-transformers/all-MiniLM-L6-v2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
deleted file mode 100644
index bd676606e..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: sentence-transformers/all-MiniLM-L6-v2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1/README.md
deleted file mode 100644
index 45bbd4308..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1/README.md
+++ /dev/null
@@ -1,187 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with tanaos/tanaos-NER-v1
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with tanaos/tanaos-NER-v1. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [tanaos/tanaos-NER-v1](https://huggingface.co/tanaos/tanaos-NER-v1).
-Suitable models can be identified by the `ForTokenClassification` suffix in the model name. NER models classify each token in the input text into entity categories (e.g., PER, ORG, LOC) or 'O' (outside any entity).
-
-tanaos/tanaos-NER-v1  is a Named Entity Recognition (NER) model, used to identify and classify named entities in text. \nIt is frequently used for information extraction, entity linking, and document analysis. Common entities include persons, organizations, locations, dates, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-tanaos-tanaos-ner-v1-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict_tokens`
-```json
-{
-  "inputs": ["Apple is looking at buying U.K. startup for $1 billion"],
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-
-response = client.batch_post(
-    route="/predict_tokens",
-    payloads=[{
-        "inputs": [["Apple is looking at buying U.K. startup for $1 billion"]],
-        "raw_scores": False,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }]
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-response = requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict_tokens",
-    json={
-        "inputs": [["Apple is looking at buying U.K. startup for $1 billion"]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-print(response.json())
-```
-Returns:
-```json
-[
-  [
-    {
-      "token": "[CLS]",
-      "token_id": 101,
-      "start": 0,
-      "end": 0,
-      "results": {
-        "O": 9.4140625,
-        "B-MISC": -1.15625,
-        "I-MISC": -0.859375,
-        "B-PER": -1.2744141,
-        "I-PER": -1.6552734,
-        "B-ORG": -0.88378906,
-        "I-ORG": -0.9345703,
-        "B-LOC": -1.2275391,
-        "I-LOC": -1.4042969
-      }
-    },
-    {
-      "token": "Apple",
-      "token_id": 6207,
-      "start": 0,
-      "end": 5,
-      "results": {
-        "B-ORG": 6.7578125,
-        "O": -1.7929688,
-        "B-LOC": 0.6015625,
-        "B-MISC": 0.2467041,
-        "B-PER": 0.17675781,
-        "I-ORG": -0.6484375,
-        "I-MISC": -1.9873047,
-        "I-LOC": -1.3808594,
-        "I-PER": -2.21875
-      }
-    }
-  ]
-]
-```
-Important, this uses the `predict_tokens` endpoint for token-level classification. The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict_tokens`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a NER endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Apple is looking at buying U.K. startup for $1 billion
-    - - John works at Google in Mountain View, California
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-tanaos-tanaos-ner-v1-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: tanaos/tanaos-NER-v1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1/config.yaml
deleted file mode 100644
index 1354122c8..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-tanaos-tanaos-ner-v1/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-      - - Apple is looking at buying U.K. startup for $1 billion
-      - - John works at Google in Mountain View, California
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-Bert-tanaos-tanaos-ner-v1-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: tanaos/tanaos-NER-v1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-  version_overrides:
-    engine_builder_version: null
-    bei_bert_version: 1.8.6.ner
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/README.md
deleted file mode 100644
index 3233c22d5..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with TaylorAI/bge-micro-v2
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with TaylorAI/bge-micro-v2. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [TaylorAI/bge-micro-v2](https://huggingface.co/TaylorAI/bge-micro-v2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-TaylorAI/bge-micro-v2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-taylorai-bge-micro-v2-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-taylorai-bge-micro-v2-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: TaylorAI/bge-micro-v2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/config.yaml
deleted file mode 100644
index e03d912f4..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-taylorai-bge-micro-v2/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-taylorai-bge-micro-v2-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: TaylorAI/bge-micro-v2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/README.md
deleted file mode 100644
index 954dc8fd7..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI-Bert (Baseten-Embeddings-Inference-BERT) with voyageai/voyage-4-nano
-
-This is a Deployment for BEI-Bert (Baseten-Embeddings-Inference-BERT) with voyageai/voyage-4-nano. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [voyageai/voyage-4-nano](https://huggingface.co/voyageai/voyage-4-nano).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-voyageai/voyage-4-nano  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-Bert-voyageai-voyage-4-nano-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-voyageai-voyage-4-nano-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: voyageai/voyage-4-nano
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/config.yaml
deleted file mode 100644
index 6d709999a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-Bert-voyageai-voyage-4-nano/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-Bert-voyageai-voyage-4-nano-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder_bert
-    checkpoint_repository:
-      repo: voyageai/voyage-4-nano
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md
deleted file mode 100644
index 18f7e2650..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with allenai/Llama-3.1-Tulu-3-8B-Reward-Model
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with allenai/Llama-3.1-Tulu-3-8B-Reward-Model. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [allenai/Llama-3.1-Tulu-3-8B-RM](https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B-RM).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-allenai/Llama-3.1-Tulu-3-8B-RM  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: allenai/Llama-3.1-Tulu-3-8B-RM
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml
deleted file mode 100644
index 7b45b33c0..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: allenai/Llama-3.1-Tulu-3-8B-RM
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/README.md
deleted file mode 100644
index 7c5af8417..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with BAAI/bge-en-icl-embedding
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with BAAI/bge-en-icl-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-en-icl](https://huggingface.co/BAAI/bge-en-icl).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-BAAI/bge-en-icl  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-baai-bge-en-icl-embedding-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-en-icl-embedding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-en-icl
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/config.yaml
deleted file mode 100644
index f5919a8f7..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-en-icl-embedding-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-en-icl-embedding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-en-icl
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/README.md
deleted file mode 100644
index 9787f7583..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with BAAI/bge-large-en-v1.5-embedding
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with BAAI/bge-large-en-v1.5-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-BAAI/bge-large-en-v1.5  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-baai-bge-large-en-v1.5-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-large-en-v1.5-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-large-en-v1.5
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/config.yaml
deleted file mode 100644
index d61622687..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-large-en-v1.5-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-large-en-v1.5-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-large-en-v1.5
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/README.md
deleted file mode 100644
index fde3603e4..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with BAAI/bge-m3-embedding-dense
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with BAAI/bge-m3-embedding-dense. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-BAAI/bge-m3  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-baai-bge-m3-embedding-dense-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-m3-embedding-dense-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-m3
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/config.yaml
deleted file mode 100644
index 80ce00928..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-m3-embedding-dense/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-m3-embedding-dense-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-m3
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/README.md
deleted file mode 100644
index 5c479a4c4..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with BAAI/bge-multilingual-gemma2-multilingual-embedding
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with BAAI/bge-multilingual-gemma2-multilingual-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-multilingual-gemma2](https://huggingface.co/BAAI/bge-multilingual-gemma2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-BAAI/bge-multilingual-gemma2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-baai-bge-multilingual-gemma2-multilingual-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-multilingual-gemma2-multilingual-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-multilingual-gemma2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml
deleted file mode 100644
index 6ec3d19f2..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-baai-bge-multilingual-gemma2-multilingual-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-multilingual-gemma2
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/README.md
deleted file mode 100644
index 3525c7256..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/README.md
+++ /dev/null
@@ -1,185 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with BAAI/bge-reranker-large
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with BAAI/bge-reranker-large. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Reranker models may have at most one label, which contains the score of the reranking.
-
-BAAI/bge-reranker-large  is a reranker model, used to re-rank a list of items, given a query. \nIt is frequently used in search engines, recommendation systems, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-baai-bge-reranker-large-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank`:
-```json
-{
-    "query": "What is Baseten?",
-    "raw_scores": true,
-    "return_text": false,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": true,
-    "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-response = client.rerank(
-    query="What is Baseten?",
-    texts=["Deep Learning is ...", "Baseten is a fast inference provider"],
-    raw_scores=True,
-    return_text=False,
-    truncate=True,
-)
-print(response.data)
-```
-
-Sometimes, you may want to apply a custom template to the texts before reranking them and call the predict endpoint instead:
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    # Custom template function to apply to the texts
-    # a popular template might be "{query}\n{document}"
-    # or also chat-style templates like "User: {query}\nDocument: {document}"
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["What is baseten? A: Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-
-### Requests python library
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank",
-    json={
-    "query": "What is Baseten?",
-    "raw_scores": True,
-    "return_text": False,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": True,
-    "truncation_direction": "Right"
-}
-```
-Returns:
-```json
-[
-  {
-    "index": 0,
-    "score": 1,
-    "text": "Deep Learning is ..."
-  }
-]
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/rerank`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI.com does not have a rerank endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-baai-bge-reranker-large-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-reranker-large
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/config.yaml
deleted file mode 100644
index 0e0ef388d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-large/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-baai-bge-reranker-large-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-reranker-large
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/README.md
deleted file mode 100644
index e437588f8..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/README.md
+++ /dev/null
@@ -1,185 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with BAAI/bge-reranker-v2-m3-multilingual
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with BAAI/bge-reranker-v2-m3-multilingual. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Reranker models may have at most one label, which contains the score of the reranking.
-
-BAAI/bge-reranker-v2-m3  is a reranker model, used to re-rank a list of items, given a query. \nIt is frequently used in search engines, recommendation systems, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-baai-bge-reranker-v2-m3-multilingual-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank`:
-```json
-{
-    "query": "What is Baseten?",
-    "raw_scores": true,
-    "return_text": false,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": true,
-    "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-response = client.rerank(
-    query="What is Baseten?",
-    texts=["Deep Learning is ...", "Baseten is a fast inference provider"],
-    raw_scores=True,
-    return_text=False,
-    truncate=True,
-)
-print(response.data)
-```
-
-Sometimes, you may want to apply a custom template to the texts before reranking them and call the predict endpoint instead:
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    # Custom template function to apply to the texts
-    # a popular template might be "{query}\n{document}"
-    # or also chat-style templates like "User: {query}\nDocument: {document}"
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["What is baseten? A: Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-
-### Requests python library
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank",
-    json={
-    "query": "What is Baseten?",
-    "raw_scores": True,
-    "return_text": False,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": True,
-    "truncation_direction": "Right"
-}
-```
-Returns:
-```json
-[
-  {
-    "index": 0,
-    "score": 1,
-    "text": "Deep Learning is ..."
-  }
-]
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/rerank`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI.com does not have a rerank endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-baai-bge-reranker-v2-m3-multilingual-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-reranker-v2-m3
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/config.yaml
deleted file mode 100644
index 78b8a92b6..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baai-bge-reranker-v2-m3-multilingual/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-baai-bge-reranker-v2-m3-multilingual-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: BAAI/bge-reranker-v2-m3
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md
deleted file mode 100644
index d65b060d7..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification](https://huggingface.co/baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml
deleted file mode 100644
index 365fb6a67..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/README.md
deleted file mode 100644
index a9d3c3f19..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with codefuse-ai/F2LLM-4B-embedding
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with codefuse-ai/F2LLM-4B-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [codefuse-ai/F2LLM-4B](https://huggingface.co/codefuse-ai/F2LLM-4B).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-codefuse-ai/F2LLM-4B  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-codefuse-ai-f2llm-4b-embedding-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-codefuse-ai-f2llm-4b-embedding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: codefuse-ai/F2LLM-4B
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/config.yaml
deleted file mode 100644
index 151d24a5b..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-codefuse-ai-f2llm-4b-embedding-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-codefuse-ai-f2llm-4b-embedding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: codefuse-ai/F2LLM-4B
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md
deleted file mode 100644
index 442d6e974..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with intfloat/e5-mistral-7b-instruct-embedding
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with intfloat/e5-mistral-7b-instruct-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [intfloat/e5-mistral-7b-instruct](https://huggingface.co/intfloat/e5-mistral-7b-instruct).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-intfloat/e5-mistral-7b-instruct  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: intfloat/e5-mistral-7b-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml
deleted file mode 100644
index 2bc0e145c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: intfloat/e5-mistral-7b-instruct
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 2
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/README.md
deleted file mode 100644
index d89fde450..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with jinaai/jina-code-embeddings-0.5b
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with jinaai/jina-code-embeddings-0.5b. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [jinaai/jina-code-embeddings-0.5b](https://huggingface.co/jinaai/jina-code-embeddings-0.5b).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-jinaai/jina-code-embeddings-0.5b  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-jinaai-jina-code-embeddings-0.5b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-jinaai-jina-code-embeddings-0.5b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: jinaai/jina-code-embeddings-0.5b
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/config.yaml
deleted file mode 100644
index 3ec46322f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-jinaai-jina-code-embeddings-0.5b-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-jinaai-jina-code-embeddings-0.5b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: jinaai/jina-code-embeddings-0.5b
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
deleted file mode 100644
index 77fa774fb..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with mixedbread-ai/mxbai-embed-large-v1-embedding
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with mixedbread-ai/mxbai-embed-large-v1-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-mixedbread-ai/mxbai-embed-large-v1  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: mixedbread-ai/mxbai-embed-large-v1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
deleted file mode 100644
index 54a55f9fe..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: mixedbread-ai/mxbai-embed-large-v1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md
deleted file mode 100644
index a422be568..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with mixedbread-ai/mxbai-rerank-base-v2-reranker
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with mixedbread-ai/mxbai-rerank-base-v2-reranker. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/mxbai-rerank-base-v2-seq](https://huggingface.co/michaelfeil/mxbai-rerank-base-v2-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/mxbai-rerank-base-v2-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/mxbai-rerank-base-v2-seq
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml
deleted file mode 100644
index af2ccf720..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/mxbai-rerank-base-v2-seq
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md
deleted file mode 100644
index 24988153f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with mixedbread-ai/mxbai-rerank-large-v2-reranker
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with mixedbread-ai/mxbai-rerank-large-v2-reranker. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/mxbai-rerank-large-v2-seq](https://huggingface.co/michaelfeil/mxbai-rerank-large-v2-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/mxbai-rerank-large-v2-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/mxbai-rerank-large-v2-seq
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml
deleted file mode 100644
index 74b3bd80c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/mxbai-rerank-large-v2-seq
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/README.md
deleted file mode 100644
index 19aea0756..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/README.md
+++ /dev/null
@@ -1,185 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with ncbi/MedCPT-Cross-Encoder-reranker
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with ncbi/MedCPT-Cross-Encoder-reranker. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [ncbi/MedCPT-Cross-Encoder](https://huggingface.co/ncbi/MedCPT-Cross-Encoder).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Reranker models may have at most one label, which contains the score of the reranking.
-
-ncbi/MedCPT-Cross-Encoder  is a reranker model, used to re-rank a list of items, given a query. \nIt is frequently used in search engines, recommendation systems, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-ncbi-medcpt-cross-encoder-reranker-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank`:
-```json
-{
-    "query": "What is Baseten?",
-    "raw_scores": true,
-    "return_text": false,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": true,
-    "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-response = client.rerank(
-    query="What is Baseten?",
-    texts=["Deep Learning is ...", "Baseten is a fast inference provider"],
-    raw_scores=True,
-    return_text=False,
-    truncate=True,
-)
-print(response.data)
-```
-
-Sometimes, you may want to apply a custom template to the texts before reranking them and call the predict endpoint instead:
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    # Custom template function to apply to the texts
-    # a popular template might be "{query}\n{document}"
-    # or also chat-style templates like "User: {query}\nDocument: {document}"
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["What is baseten? A: Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-
-### Requests python library
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank",
-    json={
-    "query": "What is Baseten?",
-    "raw_scores": True,
-    "return_text": False,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": True,
-    "truncation_direction": "Right"
-}
-```
-Returns:
-```json
-[
-  {
-    "index": 0,
-    "score": 1,
-    "text": "Deep Learning is ..."
-  }
-]
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/rerank`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI.com does not have a rerank endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-ncbi-medcpt-cross-encoder-reranker-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: ncbi/MedCPT-Cross-Encoder
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/config.yaml
deleted file mode 100644
index b283fb6ff..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-ncbi-medcpt-cross-encoder-reranker/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-ncbi-medcpt-cross-encoder-reranker-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: ncbi/MedCPT-Cross-Encoder
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /rerank
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/README.md
deleted file mode 100644
index 33d1983c9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with nomic-ai/nomic-embed-code
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with nomic-ai/nomic-embed-code. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nomic-ai/nomic-embed-code](https://huggingface.co/nomic-ai/nomic-embed-code).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-nomic-ai/nomic-embed-code  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-nomic-ai-nomic-embed-code-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-nomic-ai-nomic-embed-code-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: nomic-ai/nomic-embed-code
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/config.yaml
deleted file mode 100644
index 881343a19..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-nomic-ai-nomic-embed-code-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-nomic-ai-nomic-embed-code-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: nomic-ai/nomic-embed-code
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/README.md
deleted file mode 100644
index 3f0227fbd..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/README.md
+++ /dev/null
@@ -1,166 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with papluca/xlm-roberta-base-language-detection-classification
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with papluca/xlm-roberta-base-language-detection-classification. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [papluca/xlm-roberta-base-language-detection](https://huggingface.co/papluca/xlm-roberta-base-language-detection).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-papluca/xlm-roberta-base-language-detection  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-papluca-xlm-roberta-base-language-detection-classification-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-papluca-xlm-roberta-base-language-detection-classification-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: papluca/xlm-roberta-base-language-detection
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/config.yaml
deleted file mode 100644
index 1a97785cb..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-papluca-xlm-roberta-base-language-detection-classification/config.yaml
+++ /dev/null
@@ -1,26 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-papluca-xlm-roberta-base-language-detection-classification-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: papluca/xlm-roberta-base-language-detection
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md
deleted file mode 100644
index 02e60bd4f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-0.6B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-0.6B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-0.6B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml
deleted file mode 100644
index a80a34787..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/README.md
deleted file mode 100644
index dbfef7320..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/README.md
+++ /dev/null
@@ -1,180 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-4B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-4B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-4B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-4B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-4B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-4b-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-4b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-4B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp4
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/config.yaml
deleted file mode 100644
index 9ebddcf57..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp4/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-4b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-4B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp4
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/README.md
deleted file mode 100644
index 913dbeedd..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-4B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-4B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-4B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-4B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-4B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-4b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-4b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-4B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml
deleted file mode 100644
index 21b3da96b..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-4b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-4B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/README.md
deleted file mode 100644
index c0d510c2f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-8B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-8B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-8B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-8B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-8B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-8b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-8b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-8B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml
deleted file mode 100644
index f7b30d76a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-8b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-8B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md
deleted file mode 100644
index b5797c630..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-0.6B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-0.6B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Reranker-0.6B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-0.6B-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/Qwen3-Reranker-0.6B-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-reranker-0.6b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-0.6B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml
deleted file mode 100644
index de310dbeb..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-0.6B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/README.md
deleted file mode 100644
index 459cb3ce1..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-4B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-4B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Reranker-4B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-4B-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/Qwen3-Reranker-4B-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-reranker-4b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-4b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-4B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml
deleted file mode 100644
index 0d85c1a24..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-4b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-4B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/README.md
deleted file mode 100644
index e6ad83e09..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-8B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-8B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Reranker-8B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-8B-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/Qwen3-Reranker-8B-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-reranker-8b-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-8b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-8B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp4
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/config.yaml
deleted file mode 100644
index a15f890df..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp4/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-8b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-8B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp4
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/README.md
deleted file mode 100644
index a1d7d7d28..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-8B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-8B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Reranker-8B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-8B-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/Qwen3-Reranker-8B-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-reranker-8b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-8b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-8B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml
deleted file mode 100644
index cb254db4d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-8b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-8B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/README.md
deleted file mode 100644
index ef8e1b817..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Salesforce/SFR-Embedding-Mistral
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Salesforce/SFR-Embedding-Mistral. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Salesforce/SFR-Embedding-Mistral](https://huggingface.co/Salesforce/SFR-Embedding-Mistral).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Salesforce/SFR-Embedding-Mistral  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-salesforce-sfr-embedding-mistral-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-salesforce-sfr-embedding-mistral-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: Salesforce/SFR-Embedding-Mistral
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/config.yaml
deleted file mode 100644
index ccd3c42d9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-salesforce-sfr-embedding-mistral-fp8/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-salesforce-sfr-embedding-mistral-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: Salesforce/SFR-Embedding-Mistral
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/README.md
deleted file mode 100644
index d5bf004d5..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/README.md
+++ /dev/null
@@ -1,166 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with SamLowe/roberta-base-go_emotions-classification
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with SamLowe/roberta-base-go_emotions-classification. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [SamLowe/roberta-base-go_emotions](https://huggingface.co/SamLowe/roberta-base-go_emotions).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-SamLowe/roberta-base-go_emotions  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-samlowe-roberta-base-go_emotions-classification-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-samlowe-roberta-base-go_emotions-classification-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: SamLowe/roberta-base-go_emotions
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/config.yaml
deleted file mode 100644
index 966d5a3c5..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-samlowe-roberta-base-go_emotions-classification/config.yaml
+++ /dev/null
@@ -1,26 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-samlowe-roberta-base-go_emotions-classification-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: SamLowe/roberta-base-go_emotions
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md
deleted file mode 100644
index 398996972..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Skywork/Skywork-Reward-Llama-3.1-8B-v0.2-Reward-Model
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Skywork/Skywork-Reward-Llama-3.1-8B-v0.2-Reward-Model. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Skywork/Skywork-Reward-Llama-3.1-8B-v0.2](https://huggingface.co/Skywork/Skywork-Reward-Llama-3.1-8B-v0.2).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-Skywork/Skywork-Reward-Llama-3.1-8B-v0.2  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-### Requests python library
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: Skywork/Skywork-Reward-Llama-3.1-8B-v0.2
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml
deleted file mode 100644
index 43afb3725..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    inputs:
-    - - Baseten is a fast inference provider
-    - - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: Skywork/Skywork-Reward-Llama-3.1-8B-v0.2
-      revision: main
-      source: HF
-    max_num_tokens: 131072
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/README.md
deleted file mode 100644
index e85501e4f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Snowflake/snowflake-arctic-embed-l-v2.0
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Snowflake/snowflake-arctic-embed-l-v2.0. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Snowflake/snowflake-arctic-embed-l-v2.0](https://huggingface.co/Snowflake/snowflake-arctic-embed-l-v2.0).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Snowflake/snowflake-arctic-embed-l-v2.0  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-snowflake-snowflake-arctic-embed-l-v2.0-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-snowflake-snowflake-arctic-embed-l-v2.0-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: Snowflake/snowflake-arctic-embed-l-v2.0
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/config.yaml
deleted file mode 100644
index 20d35c634..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-snowflake-snowflake-arctic-embed-l-v2.0/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-snowflake-snowflake-arctic-embed-l-v2.0-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: Snowflake/snowflake-arctic-embed-l-v2.0
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/README.md
deleted file mode 100644
index f30fa8fdc..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with WhereIsAI/UAE-Large-V1-embedding
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with WhereIsAI/UAE-Large-V1-embedding. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [WhereIsAI/UAE-Large-V1](https://huggingface.co/WhereIsAI/UAE-Large-V1).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-WhereIsAI/UAE-Large-V1  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-whereisai-uae-large-v1-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-whereisai-uae-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: WhereIsAI/UAE-Large-V1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/config.yaml
deleted file mode 100644
index 5e490a51b..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BEI-whereisai-uae-large-v1-embedding/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-whereisai-uae-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: WhereIsAI/UAE-Large-V1
-      revision: main
-      source: HF
-    max_num_tokens: 16384
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md
deleted file mode 100644
index 0f2df2f5d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# Baseten Inference Stack with deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-
-This is a Deployment for Baseten Inference Stack with deepseek-ai/DeepSeek-R1-Distill-Llama-70B. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-deepseek-ai/DeepSeek-R1-Distill-Llama-70B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-      revision: main
-      source: HF
-    quantization_type: fp4_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml
deleted file mode 100644
index b9d714cee..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-      revision: main
-      source: HF
-    quantization_type: fp4_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md
deleted file mode 100644
index 388de80ff..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# Baseten Inference Stack with meta-llama/Llama-3.2-3B-Instruct
-
-This is a Deployment for Baseten Inference Stack with meta-llama/Llama-3.2-3B-Instruct. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.2-3B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4_mlp_only`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    quantization_type: fp4_mlp_only
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml
deleted file mode 100644
index 8f9a8bc3e..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    quantization_type: fp4_mlp_only
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/README.md
deleted file mode 100644
index 2a5883753..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# Baseten Inference Stack with meta-llama/Llama-3.2-3B-Instruct
-
-This is a Deployment for Baseten Inference Stack with meta-llama/Llama-3.2-3B-Instruct. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.2-3B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-meta-llama-llama-3.2-3b-instruct-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    quantization_type: fp8_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
deleted file mode 100644
index 4695f8ba1..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    quantization_type: fp8_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/README.md
deleted file mode 100644
index a4a9f741b..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# Baseten Inference Stack with meta-llama/Llama-3.3-70B-Instruct
-
-This is a Deployment for Baseten Inference Stack with meta-llama/Llama-3.3-70B-Instruct. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.3-70B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-meta-llama-llama-3.3-70b-instruct-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    quantization_type: fp4
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
deleted file mode 100644
index 9d864be6d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    quantization_type: fp4
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/README.md
deleted file mode 100644
index 0f07efb6d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# Baseten Inference Stack with nvidia/Llama-3.1-8B-Instruct-FP4
-
-This is a Deployment for Baseten Inference Stack with nvidia/Llama-3.1-8B-Instruct-FP4. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nvidia/Llama-3.1-8B-Instruct-FP4](https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP4).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-nvidia/Llama-3.1-8B-Instruct-FP4  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-nvidia-llama-3.1-8b-instruct-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-nvidia-llama-3.1-8b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: nvidia/Llama-3.1-8B-Instruct-FP4
-      revision: main
-      source: HF
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/config.yaml
deleted file mode 100644
index ff8e1215d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-llama-3.1-8b-instruct-fp4/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-nvidia-llama-3.1-8b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: nvidia/Llama-3.1-8B-Instruct-FP4
-      revision: main
-      source: HF
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/README.md
deleted file mode 100644
index aba79f029..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# Baseten Inference Stack with nvidia/Qwen3-30B-A3B-FP4
-
-This is a Deployment for Baseten Inference Stack with nvidia/Qwen3-30B-A3B-FP4. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nvidia/Qwen3-30B-A3B-FP4](https://huggingface.co/nvidia/Qwen3-30B-A3B-FP4).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-nvidia/Qwen3-30B-A3B-FP4  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-nvidia-qwen3-30b-a3b-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-nvidia-qwen3-30b-a3b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: nvidia/Qwen3-30B-A3B-FP4
-      revision: main
-      source: HF
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/config.yaml
deleted file mode 100644
index a4f95eb8d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-30b-a3b-fp4/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-nvidia-qwen3-30b-a3b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: nvidia/Qwen3-30B-A3B-FP4
-      revision: main
-      source: HF
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/README.md
deleted file mode 100644
index 08d12366d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# Baseten Inference Stack with nvidia/Qwen3-8B-FP4
-
-This is a Deployment for Baseten Inference Stack with nvidia/Qwen3-8B-FP4. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nvidia/Qwen3-8B-FP4](https://huggingface.co/nvidia/Qwen3-8B-FP4).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-nvidia/Qwen3-8B-FP4  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-nvidia-qwen3-8b-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-nvidia-qwen3-8b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: nvidia/Qwen3-8B-FP4
-      revision: main
-      source: HF
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/config.yaml
deleted file mode 100644
index 779a9d144..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-nvidia-qwen3-8b-fp4/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-nvidia-qwen3-8b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: nvidia/Qwen3-8B-FP4
-      revision: main
-      source: HF
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/README.md
deleted file mode 100644
index 4b4cae484..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# Baseten Inference Stack with Qwen/Qwen2.5-Coder-7B-Instruct
-
-This is a Deployment for Baseten Inference Stack with Qwen/Qwen2.5-Coder-7B-Instruct. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen2.5-Coder-7B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-qwen-qwen2.5-coder-7b-instruct-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen2.5-coder-7b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp4
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml
deleted file mode 100644
index 5b2b418d3..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen2.5-coder-7b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp4
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/README.md
deleted file mode 100644
index 90f588260..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# Baseten Inference Stack with Qwen/Qwen2.5-Coder-7B-Instruct
-
-This is a Deployment for Baseten Inference Stack with Qwen/Qwen2.5-Coder-7B-Instruct. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen2.5-Coder-7B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-qwen-qwen2.5-coder-7b-instruct-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen2.5-coder-7b-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/config.yaml
deleted file mode 100644
index fc22eaf71..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen2.5-coder-7b-instruct/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen2.5-coder-7b-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: no_quant
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/README.md
deleted file mode 100644
index f7ee35ec1..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# Baseten Inference Stack with Qwen/Qwen3-32B
-
-This is a Deployment for Baseten Inference Stack with Qwen/Qwen3-32B. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-32B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-qwen-qwen3-32b-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen3-32b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    quantization_type: fp4_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/config.yaml
deleted file mode 100644
index d52a470a0..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-32b-fp4/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen3-32b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    quantization_type: fp4_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/README.md
deleted file mode 100644
index ce82ab191..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# Baseten Inference Stack with Qwen/Qwen3-4B
-
-This is a Deployment for Baseten Inference Stack with Qwen/Qwen3-4B. Baseten Inference Stack is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Baseten Inference Stack you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-- *fp4 quantization* deployments on B200 GPUs to get even more speed
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-4B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BISV2-qwen-qwen3-4b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen3-4b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen3-4B
-      revision: main
-      source: HF
-    quantization_type: fp8_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/config.yaml
deleted file mode 100644
index be29a840a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/BISV2-qwen-qwen3-4b-fp8/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: BISV2-qwen-qwen3-4b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  inference_stack: v2
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen3-4B
-      revision: main
-      source: HF
-    quantization_type: fp8_kv
-  runtime:
-    max_batch_size: 32
-    max_num_tokens: 32768
-    max_seq_len: 32768
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md
deleted file mode 100644
index ad96330af..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# TensorRT-LLM Briton with deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-
-This is a Deployment for TensorRT-LLM Briton with deepseek-ai/DeepSeek-R1-Distill-Llama-70B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-deepseek-ai/DeepSeek-R1-Distill-Llama-70B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml
deleted file mode 100644
index 296c65dde..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md
deleted file mode 100644
index 67bf8b5a9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md
+++ /dev/null
@@ -1,173 +0,0 @@
-# TensorRT-LLM Briton with deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
-
-This is a Deployment for TensorRT-LLM Briton with deepseek-ai/DeepSeek-R1-Distill-Qwen-32B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-deepseek-ai/DeepSeek-R1-Distill-Qwen-32B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml
deleted file mode 100644
index 43b6dc35d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/README.md
deleted file mode 100644
index 80b69da8a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# TensorRT-LLM Briton with google/gemma-3-1b-it
-
-This is a Deployment for TensorRT-LLM Briton with google/gemma-3-1b-it. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [unsloth/gemma-3-1b-it](https://huggingface.co/unsloth/gemma-3-1b-it).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-unsloth/gemma-3-1b-it  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-google-gemma-3-1b-it-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-1b-it-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: unsloth/gemma-3-1b-it
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/config.yaml
deleted file mode 100644
index d051babbe..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-1b-it/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-1b-it-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: unsloth/gemma-3-1b-it
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/README.md
deleted file mode 100644
index dccff493b..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# TensorRT-LLM Briton with google/gemma-3-270m-it
-
-This is a Deployment for TensorRT-LLM Briton with google/gemma-3-270m-it. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [google/gemma-3-270m-it](https://huggingface.co/google/gemma-3-270m-it).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-google/gemma-3-270m-it  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-google-gemma-3-270m-it-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-270m-it-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: google/gemma-3-270m-it
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/config.yaml
deleted file mode 100644
index 5a29f2972..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-270m-it/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-270m-it-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: google/gemma-3-270m-it
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/README.md
deleted file mode 100644
index 1b14d9bfe..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# TensorRT-LLM Briton with google/gemma-3-27b-it-speculative-lookahead
-
-This is a Deployment for TensorRT-LLM Briton with google/gemma-3-27b-it-speculative-lookahead. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [baseten/gemma-3-27b-causallm-it](https://huggingface.co/baseten/gemma-3-27b-causallm-it).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-baseten/gemma-3-27b-causallm-it  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-google-gemma-3-27b-it-speculative-lookahead-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-27b-it-speculative-lookahead-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: baseten/gemma-3-27b-causallm-it
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 131072
-    max_seq_len: 131072
-    quantization_type: no_quant
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/config.yaml
deleted file mode 100644
index 72a8349b9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it-speculative-lookahead/config.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-27b-it-speculative-lookahead-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: baseten/gemma-3-27b-causallm-it
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 131072
-    max_seq_len: 131072
-    quantization_type: no_quant
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/README.md
deleted file mode 100644
index f0f5f39c9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# TensorRT-LLM Briton with google/gemma-3-27b-it
-
-This is a Deployment for TensorRT-LLM Briton with google/gemma-3-27b-it. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [baseten/gemma-3-27b-causallm-it](https://huggingface.co/baseten/gemma-3-27b-causallm-it).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-baseten/gemma-3-27b-causallm-it  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-google-gemma-3-27b-it-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-27b-it-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: baseten/gemma-3-27b-causallm-it
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/config.yaml
deleted file mode 100644
index c9666a8f9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-google-gemma-3-27b-it/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-google-gemma-3-27b-it-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: baseten/gemma-3-27b-causallm-it
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/README.md
deleted file mode 100644
index 44d49ec99..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.1-405B
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.1-405B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.1-405B](https://huggingface.co/meta-llama/Llama-3.1-405B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.1-405B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.1-405b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.1-405b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:8
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.1-405B
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 8
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/config.yaml
deleted file mode 100644
index 010a4e7f6..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.1-405b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:8
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.1-405B
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 8
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md
deleted file mode 100644
index a3b93f923..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.1-8B-Instruct-with-speculative-lookahead-decoding
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.1-8B-Instruct-with-speculative-lookahead-decoding. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.1-8B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.1-8B-Instruct
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 131072
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
deleted file mode 100644
index 85829cea5..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.1-8B-Instruct
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 131072
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/README.md
deleted file mode 100644
index f15bc82be..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/README.md
+++ /dev/null
@@ -1,173 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.2-1B-Instruct
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.2-1B-Instruct. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.2-1B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.2-1b-instruct-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-1b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-1B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/config.yaml
deleted file mode 100644
index 9006e0ee7..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-1b-instruct-fp8/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-1b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-1B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md
deleted file mode 100644
index 36ff57234..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md
+++ /dev/null
@@ -1,174 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.2-3B-Instruct-calib-dataset
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.2-3B-Instruct-calib-dataset. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.2-3B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_config:
-      calib_dataset: baseten/quant_calibration_dataset_v1
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml
deleted file mode 100644
index 47a5b3b1c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_config:
-      calib_dataset: baseten/quant_calibration_dataset_v1
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/README.md
deleted file mode 100644
index e70711dce..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/README.md
+++ /dev/null
@@ -1,172 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.2-3B-Instruct
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.2-3B-Instruct. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.2-3B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.2-3b-instruct-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
deleted file mode 100644
index 8d395a474..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/README.md
deleted file mode 100644
index 51edf8821..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.2-3B-Instruct
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.2-3B-Instruct. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.2-3B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.2-3b-instruct-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-3b-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/config.yaml
deleted file mode 100644
index dcfb1d8a9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.2-3b-instruct/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.2-3b-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.2-3B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/README.md
deleted file mode 100644
index aee9a97c4..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/README.md
+++ /dev/null
@@ -1,169 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.3-70B-Instruct
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.3-70B-Instruct. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.3-70B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.3-70b-instruct-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    quantization_type: fp4
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
deleted file mode 100644
index da463d4a4..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 4
-    quantization_type: fp4
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/README.md
deleted file mode 100644
index 13d8b2b65..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.3-70B-Instruct
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.3-70B-Instruct. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.3-70B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.3-70b-instruct-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.3-70b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/config.yaml
deleted file mode 100644
index 215dd074c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-fp8/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.3-70b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md
deleted file mode 100644
index c401138ee..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.3-70B-Instruct-tp4
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.3-70B-Instruct-tp4. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.3-70B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 4
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml
deleted file mode 100644
index be35731dc..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 4
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/README.md
deleted file mode 100644
index f35cf47b1..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# TensorRT-LLM Briton with microsoft/phi-4
-
-This is a Deployment for TensorRT-LLM Briton with microsoft/phi-4. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [unsloth/phi-4](https://huggingface.co/unsloth/phi-4).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-unsloth/phi-4  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-microsoft-phi-4-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-microsoft-phi-4-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: unsloth/phi-4
-      revision: main
-      source: HF
-    max_seq_len: 16384
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/config.yaml
deleted file mode 100644
index 972fad2c8..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-microsoft-phi-4-fp8/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-microsoft-phi-4-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: unsloth/phi-4
-      revision: main
-      source: HF
-    max_seq_len: 16384
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/README.md
deleted file mode 100644
index 282dd9fec..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# TensorRT-LLM Briton with mistralai/Mistral-7B-Instruct-v0.3
-
-This is a Deployment for TensorRT-LLM Briton with mistralai/Mistral-7B-Instruct-v0.3. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-mistralai/Mistral-7B-Instruct-v0.3  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-mistralai-mistral-7b-instruct-v0.3-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-mistralai-mistral-7b-instruct-v0.3-truss-example
-python_version: py39
-resources:
-  accelerator: A10G:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mistral-7B-Instruct-v0.3
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_type: no_quant
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/config.yaml
deleted file mode 100644
index eec4a13d8..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-7b-instruct-v0.3/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-mistralai-mistral-7b-instruct-v0.3-truss-example
-python_version: py39
-resources:
-  accelerator: A10G:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mistral-7B-Instruct-v0.3
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_type: no_quant
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/README.md
deleted file mode 100644
index b6e3eff92..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/README.md
+++ /dev/null
@@ -1,172 +0,0 @@
-# TensorRT-LLM Briton with mistralai/Mistral-Small-24B-Instruct-2501
-
-This is a Deployment for TensorRT-LLM Briton with mistralai/Mistral-Small-24B-Instruct-2501. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [mistralai/Mistral-Small-24B-Instruct-2501](https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-mistralai/Mistral-Small-24B-Instruct-2501  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-mistralai-mistral-small-24b-instruct-2501-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-mistralai-mistral-small-24b-instruct-2501-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mistral-Small-24B-Instruct-2501
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml
deleted file mode 100644
index 97449aeb5..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-mistralai-mistral-small-24b-instruct-2501-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mistral-Small-24B-Instruct-2501
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/README.md
deleted file mode 100644
index f11d57006..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/README.md
+++ /dev/null
@@ -1,172 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen2.5-72B-Instruct-tp2
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen2.5-72B-Instruct-tp2. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen2.5-72B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen2.5-72b-instruct-tp2-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-72b-instruct-tp2-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-72B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml
deleted file mode 100644
index f07e9d02f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-72b-instruct-tp2-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-72B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md
deleted file mode 100644
index c11eafc02..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md
+++ /dev/null
@@ -1,182 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen2.5-7B-Instruct-with-speculative-lookahead-decoding
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen2.5-7B-Instruct-with-speculative-lookahead-decoding. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen2.5-7B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-7B-Instruct
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 32768
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
deleted file mode 100644
index 7438254e5..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
+++ /dev/null
@@ -1,43 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-7B-Instruct
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 32768
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md
deleted file mode 100644
index a4e7260dc..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md
+++ /dev/null
@@ -1,172 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen2.5-Coder-7B-Instruct-calib-dataset
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen2.5-Coder-7B-Instruct-calib-dataset. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen2.5-Coder-7B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4_mlp_only`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp4_mlp_only
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml
deleted file mode 100644
index f7b8b2cfc..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp4_mlp_only
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md
deleted file mode 100644
index 850c58dc1..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md
+++ /dev/null
@@ -1,182 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen2.5-Coder-7B-Instruct-min-latency
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen2.5-Coder-7B-Instruct-min-latency. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen2.5-Coder-7B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 32768
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 32
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      num_draft_tokens: 61
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml
deleted file mode 100644
index 40f515706..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml
+++ /dev/null
@@ -1,43 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 32768
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 32
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      num_draft_tokens: 61
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md
deleted file mode 100644
index 2fa13067b..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-235B-A22B-Instruct-2507
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-235B-A22B-Instruct-2507. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-235B-A22B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-235B-A22B-Instruct-2507  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:8
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-235B-A22B-Instruct-2507
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 262144
-    max_seq_len: 262144
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 32
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      num_draft_tokens: 61
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 8
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml
deleted file mode 100644
index 3e61aa0ce..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100:8
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-235B-A22B-Instruct-2507
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 262144
-    max_seq_len: 262144
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 32
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      num_draft_tokens: 61
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 8
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/README.md
deleted file mode 100644
index 723d21364..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/README.md
+++ /dev/null
@@ -1,172 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-30B-A3B
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-30B-A3B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-30B-A3B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-30b-a3b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-30b-a3b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-30B-A3B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/config.yaml
deleted file mode 100644
index 9a2e2a2b9..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-fp8/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-30b-a3b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-30B-A3B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md
deleted file mode 100644
index 885c5b9ea..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md
+++ /dev/null
@@ -1,172 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-30B-A3B-Instruct-2507
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-30B-A3B-Instruct-2507. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-30B-A3B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-30B-A3B-Instruct-2507  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-30B-A3B-Instruct-2507
-      revision: main
-      source: HF
-    max_seq_len: 262144
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml
deleted file mode 100644
index c767998f0..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-30B-A3B-Instruct-2507
-      revision: main
-      source: HF
-    max_seq_len: 262144
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/README.md
deleted file mode 100644
index c3a55f19e..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/README.md
+++ /dev/null
@@ -1,168 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-30B-A3B
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-30B-A3B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-30B-A3B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-30b-a3b-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-30b-a3b-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-30B-A3B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    quantization_type: no_quant
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/config.yaml
deleted file mode 100644
index 994438d5c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-30b-a3b/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-30b-a3b-truss-example
-python_version: py39
-resources:
-  accelerator: H100:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-30B-A3B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    quantization_type: no_quant
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/README.md
deleted file mode 100644
index 0f5ae98a3..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/README.md
+++ /dev/null
@@ -1,170 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-32B
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-32B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-32B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-32b-fp4-mlp-only-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4_mlp_only`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-32b-fp4-mlp-only-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_type: fp4_mlp_only
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/config.yaml
deleted file mode 100644
index 5e919d604..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4-mlp-only/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-32b-fp4-mlp-only-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_type: fp4_mlp_only
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/README.md
deleted file mode 100644
index a37b059cd..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/README.md
+++ /dev/null
@@ -1,170 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-32B
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-32B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-32B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-32b-fp4-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp4_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-32b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_type: fp4_kv
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/config.yaml
deleted file mode 100644
index 94709b289..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp4/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-32b-fp4-truss-example
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_type: fp4_kv
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/README.md
deleted file mode 100644
index 265a4731b..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/README.md
+++ /dev/null
@@ -1,173 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-32B
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-32B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-32B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-32b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-32b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/config.yaml
deleted file mode 100644
index f99994354..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-32b-fp8/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-32b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/README.md
deleted file mode 100644
index c8834a83e..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/README.md
+++ /dev/null
@@ -1,181 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-8B-min-latency
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-8B-min-latency. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/Qwen3-8B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-8b-min-latency-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-8b-min-latency-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-8B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 40960
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 32
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      num_draft_tokens: 61
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/config.yaml
deleted file mode 100644
index e809dd7c4..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-8b-min-latency-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-8B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 40960
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 32
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      num_draft_tokens: 61
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/README.md
deleted file mode 100644
index 007468ceb..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/README.md
+++ /dev/null
@@ -1,173 +0,0 @@
-# TensorRT-LLM Briton with Qwen/QwQ-32B-reasoning
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/QwQ-32B-reasoning. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/QwQ-32B](https://huggingface.co/Qwen/QwQ-32B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/QwQ-32B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwq-32b-reasoning-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwq-32b-reasoning-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/QwQ-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/config.yaml
deleted file mode 100644
index 8ef63567f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-fp8/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwq-32b-reasoning-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/QwQ-32B
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/README.md
deleted file mode 100644
index 41f705373..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/README.md
+++ /dev/null
@@ -1,182 +0,0 @@
-# TensorRT-LLM Briton with Qwen/QwQ-32B-reasoning-with-speculative
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/QwQ-32B-reasoning-with-speculative. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/QwQ-32B](https://huggingface.co/Qwen/QwQ-32B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-Qwen/QwQ-32B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwq-32b-reasoning-with-speculative-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwq-32b-reasoning-with-speculative-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/QwQ-32B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 40960
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml
deleted file mode 100644
index 5dfcd2a1a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml
+++ /dev/null
@@ -1,43 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwq-32b-reasoning-with-speculative-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/QwQ-32B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 40960
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    quantization_config:
-      calib_max_seq_length: 2048
-      calib_size: 2048
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 3
-      lookahead_windows_size: 3
-      num_draft_tokens: 41
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/README.md b/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/README.md
deleted file mode 100644
index dc5f1caa2..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/README.md
+++ /dev/null
@@ -1,171 +0,0 @@
-# TensorRT-LLM Briton with tiiuae/Falcon3-10B-Instruct
-
-This is a Deployment for TensorRT-LLM Briton with tiiuae/Falcon3-10B-Instruct. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [tiiuae/Falcon3-10B-Instruct](https://huggingface.co/tiiuae/Falcon3-10B-Instruct).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-tiiuae/Falcon3-10B-Instruct  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-tiiuae-falcon3-10b-instruct-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-tiiuae-falcon3-10b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: tiiuae/Falcon3-10B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/config.yaml b/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/config.yaml
deleted file mode 100644
index 3548da52d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/Briton-tiiuae-falcon3-10b-instruct-fp8/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-tiiuae-falcon3-10b-instruct-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4:2
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: tiiuae/Falcon3-10B-Instruct
-      revision: main
-      source: HF
-    max_seq_len: 32768
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
-  runtime:
-    enable_chunked_context: true
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/README.md
deleted file mode 100644
index ea445cc61..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with Alibaba-NLP/gte-modernbert-base-embedding
-
-This is a Deployment for Huggingface's text-embeddings-inference with Alibaba-NLP/gte-modernbert-base-embedding. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Alibaba-NLP/gte-modernbert-base  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-alibaba-nlp-gte-modernbert-base-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-modernbert-base
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-alibaba-nlp-gte-modernbert-base-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/config.yaml
deleted file mode 100644
index bf026acfe..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-modernbert-base-embedding/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-modernbert-base
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-alibaba-nlp-gte-modernbert-base-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
deleted file mode 100644
index af1321660..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with Alibaba-NLP/gte-Qwen2-1.5B-instruct-embedding
-
-This is a Deployment for Huggingface's text-embeddings-inference with Alibaba-NLP/gte-Qwen2-1.5B-instruct-embedding. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-Qwen2-1.5B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Alibaba-NLP/gte-Qwen2-1.5B-instruct  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-Qwen2-1.5B-instruct
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
deleted file mode 100644
index 4b5904a87..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-Qwen2-1.5B-instruct
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
deleted file mode 100644
index 06bdaf045..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with Alibaba-NLP/gte-Qwen2-7B-instruct-embedding
-
-This is a Deployment for Huggingface's text-embeddings-inference with Alibaba-NLP/gte-Qwen2-7B-instruct-embedding. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Alibaba-NLP/gte-Qwen2-7B-instruct  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:hopper-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-Qwen2-7B-instruct
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
deleted file mode 100644
index 98e3c0c98..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:hopper-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-Qwen2-7B-instruct
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/README.md
deleted file mode 100644
index 2b6005a5c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/README.md
+++ /dev/null
@@ -1,209 +0,0 @@
-# Huggingface's text-embeddings-inference with Alibaba-NLP/gte-reranker-modernbert-base
-
-This is a Deployment for Huggingface's text-embeddings-inference with Alibaba-NLP/gte-reranker-modernbert-base. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Alibaba-NLP/gte-reranker-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Reranker models may have at most one label, which contains the score of the reranking.
-
-Alibaba-NLP/gte-reranker-modernbert-base  is a reranker model, used to re-rank a list of items, given a query. \nIt is frequently used in search engines, recommendation systems, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-alibaba-nlp-gte-reranker-modernbert-base-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank`:
-```json
-{
-    "query": "What is Baseten?",
-    "raw_scores": true,
-    "return_text": false,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": true,
-    "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-response = client.rerank(
-    query="What is Baseten?",
-    texts=["Deep Learning is ...", "Baseten is a fast inference provider"],
-    raw_scores=True,
-    return_text=False,
-    truncate=True,
-)
-print(response.data)
-```
-
-Sometimes, you may want to apply a custom template to the texts before reranking them and call the predict endpoint instead:
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    # Custom template function to apply to the texts
-    # a popular template might be "{query}\n{document}"
-    # or also chat-style templates like "User: {query}\nDocument: {document}"
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["What is baseten? A: Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-
-### Requests python library
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank",
-    json={
-    "query": "What is Baseten?",
-    "raw_scores": True,
-    "return_text": False,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": True,
-    "truncation_direction": "Right"
-}
-```
-Returns:
-```json
-[
-  {
-    "index": 0,
-    "score": 1,
-    "text": "Deep Learning is ..."
-  }
-]
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/rerank`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI.com does not have a rerank endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /rerank
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-reranker-modernbert-base
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: TEI-alibaba-nlp-gte-reranker-modernbert-base-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/config.yaml
deleted file mode 100644
index daa6fa816..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-alibaba-nlp-gte-reranker-modernbert-base/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /rerank
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: Alibaba-NLP/gte-reranker-modernbert-base
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: TEI-alibaba-nlp-gte-reranker-modernbert-base-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/README.md
deleted file mode 100644
index 600b1a9d1..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/README.md
+++ /dev/null
@@ -1,209 +0,0 @@
-# Huggingface's text-embeddings-inference with BAAI/bge-reranker-large
-
-This is a Deployment for Huggingface's text-embeddings-inference with BAAI/bge-reranker-large. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Reranker models may have at most one label, which contains the score of the reranking.
-
-BAAI/bge-reranker-large  is a reranker model, used to re-rank a list of items, given a query. \nIt is frequently used in search engines, recommendation systems, and more.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-baai-bge-reranker-large-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank`:
-```json
-{
-    "query": "What is Baseten?",
-    "raw_scores": true,
-    "return_text": false,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": true,
-    "truncation_direction": "Right"
-}
-```
-
-### Baseten Performance Client
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-response = client.rerank(
-    query="What is Baseten?",
-    texts=["Deep Learning is ...", "Baseten is a fast inference provider"],
-    raw_scores=True,
-    return_text=False,
-    truncate=True,
-)
-print(response.data)
-```
-
-Sometimes, you may want to apply a custom template to the texts before reranking them and call the predict endpoint instead:
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-def template(text: list[str]) -> list[str]:
-    # Custom template function to apply to the texts
-    # a popular template might be "{query}\n{document}"
-    # or also chat-style templates like "User: {query}\nDocument: {document}"
-    apply = lambda x: f"Custom template: {x}"
-    return [apply(t) for t in text]
-
-response = client.predict(
-    inputs=template(["What is baseten? A: Baseten is a fast inference provider", "Classify this separately."]),
-    raw_scores=True,
-    truncate=True,
-)
-print(response.data)
-```
-
-
-### Requests python library
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank",
-    json={
-    "query": "What is Baseten?",
-    "raw_scores": True,
-    "return_text": False,
-    "texts": [
-        "Deep Learning is ...", "Baseten is a fast inference provider"
-    ],
-    "truncate": True,
-    "truncation_direction": "Right"
-}
-```
-Returns:
-```json
-[
-  {
-    "index": 0,
-    "score": 1,
-    "text": "Deep Learning is ..."
-  }
-]
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/rerank`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI.com does not have a rerank endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:hopper-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /rerank
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: BAAI/bge-reranker-large
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: TEI-baai-bge-reranker-large-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/config.yaml
deleted file mode 100644
index 188e5838a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-baai-bge-reranker-large/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:hopper-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /rerank
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: BAAI/bge-reranker-large
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    query: What is Baseten?
-    raw_scores: true
-    return_text: true
-    texts:
-    - Deep Learning is ...
-    - Baseten is a fast inference provider
-    truncate: true
-    truncation_direction: Right
-model_name: TEI-baai-bge-reranker-large-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/README.md
deleted file mode 100644
index 58ed299cb..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with google/embeddinggemma-300m
-
-This is a Deployment for Huggingface's text-embeddings-inference with google/embeddinggemma-300m. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-google/embeddinggemma-300m  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-google-embeddinggemma-300m-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: google/embeddinggemma-300m
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-google-embeddinggemma-300m-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/config.yaml
deleted file mode 100644
index 9250697ef..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-google-embeddinggemma-300m/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: google/embeddinggemma-300m
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-google-embeddinggemma-300m-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/README.md
deleted file mode 100644
index ca746073c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with intfloat/multilingual-e5-large-instruct
-
-This is a Deployment for Huggingface's text-embeddings-inference with intfloat/multilingual-e5-large-instruct. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-intfloat/multilingual-e5-large-instruct  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-intfloat-multilingual-e5-large-instruct-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: intfloat/multilingual-e5-large-instruct
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-intfloat-multilingual-e5-large-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/config.yaml
deleted file mode 100644
index 98746c8d8..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-intfloat-multilingual-e5-large-instruct/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: intfloat/multilingual-e5-large-instruct
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-intfloat-multilingual-e5-large-instruct-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/README.md
deleted file mode 100644
index 6bad1f846..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with jina-ai/jina-embeddings-v2-base-en
-
-This is a Deployment for Huggingface's text-embeddings-inference with jina-ai/jina-embeddings-v2-base-en. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [jinaai/jina-embeddings-v2-base-en](https://huggingface.co/jinaai/jina-embeddings-v2-base-en).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-jinaai/jina-embeddings-v2-base-en  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-jina-ai-jina-embeddings-v2-base-en-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: jinaai/jina-embeddings-v2-base-en
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-jina-ai-jina-embeddings-v2-base-en-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/config.yaml
deleted file mode 100644
index e68e8c953..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-jina-ai-jina-embeddings-v2-base-en/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: jinaai/jina-embeddings-v2-base-en
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-jina-ai-jina-embeddings-v2-base-en-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/README.md
deleted file mode 100644
index d92a9723a..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with jinaai/jina-embeddings-v2-base-code
-
-This is a Deployment for Huggingface's text-embeddings-inference with jinaai/jina-embeddings-v2-base-code. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [jinaai/jina-embeddings-v2-base-code](https://huggingface.co/jinaai/jina-embeddings-v2-base-code).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-jinaai/jina-embeddings-v2-base-code  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-jinaai-jina-embeddings-v2-base-code-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: jinaai/jina-embeddings-v2-base-code
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-jinaai-jina-embeddings-v2-base-code-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/config.yaml
deleted file mode 100644
index ac49be155..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-jinaai-jina-embeddings-v2-base-code/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: jinaai/jina-embeddings-v2-base-code
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-jinaai-jina-embeddings-v2-base-code-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
deleted file mode 100644
index 074002645..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with mixedbread-ai/mxbai-embed-large-v1-embedding
-
-This is a Deployment for Huggingface's text-embeddings-inference with mixedbread-ai/mxbai-embed-large-v1-embedding. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-mixedbread-ai/mxbai-embed-large-v1  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: mixedbread-ai/mxbai-embed-large-v1
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
deleted file mode 100644
index 57b066b85..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: mixedbread-ai/mxbai-embed-large-v1
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/README.md
deleted file mode 100644
index 03a9dceb8..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with nomic-ai/nomic-embed-text-v1.5
-
-This is a Deployment for Huggingface's text-embeddings-inference with nomic-ai/nomic-embed-text-v1.5. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nomic-ai/nomic-embed-text-v1.5](https://huggingface.co/nomic-ai/nomic-embed-text-v1.5).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-nomic-ai/nomic-embed-text-v1.5  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-nomic-ai-nomic-embed-text-v1.5-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:86-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: nomic-ai/nomic-embed-text-v1.5
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-nomic-ai-nomic-embed-text-v1.5-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/config.yaml
deleted file mode 100644
index 60998c346..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v1.5/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:86-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: nomic-ai/nomic-embed-text-v1.5
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-nomic-ai-nomic-embed-text-v1.5-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/README.md
deleted file mode 100644
index f98793c08..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with nomic-ai/nomic-embed-text-v2-moe
-
-This is a Deployment for Huggingface's text-embeddings-inference with nomic-ai/nomic-embed-text-v2-moe. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [nomic-ai/nomic-embed-text-v2-moe](https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-nomic-ai/nomic-embed-text-v2-moe  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-nomic-ai-nomic-embed-text-v2-moe-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: nomic-ai/nomic-embed-text-v2-moe
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-nomic-ai-nomic-embed-text-v2-moe-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/config.yaml
deleted file mode 100644
index b035c847d..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-nomic-ai-nomic-embed-text-v2-moe/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: nomic-ai/nomic-embed-text-v2-moe
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-nomic-ai-nomic-embed-text-v2-moe-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/README.md
deleted file mode 100644
index 166582a3f..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with redis/langcache-embed-v2
-
-This is a Deployment for Huggingface's text-embeddings-inference with redis/langcache-embed-v2. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [redis/langcache-embed-v2](https://huggingface.co/redis/langcache-embed-v2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-redis/langcache-embed-v2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-redis-langcache-embed-v2-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: redis/langcache-embed-v2
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-redis-langcache-embed-v2-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/config.yaml
deleted file mode 100644
index 3834cc1d7..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-redis-langcache-embed-v2/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:89-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: redis/langcache-embed-v2
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-redis-langcache-embed-v2-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/README.md
deleted file mode 100644
index 676b96dc6..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with sentence-transformers/all-MiniLM-L6-v2-embedding
-
-This is a Deployment for Huggingface's text-embeddings-inference with sentence-transformers/all-MiniLM-L6-v2-embedding. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [sentence-transformers/all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-sentence-transformers/all-MiniLM-L6-v2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-sentence-transformers-all-minilm-l6-v2-embedding-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:turing-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: sentence-transformers/all-MiniLM-L6-v2
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-sentence-transformers-all-minilm-l6-v2-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: T4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
deleted file mode 100644
index 050819f1c..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:turing-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: sentence-transformers/all-MiniLM-L6-v2
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-sentence-transformers-all-minilm-l6-v2-embedding-truss-example
-python_version: py39
-resources:
-  accelerator: T4
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/README.md b/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/README.md
deleted file mode 100644
index 4facc5896..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/README.md
+++ /dev/null
@@ -1,202 +0,0 @@
-# Huggingface's text-embeddings-inference with TaylorAI/bge-micro-v2
-
-This is a Deployment for Huggingface's text-embeddings-inference with TaylorAI/bge-micro-v2. TEI is huggingface's solution for (text) embeddings, reranking models and prediction models.
-
-Supported models are tagged here: https://huggingface.co/models?other=text-embeddings-inference&sort=trending
-
-For TEI you have to perform a manual selection of the Docker Image. We have mirrored the following images:
-```
-CPU	baseten/text-embeddings-inference-mirror:cpu-1.8.3
-Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.8.3
-Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.8.3
-Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.8.3
-Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.8.3
-Hopper (H100/H100 40GB/H200)	baseten/text-embeddings-inference-mirror:hopper-1.8.3
-```
-
-As we are deploying mostly tiny models (<1GB), we are downloading the model weights into the docker image.
-For larger models, we recommend downloading the weights at runtime for faster autoscaling, as the weights don't need to go through decompression of the docker image.
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [TaylorAI/bge-micro-v2](https://huggingface.co/TaylorAI/bge-micro-v2).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-TaylorAI/bge-micro-v2  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2
-```
-
-With `11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model TEI-taylorai-bge-micro-v2-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### Baseten Performance Client
-
-```bash
-pip install baseten-performance-client
-```
-
-```python
-from baseten_performance_client import PerformanceClient
-
-client = PerformanceClient(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync"
-)
-texts = ["Hello world", "Example text", "Another sample"]
-response = client.embed(
-    input=texts,
-    model="my_model",
-    batch_size=4,
-    max_concurrent_requests=32,
-    timeout_s=360,
-    # dimensions=1536 # optional for fp8 models.
-)
-print(response.numpy())
-```
-
-Read more on the [Baseten Performance Client Blog](https://www.baseten.co/blog/your-client-code-matters-10x-higher-embedding-throughput-with-python-and-rust/)
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-    # dimensions=1536 # optional for MRL models.
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment.
-
-```yaml
-base_image:
-  image: baseten/text-embeddings-inference-mirror:86-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: TaylorAI/bge-micro-v2
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-taylorai-bge-micro-v2-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/config.yaml b/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/config.yaml
deleted file mode 100644
index 54fb71b40..000000000
--- a/11-embeddings-reranker-classification-tensorrt/TEI-taylorai-bge-micro-v2/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-base_image:
-  image: baseten/text-embeddings-inference-mirror:86-1.8.3
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/embeddings
-  readiness_endpoint: /health
-  server_port: 7997
-  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
-    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
-    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
-model_cache:
-- ignore_patterns:
-  - '*.pt'
-  - '*.ckpt'
-  - '*.onnx'
-  repo_id: TaylorAI/bge-micro-v2
-  revision: main
-  use_volume: true
-  volume_folder: cached_model
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: TEI-taylorai-bge-micro-v2-truss-example
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
index 528f049cf..3efbf29de 100644
--- a/CONTRIBUTING.md
+++ b/CONTRIBUTING.md
@@ -1,98 +1,22 @@
 # Contributing
 
-Please open a PR to add a new model. Please add new models to the `model_library` folder.
+We welcome new models and improvements to existing examples. Please open a PR.
 
-## Style guide
+## Adding a new example
 
-Model library models should follow this style guide.
+1. Place your example in the category that matches its type (`llm/`, `image/`, `audio/`, `embeddings/`, `infrastructure/`, or `tutorials/`).
+2. Include a `config.yaml` with `model_name`, `description`, and `example_model_input`.
+3. Include a `README.md` with deploy and invoke instructions.
+4. Pin all Python requirements to specific versions.
 
-General style tips:
+## Validate locally
 
-* Truss folder name should use hyphens, not underscores (e.g. `falcon-7b` not `falcon_7b`)
-* Truss folder name should not include the word Truss (e.g. `falcon-7b` not `falcon-7b-truss`)
-* Model name should include parameter count when multiple variants exist (e.g. `Falcon 7B` not `Falcon`)
-
-## README
-
-The model should have a README that follows the layout of [Stable Diffusion XL](stable-diffusion/stable-diffusion-xl-1.0).
-
-
-## Config
-
-Do not include any unnecessary or legacy config lines.
-
-Always include:
-
-* `model_name`
-* `description`
-* `model_metadata`
-  * `example_model_input`
-
-Optionally include:
-* `model_metadata`
-  * `avatar_url` (Avatars/Logos should be 128x128 PNG)
-  * `cover_image_url` (Cover images should be 452x423 PNG)
-  * `tags`
-
-### Requirements
-
-Pin versions for all Python requirements!
-
-Example:
-
-```yaml
-requirements:
-- accelerate==0.20.3
-- bitsandbytes==0.39.1
-- peft==0.3.0
-- protobuf==4.23.3
-- sentencepiece==0.1.99
-- torch==2.0.1
-- transformers==4.30.2
-```
-
-### Secrets
-
-Model library models can access secrets.
-
-If the model requires HuggingFace (e.g. Llama 2), always call the secret `hf_access_token`
-
-Example:
-
-```yaml
-secrets:
-  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
+```bash
+python _internal/bin/test_all.py
 ```
 
-### Hardware requirements
-
-Always configure a model library model with the least expensive hardware required to operate it at a reasonable degree of speed and quality. For example, Stable Diffusion XL defaults to an A10 even though performance is twice as fast on an A100. When these tradeoffs are made, note them in the README.
-
-## Model
-
-### Model I/O
+This runs all config, README, naming, and CI checks. Pre-commit hooks run the same suite automatically.
 
-* Models that support streaming should take a `stream` kwarg that defaults to false
-* Models that take any kind of text input should call it `prompt`
-
-# Testing
-
-If you would like the model to be added the CI job that tests examples very day, add a reference
-to the [ci.yaml](ci.yaml) file.
-
-# Automatic Documentation
-
-Some of the examples in this repo are used to generate automatic documentation on https://truss.baseten.co/.
-
-To add your model to this automatic documentation, add your example to one of the **top-level** categories
-in the repo, and add a `doc.yaml` file that follows the following form:
-
-```yaml
-title: "Text-to-image"
-description: "Building a text-to-image model with SDXL"
-files:
-  - model/model.py
-  - config.yaml
-```
+## Questions?
 
-See the [Introduction doc.yaml file](1_introduction/getting-started-bert/doc.yaml) for an example.
+If your model doesn't fit an existing category, open an issue to discuss placement.
diff --git a/README.md b/README.md
index a878a90dc..3e204219e 100644
--- a/README.md
+++ b/README.md
@@ -2,41 +2,183 @@
 
 [![Truss Examples CI](https://github.com/basetenlabs/truss-examples/actions/workflows/test-examples.yml/badge.svg)](https://github.com/basetenlabs/truss-examples/actions/workflows/test-examples.yml)
 
-Truss is the simplest way to serve AI/ML models in production.
+Production-ready **inference** examples for [Truss](https://truss.baseten.co/),
+the simplest way to serve AI/ML models. Each example is ready to deploy as-is or
+adapt to your own use case.
 
-To get you started with [Truss](https://truss.baseten.co/), this repository has dozens of example models, each ready to deploy as-is or adapt to your needs.
+> **Looking for training and fine-tuning?** See
+> [ml-cookbook](https://github.com/basetenlabs/ml-cookbook) for training
+> recipes, LoRA fine-tuning workflows, and end-to-end training examples on
+> Baseten.
 
-## Installation
+## Quick start
 
-Get the repository with:
+Clone the repository:
 
-```
+```bash
 git clone https://github.com/basetenlabs/truss-examples
+cd truss-examples
 ```
 
-Install Truss with:
+Install Truss:
 
-```
+```bash
 pip install --upgrade truss
 ```
 
-## Deployment
-
-
-Pick a model to deploy by passing a path to that model.
+Deploy any example:
 
 ```bash
-$ # From the truss-examples directory
-$ truss push 02-llm
+truss push tutorials/getting-started-bert
 ```
 
-This will prompt you for an API Key -- fetch one from the
-[Baseten API keys page](https://app.baseten.co/settings/account/api_keys).
-
-## Invocation
-
-Invocation depends on the model's input and output specifications. See individual model READMEs for invocation details.
-
-# Contibuting
-
-We welcome contributions of new models and improvements to existing models. See [CONTRIBUTING.md](CONTRIBUTING.md) for details.
+You will be prompted for an API key. Get one from the [Baseten API keys page](https://app.baseten.co/settings/account/api_keys).
+
+See individual example READMEs for invocation details specific to each model.
+
+## Repository structure
+
+| Category                          | Description                                                                                                                                       | Examples             | Path              |
+| --------------------------------- | ------------------------------------------------------------------------------------------------------------------------------------------------- | -------------------- | ----------------- |
+| [Tutorials](tutorials/)           | Getting-started guides covering core Truss features: BERT, LLMs, streaming, image generation, speech-to-text, caching, dynamic batching, and more | 9                    | `tutorials/`      |
+| [LLMs](llm/)                      | Large language model families including Llama, Qwen, DeepSeek, Mistral, Gemma, Phi, Falcon, and others                                            | 19 families          | `llm/`            |
+| [Embeddings](embeddings/)         | Embedding and reranking models via BEI (48 models), TEI (15 models), CLIP, and text-embeddings-inference                                          | 4 engines, 65 models | `embeddings/`     |
+| [Image](image/)                   | Image generation, editing, and segmentation: Stable Diffusion, Flux, SDXL, ComfyUI, ControlNet, SAM, and more                                     | 15                   | `image/`          |
+| [Audio](audio/)                   | Speech-to-text, text-to-speech, and audio generation: Whisper, Kokoro, Chatterbox, Orpheus, XTTS, MusicGen, and more                              | 19                   | `audio/`          |
+| [Optimized](optimized/)           | Production-grade optimized configs via Briton (33 TRT-LLM models) and BISv2 (11 models)                                                           | 2 engines, 44 models | `optimized/`      |
+| [Infrastructure](infrastructure/) | Patterns and techniques: custom servers, gRPC, Chains, multiprocessing, metrics, model caching, and more                                          | 14                   | `infrastructure/` |
+
+### Tutorials
+
+Step-by-step introductions to Truss concepts and features.
+
+| Example                                                 | Description                                 |
+| ------------------------------------------------------- | ------------------------------------------- |
+| [getting-started-bert](tutorials/getting-started-bert/) | Deploy a BERT model for text classification |
+| [llm-basics](tutorials/llm-basics/)                     | Serve an LLM with Truss                     |
+| [llm-streaming](tutorials/llm-streaming/)               | Stream LLM responses token-by-token         |
+| [image-generation](tutorials/image-generation/)         | Serve an image generation model             |
+| [speech-to-text](tutorials/speech-to-text/)             | Deploy a speech-to-text pipeline            |
+| [cached-weights](tutorials/cached-weights/)             | Cache model weights for faster cold starts  |
+| [private-huggingface](tutorials/private-huggingface/)   | Access private Hugging Face models          |
+| [dynamic-batching](tutorials/dynamic-batching/)         | Enable dynamic batching for throughput      |
+| [system-packages](tutorials/system-packages/)           | Add system-level dependencies               |
+
+### LLMs
+
+Deploy popular large language model families with optimized serving configurations.
+
+| Model family      | Path                                               |
+| ----------------- | -------------------------------------------------- |
+| Llama             | [`llm/llama/`](llm/llama/)                         |
+| Qwen              | [`llm/qwen/`](llm/qwen/)                           |
+| DeepSeek          | [`llm/deepseek/`](llm/deepseek/)                   |
+| Mistral           | [`llm/mistral/`](llm/mistral/)                     |
+| Gemma             | [`llm/gemma/`](llm/gemma/)                         |
+| Phi               | [`llm/phi/`](llm/phi/)                             |
+| Falcon            | [`llm/falcon/`](llm/falcon/)                       |
+| Cogito            | [`llm/cogito/`](llm/cogito/)                       |
+| CogVLM            | [`llm/cogvlm/`](llm/cogvlm/)                       |
+| LLaVA             | [`llm/llava/`](llm/llava/)                         |
+| LoRA              | [`llm/lora/`](llm/lora/)                           |
+| OpenAI-compatible | [`llm/openai/`](llm/openai/)                       |
+| Midnight          | [`llm/midnight/`](llm/midnight/)                   |
+| MiniMax           | [`llm/minimax/`](llm/minimax/)                     |
+| Nemotron          | [`llm/nemotron/`](llm/nemotron/)                   |
+| NSQL              | [`llm/nsql/`](llm/nsql/)                           |
+| Personaplex       | [`llm/personaplex-7b-v1/`](llm/personaplex-7b-v1/) |
+| Seed              | [`llm/seed/`](llm/seed/)                           |
+| Z-AI              | [`llm/z-ai/`](llm/z-ai/)                           |
+
+### Embeddings
+
+Embedding, reranking, and classification models across multiple serving engines.
+
+| Engine                    | Models | Path                                                                             |
+| ------------------------- | ------ | -------------------------------------------------------------------------------- |
+| BEI                       | 48     | [`embeddings/bei/`](embeddings/bei/)                                             |
+| TEI                       | 15     | [`embeddings/tei/`](embeddings/tei/)                                             |
+| CLIP                      | 1      | [`embeddings/clip/`](embeddings/clip/)                                           |
+| text-embeddings-inference | 1      | [`embeddings/text-embeddings-inference/`](embeddings/text-embeddings-inference/) |
+
+### Image
+
+Image generation, editing, upscaling, and segmentation models.
+
+| Example             | Path                                                               |
+| ------------------- | ------------------------------------------------------------------ |
+| Stable Diffusion    | [`image/stable-diffusion/`](image/stable-diffusion/)               |
+| Flux                | [`image/flux/`](image/flux/)                                       |
+| Flux Dev TRT (B200) | [`image/flux-dev-trt-b200/`](image/flux-dev-trt-b200/)             |
+| Sana                | [`image/sana/`](image/sana/)                                       |
+| ComfyUI             | [`image/comfyui/`](image/comfyui/)                                 |
+| ControlNet QR Code  | [`image/control-net-qrcode/`](image/control-net-qrcode/)           |
+| DeepFloyd XL        | [`image/deepfloyd-xl/`](image/deepfloyd-xl/)                       |
+| Fotographer         | [`image/fotographer/`](image/fotographer/)                         |
+| GFP-GAN             | [`image/gfp-gan/`](image/gfp-gan/)                                 |
+| IP-Adapter          | [`image/ip-adapter/`](image/ip-adapter/)                           |
+| Magic Animate       | [`image/magic-animate/`](image/magic-animate/)                     |
+| Playground v2       | [`image/playground-v2-aesthetic/`](image/playground-v2-aesthetic/) |
+| Segment Anything    | [`image/segment-anything/`](image/segment-anything/)               |
+| DIS Segmentation    | [`image/dis-segmentation/`](image/dis-segmentation/)               |
+| Image Segmentation  | [`image/image-segmentation/`](image/image-segmentation/)           |
+
+### Audio
+
+Speech-to-text, text-to-speech, and audio/music generation models.
+
+| Example                    | Path                                                                 |
+| -------------------------- | -------------------------------------------------------------------- |
+| Whisper                    | [`audio/whisper/`](audio/whisper/)                                   |
+| Kokoro                     | [`audio/kokoro/`](audio/kokoro/)                                     |
+| Chatterbox TTS             | [`audio/chatterbox-tts/`](audio/chatterbox-tts/)                     |
+| Piper TTS                  | [`audio/piper-tts/`](audio/piper-tts/)                               |
+| XTTS v2                    | [`audio/xtts-v2/`](audio/xtts-v2/)                                   |
+| XTTS Streaming             | [`audio/xtts-streaming/`](audio/xtts-streaming/)                     |
+| Orpheus 3B (WebSockets)    | [`audio/orpheus-3b-websockets/`](audio/orpheus-3b-websockets/)       |
+| Orpheus (Best Performance) | [`audio/orpheus-best-performance/`](audio/orpheus-best-performance/) |
+| Sesame CSM 1B              | [`audio/sesame-csm-1b/`](audio/sesame-csm-1b/)                       |
+| MetaVoice 1B               | [`audio/metavoice-1b/`](audio/metavoice-1b/)                         |
+| Ultravox                   | [`audio/ultravox/`](audio/ultravox/)                                 |
+| AudioGen Medium            | [`audio/audiogen-medium/`](audio/audiogen-medium/)                   |
+| MusicGen Large             | [`audio/musicgen-large/`](audio/musicgen-large/)                     |
+| MusicGen Melody            | [`audio/musicgen-melody/`](audio/musicgen-melody/)                   |
+| NVIDIA Parakeet            | [`audio/nvidia-parakeet/`](audio/nvidia-parakeet/)                   |
+| Qwen ASR                   | [`audio/qwen-asr/`](audio/qwen-asr/)                                 |
+| Qwen Omni                  | [`audio/qwen-omni/`](audio/qwen-omni/)                               |
+| Qwen Omni Thinker          | [`audio/qwen-omni-thinker/`](audio/qwen-omni-thinker/)               |
+| Voxtral Streaming 4B       | [`audio/voxtral-streaming-4b/`](audio/voxtral-streaming-4b/)         |
+
+### Optimized
+
+Production-grade, autogenerated model configurations using TRT-LLM and other optimization engines.
+
+| Engine           | Models | Path                                     |
+| ---------------- | ------ | ---------------------------------------- |
+| Briton (TRT-LLM) | 33     | [`optimized/briton/`](optimized/briton/) |
+| BISv2            | 11     | [`optimized/bisv2/`](optimized/bisv2/)   |
+
+### Infrastructure
+
+Patterns, techniques, and advanced serving configurations.
+
+| Example                       | Path                                                                                             |
+| ----------------------------- | ------------------------------------------------------------------------------------------------ |
+| Custom Server                 | [`infrastructure/custom-server/`](infrastructure/custom-server/)                                 |
+| gRPC                          | [`infrastructure/grpc/`](infrastructure/grpc/)                                                   |
+| Chains                        | [`infrastructure/chains-examples/`](infrastructure/chains-examples/)                             |
+| Multiprocessing               | [`infrastructure/multiprocessing/`](infrastructure/multiprocessing/)                             |
+| Metrics                       | [`infrastructure/metrics/`](infrastructure/metrics/)                                             |
+| Model Cache                   | [`infrastructure/model-cache/`](infrastructure/model-cache/)                                     |
+| Custom Engine Builder Control | [`infrastructure/custom-engine-builder-control/`](infrastructure/custom-engine-builder-control/) |
+| LLama.cpp Server              | [`infrastructure/llama-cpp-server/`](infrastructure/llama-cpp-server/)                           |
+| JSON Formatter                | [`infrastructure/jsonformatter/`](infrastructure/jsonformatter/)                                 |
+| LayoutLM Document QA          | [`infrastructure/layoutlm-document-qa/`](infrastructure/layoutlm-document-qa/)                   |
+| N-gram Speculator             | [`infrastructure/ngram-speculator/`](infrastructure/ngram-speculator/)                           |
+| PaddlePaddle                  | [`infrastructure/paddlepaddle/`](infrastructure/paddlepaddle/)                                   |
+| Autodesk WALA                 | [`infrastructure/autodesk-wala/`](infrastructure/autodesk-wala/)                                 |
+| Binocular                     | [`infrastructure/binocular/`](infrastructure/binocular/)                                         |
+
+## Contributing
+
+We welcome new models and improvements to existing examples. See [CONTRIBUTING.md](CONTRIBUTING.md) for details.
diff --git a/11-embeddings-reranker-classification-tensorrt/README.md b/_internal/11-trt-original-README.md
similarity index 100%
rename from 11-embeddings-reranker-classification-tensorrt/README.md
rename to _internal/11-trt-original-README.md
diff --git a/_internal/README.md b/_internal/README.md
new file mode 100644
index 000000000..ec6a22ebf
--- /dev/null
+++ b/_internal/README.md
@@ -0,0 +1,116 @@
+# _internal/
+
+Development tooling, CI infrastructure, and codegen scripts for the truss-examples repository. Nothing in this directory is a user-facing example.
+
+## Directory layout
+
+```
+_internal/
+├── bin/                        # CLI scripts for testing, validation, and README generation
+├── templates/                  # Legacy TRT-LLM / Whisper codegen templates
+├── templating/                 # Programmatic config generation for BEI, TEI, Briton, BISV2
+├── baseten-inference-stack-v2-templates/  # BIS V2 model configs (DeepSeek, Llama 4, Qwen, etc.)
+├── trt-llm-engine-builder-templates/     # TRT-LLM engine builder configs (Llama 3.1 variants)
+├── assets/                     # Static assets (ComfyUI screenshots, workflow JSON)
+├── dockerfiles/                # Custom Dockerfiles (ComfyUI)
+├── essential/                  # Config for Essential AI model (vLLM-based)
+├── internal/                   # Config for internal Briton speculative decoding test
+└── validation_report.json      # Output from validate_all.py
+```
+
+## Scripts
+
+### bin/
+
+| Script | Description |
+|--------|-------------|
+| `test_all.py` | Comprehensive local test suite: config validation, README checks, naming conventions, link validation, CI completeness |
+| `validate_all.py` | Walks every config.yaml and produces a structured validation report (markdown, JSON, or CSV) |
+| `discover_examples.py` | Auto-discovers all testable example directories and outputs a JSON array; used by CI to generate the test matrix |
+| `test_example.py` | CI script: detects changed model from git diff, pushes to staging, runs inference, cleans up old deployments |
+| `test_truss_deploy.py` | Deploys a truss to staging via `truss push`, invokes inference with `example_model_input`, deactivates after |
+| `generate_readmes.py` | Auto-generates standardized README.md files for every non-archived example from config.yaml metadata |
+| `image.txt` | Base64-encoded test image used by `test_example.py` for image model inference |
+
+### templating/
+
+| Script | Description |
+|--------|-------------|
+| `generate_templates.py` | Programmatically generates truss configs for BEI, TEI, Briton, and BIS V2 deployments from model definitions |
+| `deploy_all.py` | Bulk deploy/delete/test operations for generated templates (supports `--action deploy\|delete\|britontest` with `--filter`) |
+
+### templates/
+
+| File | Description |
+|------|-------------|
+| `generate.py` | Generates truss directories from base templates + config overrides defined in `generate.yaml` |
+| `generate.yaml` | Declares legacy TRT-LLM and Transformers model variants (Mistral, Mixtral, Llama 2, Zephyr, Whisper) |
+| `faster-whisper-truss/` | Base template for Faster Whisper models |
+| `transformers-openai-compatible/` | Base template for HuggingFace Transformers with OpenAI-compatible API |
+| `trt-llm/` | Base template for TRT-LLM models (includes Triton server configs) |
+
+## Running the test suite
+
+All scripts assume you are in the repository root.
+
+```sh
+# Run all local validation checks
+python _internal/bin/test_all.py
+
+# Verbose output (prints passing tests too)
+python _internal/bin/test_all.py --verbose
+
+# Filter to a single category
+python _internal/bin/test_all.py --category llm
+```
+
+The test suite runs 11 checks:
+
+1. Config validation (YAML parse + `truss.load()`)
+2. README existence for every example
+3. README-to-config consistency (endpoints, secrets, tags)
+4. Directory naming conventions (hyphens, no underscores)
+5. README link/path validation
+6. `example_model_input` format validation
+7. Requirements version pinning
+8. CI excludes validation against `ci_excludes.yaml`
+9. TRT-LLM OpenAI-compatible tag presence
+10. `model.py` syntax validation (AST parse)
+11. Discovery sanity check (auto-discovery finds >100 examples)
+
+### Other validation commands
+
+```sh
+# Structured validation report (writes validation_report.json)
+python _internal/bin/validate_all.py
+python _internal/bin/validate_all.py --json
+python _internal/bin/validate_all.py --csv
+
+# List all auto-discovered examples
+python _internal/bin/discover_examples.py --pretty
+```
+
+### README generation
+
+```sh
+# Regenerate all example READMEs from config.yaml metadata
+python _internal/bin/generate_readmes.py
+```
+
+### Template generation (legacy)
+
+```sh
+cd _internal/templates
+python generate.py --root ../.. --templates . --config generate.yaml
+
+# Check-only mode (fails if generated output would differ)
+python generate.py --only_check --root ../.. --templates . --config generate.yaml
+```
+
+## Dependencies
+
+- **truss** -- required by all scripts. Install with `uv pip install -e ../truss` or `pip install truss`.
+- **PyYAML** -- used for config parsing (typically installed with truss).
+- `templating/generate_templates.py` additionally requires `transformers`, `pydantic`, and `requests`.
+- `templating/deploy_all.py` requires `BASETEN_API_KEY` environment variable and optionally `openai`.
+- `templates/generate.py` requires `jinja2` and `pydantic`.
diff --git a/assets/comfyui-screenshot-1.png b/_internal/assets/comfyui-screenshot-1.png
similarity index 100%
rename from assets/comfyui-screenshot-1.png
rename to _internal/assets/comfyui-screenshot-1.png
diff --git a/assets/comfyui-screenshot-2.png b/_internal/assets/comfyui-screenshot-2.png
similarity index 100%
rename from assets/comfyui-screenshot-2.png
rename to _internal/assets/comfyui-screenshot-2.png
diff --git a/assets/comfyui-screenshot-3.png b/_internal/assets/comfyui-screenshot-3.png
similarity index 100%
rename from assets/comfyui-screenshot-3.png
rename to _internal/assets/comfyui-screenshot-3.png
diff --git a/assets/sdxl-controlnet-workflow.json b/_internal/assets/sdxl-controlnet-workflow.json
similarity index 100%
rename from assets/sdxl-controlnet-workflow.json
rename to _internal/assets/sdxl-controlnet-workflow.json
diff --git a/baseten-inference-stack-v2-templates/deepseek-v3-0324/README.md b/_internal/baseten-inference-stack-v2-templates/deepseek-v3-0324/README.md
similarity index 100%
rename from baseten-inference-stack-v2-templates/deepseek-v3-0324/README.md
rename to _internal/baseten-inference-stack-v2-templates/deepseek-v3-0324/README.md
diff --git a/baseten-inference-stack-v2-templates/deepseek-v3-0324/config.yaml b/_internal/baseten-inference-stack-v2-templates/deepseek-v3-0324/config.yaml
similarity index 100%
rename from baseten-inference-stack-v2-templates/deepseek-v3-0324/config.yaml
rename to _internal/baseten-inference-stack-v2-templates/deepseek-v3-0324/config.yaml
diff --git a/baseten-inference-stack-v2-templates/glm47/config.yaml b/_internal/baseten-inference-stack-v2-templates/glm47/config.yaml
similarity index 100%
rename from baseten-inference-stack-v2-templates/glm47/config.yaml
rename to _internal/baseten-inference-stack-v2-templates/glm47/config.yaml
diff --git a/baseten-inference-stack-v2-templates/gpt-oss-120b/README.md b/_internal/baseten-inference-stack-v2-templates/gpt-oss-120b/README.md
similarity index 100%
rename from baseten-inference-stack-v2-templates/gpt-oss-120b/README.md
rename to _internal/baseten-inference-stack-v2-templates/gpt-oss-120b/README.md
diff --git a/baseten-inference-stack-v2-templates/gpt-oss-120b/config.yaml b/_internal/baseten-inference-stack-v2-templates/gpt-oss-120b/config.yaml
similarity index 100%
rename from baseten-inference-stack-v2-templates/gpt-oss-120b/config.yaml
rename to _internal/baseten-inference-stack-v2-templates/gpt-oss-120b/config.yaml
diff --git a/baseten-inference-stack-v2-templates/kimi-k2-instruct/README.md b/_internal/baseten-inference-stack-v2-templates/kimi-k2-instruct/README.md
similarity index 100%
rename from baseten-inference-stack-v2-templates/kimi-k2-instruct/README.md
rename to _internal/baseten-inference-stack-v2-templates/kimi-k2-instruct/README.md
diff --git a/baseten-inference-stack-v2-templates/kimi-k2-instruct/config.yaml b/_internal/baseten-inference-stack-v2-templates/kimi-k2-instruct/config.yaml
similarity index 100%
rename from baseten-inference-stack-v2-templates/kimi-k2-instruct/config.yaml
rename to _internal/baseten-inference-stack-v2-templates/kimi-k2-instruct/config.yaml
diff --git a/baseten-inference-stack-v2-templates/llama-4-maverick/config.yaml b/_internal/baseten-inference-stack-v2-templates/llama-4-maverick/config.yaml
similarity index 100%
rename from baseten-inference-stack-v2-templates/llama-4-maverick/config.yaml
rename to _internal/baseten-inference-stack-v2-templates/llama-4-maverick/config.yaml
diff --git a/baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/README.md b/_internal/baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/README.md
similarity index 100%
rename from baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/README.md
rename to _internal/baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/README.md
diff --git a/baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/config.yaml b/_internal/baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/config.yaml
similarity index 100%
rename from baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/config.yaml
rename to _internal/baseten-inference-stack-v2-templates/qwen3-Coder-30b-instruct/config.yaml
diff --git a/_internal/bin/discover_examples.py b/_internal/bin/discover_examples.py
new file mode 100644
index 000000000..a4dcdad5f
--- /dev/null
+++ b/_internal/bin/discover_examples.py
@@ -0,0 +1,65 @@
+#!/usr/bin/env python3
+"""Auto-discover all testable example directories in the repository.
+
+Walks the repo looking for directories containing a config.yaml,
+skipping internal/archived paths and any paths listed in ci_excludes.yaml.
+
+Outputs a sorted JSON array of relative paths to stdout.
+Used by the GitHub Actions workflow to generate the CI matrix.
+
+Usage:
+    python _internal/bin/discover_examples.py           # compact JSON
+    python _internal/bin/discover_examples.py --pretty   # indented JSON
+"""
+
+import json
+import os
+import sys
+from pathlib import Path
+
+import yaml
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+
+SKIP_DIRS = {".git", ".github", "__pycache__", "node_modules", ".venv", "venv"}
+
+SKIP_PREFIXES = ("_archive", "_internal")
+
+
+def load_excludes(repo_root: Path) -> set[str]:
+    """Load excluded paths from ci_excludes.yaml if it exists."""
+    excludes_path = repo_root / "ci_excludes.yaml"
+    if not excludes_path.exists():
+        return set()
+    with open(excludes_path) as f:
+        data = yaml.safe_load(f)
+    if not data or not isinstance(data, dict):
+        return set()
+    return set(data.get("exclude", []) or [])
+
+
+def discover_examples(repo_root: Path) -> list[str]:
+    """Find all directories containing a config.yaml, minus excludes."""
+    excludes = load_excludes(repo_root)
+    examples = []
+    for dirpath, dirnames, filenames in os.walk(repo_root):
+        dirnames[:] = [d for d in dirnames if d not in SKIP_DIRS]
+        if "config.yaml" in filenames:
+            rel = str(Path(dirpath).relative_to(repo_root))
+            if any(rel.startswith(prefix) for prefix in SKIP_PREFIXES):
+                continue
+            if rel in excludes:
+                continue
+            examples.append(rel)
+    return sorted(examples)
+
+
+def main():
+    pretty = "--pretty" in sys.argv
+    examples = discover_examples(REPO_ROOT)
+    indent = 2 if pretty else None
+    print(json.dumps(examples, indent=indent))
+
+
+if __name__ == "__main__":
+    main()
diff --git a/_internal/bin/generate_readmes.py b/_internal/bin/generate_readmes.py
new file mode 100644
index 000000000..14dc36d9d
--- /dev/null
+++ b/_internal/bin/generate_readmes.py
@@ -0,0 +1,680 @@
+#!/usr/bin/env python3
+"""Generate standardized README.md for every non-archived example directory."""
+
+from __future__ import annotations
+
+import json
+import os
+import re
+from pathlib import Path
+
+import yaml
+
+REPO_ROOT = Path(__file__).resolve().parents[2]
+SKIP_DIRS = {"_archive", "_internal", ".git", ".github", ".venv", "__pycache__"}
+
+
+# ---------------------------------------------------------------------------
+# Metadata extraction helpers
+# ---------------------------------------------------------------------------
+
+
+def extract_hf_id(config: dict) -> str | None:
+    # 1. model_metadata.repo_id
+    repo_id = config.get("model_metadata", {}).get("repo_id")
+    if repo_id:
+        return repo_id
+
+    # 2. trt_llm checkpoint repo
+    trt = config.get("trt_llm", {})
+    repo = trt.get("build", {}).get("checkpoint_repository", {}).get("repo")
+    if repo:
+        return repo
+
+    # 3. model_cache first entry
+    caches = config.get("model_cache", [])
+    if caches and isinstance(caches, list):
+        first = caches[0] if caches else {}
+        rid = first.get("repo_id")
+        if rid:
+            return rid
+
+    # 4. Parse from docker_server.start_command
+    cmd = config.get("docker_server", {}).get("start_command", "")
+    for token in cmd.split():
+        if "/" in token and not token.startswith("-") and not token.startswith("/"):
+            # Looks like org/model — strip any trailing punctuation
+            cleaned = token.strip("\"'")
+            if re.match(r"^[\w.-]+/[\w.-]+", cleaned):
+                return cleaned
+
+    return None
+
+
+def detect_engine(config: dict) -> str:
+    base_image = str(
+        config.get("base_image", {}).get("image", "")
+        if isinstance(config.get("base_image"), dict)
+        else ""
+    )
+    start_cmd = config.get("docker_server", {}).get("start_command", "")
+    requirements = [str(r) for r in config.get("requirements", [])]
+    req_str = " ".join(requirements)
+
+    if config.get("trt_llm"):
+        build = config["trt_llm"].get("build", {})
+        base_model = build.get("base_model", "")
+        if base_model in ("encoder", "encoder_bert"):
+            return "BEI (TensorRT)"
+        return "TRT-LLM"
+    if "vllm" in base_image.lower() or "vllm" in start_cmd.lower():
+        return "vLLM"
+    if "sglang" in start_cmd.lower() or "sglang" in req_str.lower():
+        return "SGLang"
+    if (
+        "text-embeddings" in base_image.lower()
+        or "text-embeddings-router" in start_cmd.lower()
+    ):
+        return "TEI (HuggingFace)"
+    if config.get("docker_server"):
+        return "Docker Server"
+    return "Custom (Truss)"
+
+
+def infer_task_type(config: dict, category: str, dir_name: str) -> str:
+    tags = config.get("model_metadata", {}).get("tags", [])
+    if not isinstance(tags, list):
+        tags = []
+
+    # Check tags first
+    for tag in tags:
+        tag_l = tag.lower()
+        if "text-to-speech" in tag_l or "tts" in tag_l:
+            return "Text-to-speech"
+        if "speech-to-text" in tag_l or "stt" in tag_l or "transcription" in tag_l:
+            return "Speech-to-text"
+        if "image-generation" in tag_l:
+            return "Image generation"
+        if "embedding" in tag_l:
+            return "Embeddings"
+        if "rerank" in tag_l:
+            return "Reranking"
+        if "classification" in tag_l:
+            return "Classification"
+
+    # Infer from category path
+    cat_l = category.lower()
+    if "llm" in cat_l or "optimized" in cat_l:
+        return "Text generation"
+    if "embedding" in cat_l:
+        # Check if it's a reranker
+        if "rerank" in dir_name.lower():
+            return "Reranking"
+        if (
+            "classification" in dir_name.lower()
+            or "reward" in dir_name.lower()
+            or "ner" in dir_name.lower()
+        ):
+            return "Classification"
+        return "Embeddings"
+    if "image" in cat_l:
+        return "Image generation"
+    if "audio" in cat_l:
+        if (
+            "tts" in dir_name.lower()
+            or "voice" in dir_name.lower()
+            or "speech" in dir_name.lower()
+            or "kokoro" in dir_name.lower()
+            or "chatterbox" in dir_name.lower()
+            or "metavoice" in dir_name.lower()
+            or "sesame" in dir_name.lower()
+        ):
+            return "Text-to-speech"
+        if "whisper" in dir_name.lower() or "transcri" in dir_name.lower():
+            return "Speech-to-text"
+        if "music" in dir_name.lower() or "audiogen" in dir_name.lower():
+            return "Audio generation"
+        if "ultravox" in dir_name.lower():
+            return "Audio understanding"
+        return "Audio"
+    if "infrastructure" in cat_l:
+        return "Infrastructure / Custom server"
+    if "tutorial" in cat_l:
+        return "Tutorial"
+    return "ML inference"
+
+
+def extract_quantization(config: dict, dir_name: str) -> str | None:
+    # From config
+    trt = config.get("trt_llm", {})
+    qt = trt.get("build", {}).get("quantization_type")
+    if qt:
+        return qt.upper().replace("_", " ")
+
+    # From directory name
+    name_l = dir_name.lower()
+    for pat, label in [
+        ("fp4", "FP4"),
+        ("fp8", "FP8"),
+        ("int8", "INT8"),
+        ("int4", "INT4"),
+        ("awq", "AWQ"),
+        ("gptq", "GPTQ"),
+        ("bnb", "BnB 4-bit"),
+    ]:
+        if pat in name_l:
+            return label
+    return None
+
+
+def extract_gpu(config: dict) -> str:
+    acc = config.get("resources", {}).get("accelerator")
+    if acc:
+        return str(acc)
+    if config.get("resources", {}).get("use_gpu"):
+        return "GPU (unspecified)"
+    return "CPU"
+
+
+def requires_hf_token(config: dict) -> bool:
+    secrets = config.get("secrets", {})
+    if isinstance(secrets, dict) and "hf_access_token" in secrets:
+        return True
+    env = config.get("environment_variables", {})
+    if isinstance(env, dict) and "hf_access_token" in env:
+        return True
+    return False
+
+
+def get_api_endpoint(config: dict) -> str:
+    ep = config.get("docker_server", {}).get("predict_endpoint")
+    if ep:
+        return ep
+    rt = config.get("trt_llm", {}).get("runtime", {})
+    route = rt.get("webserver_default_route")
+    if route:
+        return route
+    # OpenAI-compatible TRT-LLM defaults
+    tags = config.get("model_metadata", {}).get("tags", [])
+    if isinstance(tags, list) and "openai-compatible" in tags:
+        return "/v1/chat/completions"
+    return "/predict"
+
+
+def is_openai_compatible(config: dict) -> bool:
+    tags = config.get("model_metadata", {}).get("tags", [])
+    if not isinstance(tags, list):
+        return False
+    return "openai-compatible" in tags
+
+
+def get_category(example_dir: Path) -> str:
+    """Return the top-level category (llm, embeddings, image, audio, etc.)."""
+    rel = example_dir.relative_to(REPO_ROOT)
+    parts = rel.parts
+    if len(parts) >= 1:
+        return parts[0]
+    return "unknown"
+
+
+def get_subcategory(example_dir: Path) -> str:
+    """Return deeper subcategory path for richer context."""
+    rel = example_dir.relative_to(REPO_ROOT)
+    parts = rel.parts
+    if len(parts) >= 2:
+        return "/".join(parts[:2])
+    return parts[0] if parts else "unknown"
+
+
+# ---------------------------------------------------------------------------
+# Config highlights
+# ---------------------------------------------------------------------------
+
+
+def build_config_highlights(config: dict, engine: str) -> list[str]:
+    highlights = []
+
+    # Quantization
+    trt = config.get("trt_llm", {})
+    build = trt.get("build", {})
+    qt = build.get("quantization_type")
+    if qt:
+        highlights.append(f"Quantization: **{qt}**")
+
+    # Tensor parallelism
+    tp = build.get("tensor_parallel_count")
+    if tp and tp > 1:
+        highlights.append(f"Tensor parallelism: **{tp}** GPUs")
+
+    # Speculative decoding
+    spec = build.get("speculator", {})
+    if spec:
+        mode = spec.get("speculative_decoding_mode", "enabled")
+        highlights.append(f"Speculative decoding: **{mode}**")
+
+    # Max sequence length
+    max_seq = build.get("max_seq_len")
+    if max_seq:
+        highlights.append(f"Max sequence length: **{max_seq:,}**")
+
+    # Chunked context
+    if trt.get("runtime", {}).get("enable_chunked_context"):
+        highlights.append("Chunked context: **enabled**")
+
+    # Batch scheduler policy
+    bsp = trt.get("runtime", {}).get("batch_scheduler_policy")
+    if bsp:
+        highlights.append(f"Batch scheduler policy: **{bsp}**")
+
+    # Plugin configuration
+    plugins = build.get("plugin_configuration", {})
+    if plugins:
+        for k, v in plugins.items():
+            if v:
+                highlights.append(f"Plugin: **{k}**")
+
+    # Custom base image
+    base_image = config.get("base_image", {})
+    if isinstance(base_image, dict) and base_image.get("image"):
+        highlights.append(f"Base image: `{base_image['image']}`")
+
+    # Model cache / volume mounting
+    caches = config.get("model_cache", [])
+    if caches and isinstance(caches, list):
+        for c in caches:
+            if c.get("use_volume"):
+                highlights.append(
+                    "Model cache: **volume-mounted** for fast cold starts"
+                )
+                break
+
+    # Concurrency
+    conc = config.get("runtime", {}).get("predict_concurrency")
+    if conc:
+        highlights.append(f"Predict concurrency: **{conc}**")
+
+    # System packages
+    sys_pkgs = config.get("system_packages", [])
+    if sys_pkgs:
+        highlights.append(f"System packages: `{', '.join(sys_pkgs)}`")
+
+    # Streaming via example input
+    example = config.get("model_metadata", {}).get("example_model_input", {})
+    if isinstance(example, dict) and example.get("stream"):
+        highlights.append("Streaming: **enabled**")
+
+    # Environment variables (non-secret)
+    envs = config.get("environment_variables", {})
+    if isinstance(envs, dict):
+        notable = {
+            k: v for k, v in envs.items() if k != "hf_access_token" and v is not None
+        }
+        if notable:
+            highlights.append(
+                f"Environment variables: {', '.join(f'`{k}`' for k in notable)}"
+            )
+
+    if not highlights:
+        highlights.append(f"Engine: **{engine}**")
+
+    return highlights
+
+
+# ---------------------------------------------------------------------------
+# Invoke section
+# ---------------------------------------------------------------------------
+
+
+def build_invoke_section(
+    config: dict,
+    engine: str,
+    task: str,
+    endpoint: str,
+    hf_id: str | None,
+    openai_compat: bool,
+) -> str:
+    example_input = config.get("model_metadata", {}).get("example_model_input")
+
+    # --- OpenAI-compatible LLM ---
+    if openai_compat and task == "Text generation":
+        model_name = hf_id or "model"
+        return f"""\
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="{model_name}",
+    messages=[{{"role": "user", "content": "What is machine learning?"}}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{{"model": "{model_name}", "messages": [{{"role": "user", "content": "What is machine learning?"}}], "max_tokens": 512}}'
+```"""
+
+    # --- BEI / TEI embeddings ---
+    if endpoint == "/v1/embeddings":
+        model_name = hf_id or "model"
+        return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{{"input": "What is deep learning?", "model": "{model_name}"}}'
+```"""
+
+    # --- Reranker ---
+    if endpoint == "/rerank" or task == "Reranking":
+        return """\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```"""
+
+    # --- Use example_model_input if available ---
+    if example_input:
+        if isinstance(example_input, str):
+            try:
+                formatted = example_input
+            except Exception:
+                formatted = example_input
+            return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co{endpoint} \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{formatted}'
+```"""
+        else:
+            formatted = json.dumps(example_input, indent=2)
+            return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co{endpoint} \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{formatted}'
+```"""
+
+    # --- Fallback by task ---
+    if task == "Text generation":
+        return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co{endpoint} \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{{"prompt": "What is machine learning?", "max_tokens": 512}}'
+```"""
+
+    if task == "Image generation":
+        return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co{endpoint} \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{{"prompt": "A photo of a cat in a field of sunflowers"}}'
+```
+
+> The response may contain base64-encoded image data."""
+
+    if task in ("Text-to-speech", "Audio generation"):
+        return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co{endpoint} \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{{"text": "Hello, this is a test of text to speech."}}'
+```"""
+
+    if task == "Speech-to-text":
+        return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co{endpoint} \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{{"url": "https://example.com/audio.wav"}}'
+```"""
+
+    # Generic fallback
+    return f"""\
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co{endpoint} \\
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \\
+  -H "Content-Type: application/json" \\
+  -d '{{}}'
+```"""
+
+
+# ---------------------------------------------------------------------------
+# Description generation
+# ---------------------------------------------------------------------------
+
+
+def build_description(
+    model_name: str,
+    config: dict,
+    category: str,
+    engine: str,
+    hf_id: str | None,
+    task: str,
+) -> str:
+    desc = config.get("description")
+    if desc:
+        return desc
+
+    hf_part = f"[{hf_id}](https://huggingface.co/{hf_id})" if hf_id else model_name
+    engine_article = "an" if engine[0] in "AEIOU" else "a"
+
+    if task == "Text generation":
+        return f"Deploy {hf_part} for text generation using {engine_article} {engine} engine on Baseten."
+    if task == "Embeddings":
+        return f"Deploy {hf_part} for generating text embeddings using {engine_article} {engine} engine on Baseten."
+    if task == "Reranking":
+        return f"Deploy {hf_part} as a reranker using {engine_article} {engine} engine on Baseten."
+    if task == "Classification":
+        return f"Deploy {hf_part} for classification using {engine_article} {engine} engine on Baseten."
+    if task == "Image generation":
+        return f"Deploy {hf_part} for image generation on Baseten."
+    if task == "Text-to-speech":
+        return f"Deploy {hf_part} for text-to-speech on Baseten."
+    if task == "Speech-to-text":
+        return f"Deploy {hf_part} for speech-to-text transcription on Baseten."
+    if task == "Audio generation":
+        return f"Deploy {hf_part} for audio generation on Baseten."
+    if task == "Audio understanding":
+        return f"Deploy {hf_part} for audio understanding on Baseten."
+    if task == "Tutorial":
+        return f"A tutorial example showing how to deploy {model_name} on Baseten."
+    if task == "Infrastructure / Custom server":
+        return f"Deploy {hf_part} using a custom server configuration on Baseten."
+
+    return f"Deploy {model_name} on Baseten using {engine_article} {engine} engine."
+
+
+# ---------------------------------------------------------------------------
+# README rendering
+# ---------------------------------------------------------------------------
+
+
+def render_readme(
+    model_name: str,
+    description: str,
+    hf_id: str | None,
+    task: str,
+    engine: str,
+    gpu: str,
+    quantization: str | None,
+    openai_compat: bool,
+    hf_token: bool,
+    endpoint: str,
+    highlights: list[str],
+    invoke_section: str,
+    has_model_py: bool,
+    python_version: str | None,
+) -> str:
+    lines = []
+    lines.append(f"# {model_name}\n")
+    lines.append(f"{description}\n")
+
+    # Properties table
+    lines.append("| Property | Value |")
+    lines.append("|----------|-------|")
+    if hf_id:
+        lines.append(f"| Model | [{hf_id}](https://huggingface.co/{hf_id}) |")
+    lines.append(f"| Task | {task} |")
+    lines.append(f"| Engine | {engine} |")
+    lines.append(f"| GPU | {gpu} |")
+    if quantization:
+        lines.append(f"| Quantization | {quantization} |")
+    if openai_compat:
+        lines.append("| OpenAI compatible | Yes |")
+    if python_version:
+        lines.append(f"| Python | {python_version} |")
+    lines.append("")
+
+    # Deploy
+    lines.append("## Deploy\n")
+    if hf_token:
+        lines.append(
+            "> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.\n"
+        )
+    lines.append("```sh\ntruss push\n```\n")
+
+    # Invoke
+    lines.append("## Invoke\n")
+    lines.append(invoke_section)
+    lines.append("")
+
+    # Config highlights
+    lines.append("## Configuration highlights\n")
+    for h in highlights:
+        lines.append(f"- {h}")
+    lines.append("")
+
+    return "\n".join(lines)
+
+
+# ---------------------------------------------------------------------------
+# Main
+# ---------------------------------------------------------------------------
+
+
+def find_example_dirs() -> list[Path]:
+    """Find all directories containing config.yaml, excluding archive/internal."""
+    examples = []
+    for root, dirs, files in os.walk(REPO_ROOT):
+        # Prune skip dirs
+        dirs[:] = [d for d in dirs if d not in SKIP_DIRS]
+        if "config.yaml" in files:
+            examples.append(Path(root))
+    return sorted(examples)
+
+
+def process_example(example_dir: Path) -> dict:
+    """Process a single example directory and return metadata + generated README."""
+    config_path = example_dir / "config.yaml"
+    with open(config_path) as f:
+        config = yaml.safe_load(f) or {}
+
+    dir_name = example_dir.name
+    category = get_category(example_dir)
+
+    model_name = (
+        config.get("model_name") or dir_name.replace("-", " ").replace("_", " ").title()
+    )
+    hf_id = extract_hf_id(config)
+    engine = detect_engine(config)
+    task = infer_task_type(config, category, dir_name)
+    gpu = extract_gpu(config)
+    quantization = extract_quantization(config, dir_name)
+    openai_compat = is_openai_compatible(config)
+    hf_token = requires_hf_token(config)
+    endpoint = get_api_endpoint(config)
+    python_version = config.get("python_version")
+    has_model_py = (example_dir / "model" / "model.py").exists()
+    highlights = build_config_highlights(config, engine)
+
+    description = build_description(model_name, config, category, engine, hf_id, task)
+    invoke = build_invoke_section(config, engine, task, endpoint, hf_id, openai_compat)
+
+    readme = render_readme(
+        model_name=model_name,
+        description=description,
+        hf_id=hf_id,
+        task=task,
+        engine=engine,
+        gpu=gpu,
+        quantization=quantization,
+        openai_compat=openai_compat,
+        hf_token=hf_token,
+        endpoint=endpoint,
+        highlights=highlights,
+        invoke_section=invoke,
+        has_model_py=has_model_py,
+        python_version=python_version,
+    )
+
+    return {
+        "dir": str(example_dir),
+        "model_name": model_name,
+        "hf_id": hf_id,
+        "engine": engine,
+        "task": task,
+        "readme": readme,
+    }
+
+
+def main():
+    examples = find_example_dirs()
+    print(f"Found {len(examples)} example directories\n")
+
+    generated = 0
+    missing_hf = []
+
+    for example_dir in examples:
+        try:
+            result = process_example(example_dir)
+        except Exception as e:
+            print(f"  ERROR: {example_dir.relative_to(REPO_ROOT)}: {e}")
+            continue
+
+        readme_path = example_dir / "README.md"
+        readme_path.write_text(result["readme"])
+        generated += 1
+
+        rel = example_dir.relative_to(REPO_ROOT)
+        status = "ok" if result["hf_id"] else "no HF ID"
+        print(f"  {rel} [{result['engine']}] — {status}")
+
+        if not result["hf_id"]:
+            missing_hf.append(str(rel))
+
+    print("\n--- Summary ---")
+    print(f"Generated: {generated}/{len(examples)}")
+    print(f"Missing HuggingFace ID: {len(missing_hf)}")
+    if missing_hf:
+        print("Directories without HF ID:")
+        for d in missing_hf:
+            print(f"  - {d}")
+
+
+if __name__ == "__main__":
+    main()
diff --git a/bin/image.txt b/_internal/bin/image.txt
similarity index 100%
rename from bin/image.txt
rename to _internal/bin/image.txt
diff --git a/_internal/bin/test_all.py b/_internal/bin/test_all.py
new file mode 100644
index 000000000..5a41a628b
--- /dev/null
+++ b/_internal/bin/test_all.py
@@ -0,0 +1,707 @@
+#!/usr/bin/env python3
+"""Comprehensive test suite for truss-examples repository.
+
+Runs all validation that can be done locally without deploying models:
+  1. Config validation (YAML parsing, truss.load())
+  2. README existence and structure
+  3. README ↔ config consistency (endpoints, secrets, model names)
+  4. Directory naming conventions
+  5. Link/path validation (README links point to real dirs)
+  6. example_model_input format validation
+  7. Requirements pinning check
+  8. CI excludes validation
+
+Usage:
+    python _internal/bin/test_all.py
+    python _internal/bin/test_all.py --verbose
+    python _internal/bin/test_all.py --category llm
+"""
+
+import os
+import re
+import sys
+from pathlib import Path
+
+import yaml
+
+try:
+    import truss
+except ImportError:
+    print(
+        "ERROR: truss not installed. Run: uv pip install -e ../truss", file=sys.stderr
+    )
+    sys.exit(1)
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+VERBOSE = "--verbose" in sys.argv or "-v" in sys.argv
+CATEGORY_FILTER = None
+for i, arg in enumerate(sys.argv):
+    if arg == "--category" and i + 1 < len(sys.argv):
+        CATEGORY_FILTER = sys.argv[i + 1]
+
+SKIP_DIRS = {
+    ".git",
+    ".github",
+    "__pycache__",
+    "node_modules",
+    ".venv",
+    "venv",
+    "templating",
+    "assets",
+    "bin",
+    "dockerfiles",
+    "packages",
+    "sample_images",
+    "examples",  # skip nested examples dirs for dir naming checks
+}
+
+CUSTOMER_CATEGORIES = [
+    "tutorials",
+    "llm",
+    "embeddings",
+    "image",
+    "audio",
+    "optimized",
+    "infrastructure",
+]
+
+# Counters
+passed = 0
+failed = 0
+warnings = 0
+failures = []
+
+
+def log_pass(msg):
+    global passed
+    passed += 1
+    if VERBOSE:
+        print(f"  PASS  {msg}")
+
+
+def log_fail(msg):
+    global failed
+    failed += 1
+    failures.append(msg)
+    print(f"  FAIL  {msg}")
+
+
+def log_warn(msg):
+    global warnings
+    warnings += 1
+    if VERBOSE:
+        print(f"  WARN  {msg}")
+
+
+def find_all_configs(root: Path) -> list[Path]:
+    """Find all directories containing a config.yaml."""
+    configs = []
+    for dirpath, dirnames, filenames in os.walk(root):
+        dirnames[:] = [
+            d
+            for d in dirnames
+            if d
+            not in {".git", ".github", "__pycache__", "node_modules", ".venv", "venv"}
+        ]
+        if "config.yaml" in filenames:
+            rel = Path(dirpath).relative_to(root)
+            # Skip _archive
+            if str(rel).startswith("_archive"):
+                continue
+            if CATEGORY_FILTER and not str(rel).startswith(CATEGORY_FILTER):
+                continue
+            configs.append(Path(dirpath))
+    return sorted(configs)
+
+
+# ─── Test 1: Config validation ───────────────────────────────────────────────
+
+
+def test_config_validation():
+    print("\n== Test 1: Config Validation (YAML + truss.load) ==")
+    configs = find_all_configs(REPO_ROOT)
+    print(f"   Found {len(configs)} configs to validate")
+
+    broken = []
+    for config_dir in configs:
+        rel = config_dir.relative_to(REPO_ROOT)
+        config_path = config_dir / "config.yaml"
+        try:
+            raw = yaml.safe_load(config_path.read_text())
+        except Exception as e:
+            log_fail(f"{rel}: YAML parse error: {e}")
+            broken.append(str(rel))
+            continue
+
+        try:
+            truss.load(str(config_dir))
+            log_pass(f"{rel}: config loads")
+        except Exception as e:
+            err = str(e)[:150]
+            # Allow known internal/archived breakages
+            if str(rel).startswith("_internal") or "/_archive/" in str(rel):
+                log_warn(f"{rel}: truss.load() failed (internal/archived): {err}")
+            else:
+                log_fail(f"{rel}: truss.load() failed: {err}")
+                broken.append(str(rel))
+
+    return broken
+
+
+# ─── Test 2: README existence ────────────────────────────────────────────────
+
+
+def test_readme_existence():
+    print("\n== Test 2: README Existence ==")
+    configs = find_all_configs(REPO_ROOT)
+    missing = []
+
+    for config_dir in configs:
+        rel = config_dir.relative_to(REPO_ROOT)
+        if (
+            str(rel).startswith("_internal")
+            or "/_archive/" in str(rel)
+            or str(rel).startswith("_archive")
+        ):
+            continue
+        # Skip _archive at any level
+        if "_archive" in rel.parts:
+            continue
+        readme = config_dir / "README.md"
+        if readme.exists():
+            log_pass(f"{rel}: has README.md")
+        else:
+            log_fail(f"{rel}: missing README.md")
+            missing.append(str(rel))
+
+    # Check category READMEs
+    for cat in CUSTOMER_CATEGORIES:
+        cat_readme = REPO_ROOT / cat / "README.md"
+        if cat_readme.exists():
+            log_pass(f"{cat}/README.md exists")
+        else:
+            log_fail(f"{cat}/README.md missing")
+
+    return missing
+
+
+# ─── Test 3: README ↔ Config consistency ─────────────────────────────────────
+
+
+def test_readme_config_consistency():
+    print("\n== Test 3: README ↔ Config Consistency ==")
+    configs = find_all_configs(REPO_ROOT)
+
+    for config_dir in configs:
+        rel = config_dir.relative_to(REPO_ROOT)
+        if str(rel).startswith("_internal"):
+            continue
+        readme_path = config_dir / "README.md"
+        config_path = config_dir / "config.yaml"
+
+        if not readme_path.exists():
+            continue
+
+        try:
+            raw = yaml.safe_load(config_path.read_text())
+        except Exception:
+            continue
+
+        if not raw or not isinstance(raw, dict):
+            continue
+
+        readme_text = readme_path.read_text()
+
+        # Check: if config has hf_access_token secret, README should mention it
+        secrets = raw.get("secrets", {})
+        if secrets and isinstance(secrets, dict) and "hf_access_token" in secrets:
+            if (
+                "hf_access_token" in readme_text
+                or "HuggingFace access token" in readme_text
+            ):
+                log_pass(f"{rel}: README mentions HF token (config requires it)")
+            else:
+                log_warn(
+                    f"{rel}: config has hf_access_token but README doesn't mention it"
+                )
+
+        # Check: endpoint consistency
+        docker_server = raw.get("docker_server", {})
+        if isinstance(docker_server, dict):
+            predict_endpoint = docker_server.get("predict_endpoint", "")
+            if predict_endpoint and predict_endpoint != "/predict":
+                # README should reference this endpoint
+                if predict_endpoint in readme_text:
+                    log_pass(f"{rel}: README uses correct endpoint {predict_endpoint}")
+                elif "/predict" in readme_text and predict_endpoint not in readme_text:
+                    log_warn(
+                        f"{rel}: README uses /predict but config has {predict_endpoint}"
+                    )
+
+        # Check: OpenAI-compatible tag should have /v1/chat/completions in README
+        tags = []
+        metadata = raw.get("model_metadata", {})
+        if isinstance(metadata, dict):
+            tags = metadata.get("tags", [])
+        if isinstance(tags, list) and "openai-compatible" in tags:
+            if "/v1/chat/completions" in readme_text or "OpenAI" in readme_text:
+                log_pass(f"{rel}: OpenAI-compatible model has correct invoke style")
+            else:
+                log_warn(
+                    f"{rel}: tagged openai-compatible but README lacks /v1/chat/completions"
+                )
+
+
+# ─── Test 4: Directory naming conventions ────────────────────────────────────
+
+
+def test_directory_naming():
+    print("\n== Test 4: Directory Naming Conventions ==")
+
+    for cat in CUSTOMER_CATEGORIES:
+        cat_dir = REPO_ROOT / cat
+        if not cat_dir.exists():
+            continue
+        for dirpath, dirnames, filenames in os.walk(cat_dir):
+            dirnames[:] = [
+                d for d in dirnames if d not in SKIP_DIRS and not d.startswith(".")
+            ]
+            rel = Path(dirpath).relative_to(REPO_ROOT)
+
+            # If this directory contains config.yaml, it's an example root.
+            # Don't check naming of its children (model/, data/, etc. are internal).
+            if "config.yaml" in filenames:
+                dirnames.clear()
+                continue
+
+            for d in dirnames:
+                if d.startswith("_"):
+                    continue  # _archive is OK
+
+                # Check underscores (allow go_emotions which is upstream model name)
+                if "_" in d and "go_emotions" not in d:
+                    log_fail(
+                        f"{rel}/{d}: directory name contains underscore (use hyphens)"
+                    )
+
+                # Check for "truss" in name
+                if (
+                    "truss" in d.lower() and d != "truss"
+                ):  # allow ngram-speculator/truss
+                    log_warn(f"{rel}/{d}: directory name contains 'truss'")
+
+                # Check PascalCase (first char uppercase suggests PascalCase)
+                if d[0].isupper() and "-" in d:
+                    log_warn(
+                        f"{rel}/{d}: directory name uses PascalCase (prefer lowercase)"
+                    )
+
+
+# ─── Test 5: Link/path validation in READMEs ────────────────────────────────
+
+
+def test_readme_links():
+    print("\n== Test 5: README Link Validation ==")
+
+    # Check root README
+    root_readme = REPO_ROOT / "README.md"
+    if root_readme.exists():
+        text = root_readme.read_text()
+        # Find all relative links like [text](path/)
+        links = re.findall(r"\[.*?\]\(([^)]+)\)", text)
+        for link in links:
+            if link.startswith("http") or link.startswith("#"):
+                continue
+            # Strip trailing /
+            clean = link.rstrip("/")
+            target = REPO_ROOT / clean
+            if target.exists() or (REPO_ROOT / clean.split("/")[0]).exists():
+                log_pass(f"README.md: link to {clean} is valid")
+            else:
+                log_fail(f"README.md: broken link to {clean}")
+
+    # Check category READMEs
+    for cat in CUSTOMER_CATEGORIES:
+        cat_readme = REPO_ROOT / cat / "README.md"
+        if not cat_readme.exists():
+            continue
+        text = cat_readme.read_text()
+        links = re.findall(r"\[.*?\]\(([^)]+)\)", text)
+        for link in links:
+            if link.startswith("http") or link.startswith("#"):
+                continue
+            clean = link.rstrip("/")
+            target = REPO_ROOT / cat / clean
+            if target.exists():
+                log_pass(f"{cat}/README.md: link to {clean} is valid")
+            else:
+                log_fail(f"{cat}/README.md: broken link to {clean}")
+
+    # Check CONTRIBUTING.md
+    contrib = REPO_ROOT / "CONTRIBUTING.md"
+    if contrib.exists():
+        text = contrib.read_text()
+        # Find paths in the table (backtick-wrapped)
+        paths = re.findall(r"`([a-z][\w/-]+)`", text)
+        for p in paths:
+            if (
+                "/" in p
+                and not p.endswith(".py")
+                and not p.endswith(".yaml")
+                and not p.endswith(".yml")
+            ):
+                target = REPO_ROOT / p
+                if target.exists():
+                    log_pass(f"CONTRIBUTING.md: path {p} exists")
+                elif not any(
+                    p.endswith(ext) for ext in [".yaml", ".py", ".md", ".json"]
+                ):
+                    # Check if it exists as a subdir within any category
+                    found = any(
+                        (REPO_ROOT / cat / p).exists() for cat in CUSTOMER_CATEGORIES
+                    )
+                    if found:
+                        log_pass(
+                            f"CONTRIBUTING.md: path {p} exists (relative to category)"
+                        )
+                    else:
+                        log_warn(f"CONTRIBUTING.md: path {p} may not exist")
+
+
+# ─── Test 6: example_model_input validation ──────────────────────────────────
+
+
+def test_example_model_input():
+    print("\n== Test 6: example_model_input Validation ==")
+    configs = find_all_configs(REPO_ROOT)
+    missing_count = 0
+
+    for config_dir in configs:
+        rel = config_dir.relative_to(REPO_ROOT)
+        if str(rel).startswith("_internal") or "_archive" in rel.parts:
+            continue
+
+        config_path = config_dir / "config.yaml"
+        try:
+            raw = yaml.safe_load(config_path.read_text())
+        except Exception:
+            continue
+
+        if not raw or not isinstance(raw, dict):
+            continue
+
+        metadata = raw.get("model_metadata", {})
+        if not isinstance(metadata, dict):
+            continue
+
+        example_input = metadata.get("example_model_input")
+        if example_input is None:
+            missing_count += 1
+            log_warn(f"{rel}: missing example_model_input")
+            continue
+
+        # Validate input format based on category
+        rel_str = str(rel)
+        if rel_str.startswith("llm/"):
+            # LLMs: prompt, messages, query, text, message, image_url, queries all valid
+            if isinstance(example_input, dict):
+                valid_keys = {
+                    "prompt",
+                    "messages",
+                    "query",
+                    "text",
+                    "message",
+                    "image_url",
+                    "queries",
+                    "model",
+                }
+                if valid_keys & set(example_input.keys()):
+                    log_pass(f"{rel}: example_model_input has valid LLM format")
+                else:
+                    log_warn(
+                        f"{rel}: LLM example_model_input missing prompt/messages/query"
+                    )
+            elif isinstance(example_input, str):
+                log_pass(f"{rel}: example_model_input is a string (direct input)")
+
+        elif rel_str.startswith("embeddings/"):
+            # Embeddings: input (standard), query/texts (rerankers), text/inputs (classifiers/NER), url (CLIP)
+            if isinstance(example_input, dict):
+                valid_keys = {
+                    "input",
+                    "inputs",
+                    "encoding_format",
+                    "query",
+                    "texts",
+                    "text",
+                    "model",
+                    "sentences",
+                    "url",
+                }
+                if valid_keys & set(example_input.keys()):
+                    log_pass(f"{rel}: example_model_input has valid embedding format")
+                else:
+                    log_warn(
+                        f"{rel}: embedding example_model_input may have wrong format"
+                    )
+            elif isinstance(example_input, str):
+                log_pass(f"{rel}: example_model_input is a string (direct input)")
+            else:
+                log_warn(f"{rel}: embedding example_model_input may have wrong format")
+
+        elif rel_str.startswith("image/"):
+            # Image: prompt, image, instances, workflow_values (comfyui), input_image, reference_image, image_url
+            if isinstance(example_input, dict):
+                valid_keys = {
+                    "prompt",
+                    "image",
+                    "instances",
+                    "workflow",
+                    "workflow_values",
+                    "url",
+                    "input_image",
+                    "reference_image",
+                    "bbox",
+                    "image_url",
+                    "text",
+                }
+                if valid_keys & set(example_input.keys()):
+                    log_pass(f"{rel}: example_model_input has valid image format")
+                else:
+                    log_warn(f"{rel}: image example_model_input may have wrong format")
+            else:
+                log_warn(f"{rel}: image example_model_input may have wrong format")
+
+        elif rel_str.startswith("audio/"):
+            if isinstance(example_input, dict):
+                log_pass(f"{rel}: example_model_input is a dict")
+            elif isinstance(example_input, str):
+                log_pass(f"{rel}: example_model_input is a string")
+
+        else:
+            log_pass(f"{rel}: has example_model_input")
+
+    if missing_count:
+        print(f"   {missing_count} configs still missing example_model_input")
+
+
+# ─── Test 7: Requirements pinning ────────────────────────────────────────────
+
+
+def test_requirements_pinning():
+    print("\n== Test 7: Requirements Pinning ==")
+    configs = find_all_configs(REPO_ROOT)
+    unpinned_count = 0
+
+    for config_dir in configs:
+        rel = config_dir.relative_to(REPO_ROOT)
+        if str(rel).startswith("_internal") or str(rel).startswith("_archive"):
+            continue
+
+        config_path = config_dir / "config.yaml"
+        try:
+            raw = yaml.safe_load(config_path.read_text())
+        except Exception:
+            continue
+
+        if not raw or not isinstance(raw, dict):
+            continue
+
+        requirements = raw.get("requirements", [])
+        if not isinstance(requirements, list):
+            continue
+
+        # Check if using requirements_file instead
+        if raw.get("requirements_file"):
+            continue
+
+        unpinned = []
+        for req in requirements:
+            if not isinstance(req, str):
+                continue
+            req = req.strip()
+            if not req or req.startswith("#") or req.startswith("-"):
+                continue
+            if req.startswith("git+"):
+                # git+ URLs should pin to a commit hash, not @master or @main
+                if "@master" in req or "@main" in req:
+                    unpinned.append(req)
+                continue
+            # Check for version pin
+            if (
+                "==" not in req
+                and ">=" not in req
+                and "<=" not in req
+                and "~=" not in req
+            ):
+                unpinned.append(req)
+
+        if unpinned:
+            unpinned_count += 1
+            for u in unpinned:
+                log_warn(f"{rel}: unpinned requirement: {u}")
+        elif requirements:
+            log_pass(f"{rel}: all requirements pinned")
+
+    if unpinned_count:
+        print(f"   {unpinned_count} configs have unpinned requirements")
+
+
+# ─── Test 8: CI excludes validation ──────────────────────────────────────────
+
+
+def test_ci_excludes():
+    print("\n== Test 8: CI Excludes Validation ==")
+    excludes_path = REPO_ROOT / "ci_excludes.yaml"
+    if not excludes_path.exists():
+        log_pass("ci_excludes.yaml not found (no excludes, OK)")
+        return
+
+    with open(excludes_path) as f:
+        data = yaml.safe_load(f)
+
+    if not data or not isinstance(data, dict):
+        log_pass("ci_excludes.yaml is empty (OK)")
+        return
+
+    excludes = data.get("exclude", []) or []
+    print(f"   {len(excludes)} excluded paths")
+
+    for path in excludes:
+        full_path = REPO_ROOT / path
+        config_path = full_path / "config.yaml"
+        if not full_path.exists():
+            log_fail(f"ci_excludes.yaml: {path} does not exist")
+        elif not config_path.exists():
+            log_fail(f"ci_excludes.yaml: {path} has no config.yaml")
+        else:
+            log_pass(f"ci_excludes.yaml: {path} is a valid example")
+
+
+# ─── Test 9: TRT-LLM openai-compatible tag ──────────────────────────────────
+
+
+def test_trt_llm_tags():
+    print("\n== Test 9: TRT-LLM openai-compatible Tag ==")
+    configs = find_all_configs(REPO_ROOT)
+    missing = 0
+
+    for config_dir in configs:
+        rel = config_dir.relative_to(REPO_ROOT)
+        if "_archive" in rel.parts or "_internal" in rel.parts:
+            continue
+
+        config_path = config_dir / "config.yaml"
+        try:
+            raw = yaml.safe_load(config_path.read_text())
+        except Exception:
+            continue
+
+        if not raw or not isinstance(raw, dict):
+            continue
+
+        if not raw.get("trt_llm"):
+            continue
+
+        tags = raw.get("model_metadata", {}).get("tags", [])
+        if not isinstance(tags, list):
+            tags = []
+
+        has_oai = "openai-compatible" in tags
+        has_legacy = "force-legacy-api-non-openai-compatible" in tags
+        if has_oai or has_legacy:
+            log_pass(f"{rel}: TRT-LLM has API compatibility tag")
+        else:
+            log_fail(f"{rel}: TRT-LLM missing openai-compatible or force-legacy tag")
+            missing += 1
+
+    if missing:
+        print(f"   {missing} TRT-LLM configs missing API compatibility tag")
+
+
+# ─── Test 10: model.py syntax validation ────────────────────────────────────
+
+
+def test_model_py_syntax():
+    print("\n== Test 10: model.py Syntax Validation ==")
+    import ast
+
+    model_files = sorted(REPO_ROOT.glob("**/model/model.py"))
+    checked = 0
+
+    for mf in model_files:
+        rel = mf.relative_to(REPO_ROOT)
+        if "_archive" in rel.parts or "_internal" in rel.parts:
+            continue
+        checked += 1
+        try:
+            ast.parse(mf.read_text())
+            log_pass(f"{rel}: syntax OK")
+        except SyntaxError as e:
+            log_fail(f"{rel}: syntax error: {e}")
+
+    print(f"   Checked {checked} model.py files")
+
+
+# ─── Test 11: Discovery sanity ──────────────────────────────────────────────
+
+
+def test_discovery_sanity():
+    print("\n== Test 11: Discovery Sanity Check ==")
+    configs = find_all_configs(REPO_ROOT)
+    # Filter same way as discovery script
+    examples = [
+        str(c.relative_to(REPO_ROOT))
+        for c in configs
+        if not str(c.relative_to(REPO_ROOT)).startswith("_internal")
+        and "_archive" not in c.relative_to(REPO_ROOT).parts
+    ]
+    count = len(examples)
+    print(f"   Auto-discovery found {count} examples")
+
+    if count > 100:
+        log_pass(f"Discovery found {count} examples (> 100 threshold)")
+    else:
+        log_fail(
+            f"Discovery found only {count} examples (expected > 100) — "
+            "discovery logic may be broken"
+        )
+
+
+# ─── Main ────────────────────────────────────────────────────────────────────
+
+
+def main():
+    print("=" * 60)
+    print("  Truss Examples - Comprehensive Test Suite")
+    print("=" * 60)
+
+    test_config_validation()
+    test_readme_existence()
+    test_readme_config_consistency()
+    test_directory_naming()
+    test_readme_links()
+    test_example_model_input()
+    test_requirements_pinning()
+    test_ci_excludes()
+    test_trt_llm_tags()
+    test_model_py_syntax()
+    test_discovery_sanity()
+
+    print("\n" + "=" * 60)
+    print(f"  Results: {passed} passed, {failed} failed, {warnings} warnings")
+    print("=" * 60)
+
+    if failures:
+        print(f"\n  {len(failures)} failure(s):")
+        for f in failures:
+            print(f"    - {f}")
+        print()
+        sys.exit(1)
+    else:
+        print("\n  All tests passed!\n")
+        sys.exit(0)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/bin/test_example.py b/_internal/bin/test_example.py
similarity index 100%
rename from bin/test_example.py
rename to _internal/bin/test_example.py
diff --git a/bin/test_truss_deploy.py b/_internal/bin/test_truss_deploy.py
similarity index 98%
rename from bin/test_truss_deploy.py
rename to _internal/bin/test_truss_deploy.py
index 33edb2938..081fadfd8 100644
--- a/bin/test_truss_deploy.py
+++ b/_internal/bin/test_truss_deploy.py
@@ -153,7 +153,7 @@ def get_time_in_ms():
 if __name__ == "__main__":
     model_dir = get_model_dir()
 
-    image_str = open("bin/image.txt", "r").read()
+    image_str = open(os.path.join(os.path.dirname(__file__), "image.txt"), "r").read()
 
     os.chdir(model_dir)
     example_input = get_example_input(image_str)
diff --git a/_internal/bin/validate_all.py b/_internal/bin/validate_all.py
new file mode 100644
index 000000000..d05f8af76
--- /dev/null
+++ b/_internal/bin/validate_all.py
@@ -0,0 +1,349 @@
+#!/usr/bin/env python3
+"""Validate every truss example config in the repository.
+
+Walks the repo, finds every directory containing a config.yaml,
+and runs validation checks. Outputs a markdown report to stdout.
+
+Usage:
+    python bin/validate_all.py [--json] [--csv]
+"""
+
+import ast
+import json
+import os
+import sys
+from pathlib import Path
+
+import yaml
+
+# Ensure we can import truss
+try:
+    import truss
+    from truss.base.truss_config import TrussConfig
+except ImportError:
+    print(
+        "ERROR: truss not installed. Run: uv pip install -e ../truss", file=sys.stderr
+    )
+    sys.exit(1)
+
+REPO_ROOT = Path(__file__).resolve().parent.parent.parent
+
+# Directories to skip entirely (not examples)
+SKIP_DIRS = {
+    ".git",
+    ".github",
+    "__pycache__",
+    "node_modules",
+    ".venv",
+    "venv",
+    "templating",
+    "assets",
+    "bin",
+    "dockerfiles",
+}
+
+# Patterns that indicate a deprecated/archivable example
+ARCHIVE_PATTERNS = [
+    "llama-2-",
+    "llama-7b",
+    "mistral-7b-chat",
+    "mistral-7b-instruct",
+    "whisper-torchserve",
+    "deepspeed-mii",
+    "nous-capybara",
+    "model_cach_gcs",
+    "qwen-7b-chat",
+]
+
+# Deprecated config fields
+DEPRECATED_FIELDS_IN_YAML = ["hf_cache"]
+
+
+class ValidationResult:
+    def __init__(self, path: str):
+        self.path = path
+        self.status = "VALID"
+        self.issues: list[str] = []
+        self.warnings: list[str] = []
+        self.has_model_py = False
+        self.has_readme = False
+        self.has_example_input = False
+        self.has_docker_server = False
+        self.has_trt_llm = False
+        self.config_loads = False
+
+    def add_issue(self, msg: str):
+        self.issues.append(msg)
+
+    def add_warning(self, msg: str):
+        self.warnings.append(msg)
+
+    def determine_status(self):
+        if not self.config_loads:
+            self.status = "BROKEN"
+            return
+
+        # Check archive patterns
+        dir_name = Path(self.path).name.lower()
+        for pattern in ARCHIVE_PATTERNS:
+            if pattern in dir_name:
+                self.status = "ARCHIVE"
+                return
+
+        if self.issues:
+            self.status = "DEPRECATED"
+            return
+
+        if not self.has_example_input and not self.has_readme:
+            self.status = "VALID_NO_INPUT"
+        elif not self.has_readme:
+            self.status = "VALID_NO_README"
+        elif not self.has_example_input:
+            self.status = "VALID_NO_INPUT"
+        else:
+            self.status = "VALID"
+
+
+def find_all_configs(root: Path) -> list[Path]:
+    """Find all directories containing a config.yaml."""
+    configs = []
+    for dirpath, dirnames, filenames in os.walk(root):
+        # Prune skip dirs
+        dirnames[:] = [d for d in dirnames if d not in SKIP_DIRS]
+        if "config.yaml" in filenames:
+            configs.append(Path(dirpath))
+    return sorted(configs)
+
+
+def check_model_py(example_dir: Path) -> tuple[bool, list[str]]:
+    """Check if model.py exists and has load/predict methods via AST."""
+    issues = []
+    model_dir = example_dir / "model"
+    model_py = model_dir / "model.py"
+
+    if not model_py.exists():
+        # Also check root level
+        model_py = example_dir / "model.py"
+
+    if not model_py.exists():
+        return False, []
+
+    try:
+        source = model_py.read_text()
+        tree = ast.parse(source)
+    except SyntaxError as e:
+        return True, [f"model.py has syntax error: {e}"]
+
+    # Find class with load/predict
+    has_load = False
+    has_predict = False
+    for node in ast.walk(tree):
+        if isinstance(node, ast.ClassDef):
+            for item in node.body:
+                if isinstance(item, (ast.FunctionDef, ast.AsyncFunctionDef)):
+                    if item.name == "load":
+                        has_load = True
+                    if item.name == "predict":
+                        has_predict = True
+
+    if not has_load:
+        issues.append("model.py: no load() method found")
+    if not has_predict:
+        issues.append("model.py: no predict() method found")
+
+    return True, issues
+
+
+def check_config_yaml(example_dir: Path) -> ValidationResult:
+    """Run all validation checks on a single example."""
+    result = ValidationResult(str(example_dir.relative_to(REPO_ROOT)))
+    config_path = example_dir / "config.yaml"
+
+    # 1. Try loading with truss
+    try:
+        raw = yaml.safe_load(config_path.read_text())
+    except Exception as e:
+        result.add_issue(f"YAML parse error: {e}")
+        result.determine_status()
+        return result
+
+    # Check for deprecated fields in raw YAML
+    if raw and isinstance(raw, dict):
+        for field in DEPRECATED_FIELDS_IN_YAML:
+            if field in raw:
+                result.add_issue(f"Deprecated field '{field}' found in config")
+
+        # Check for old base_model values
+        trt_llm = raw.get("trt_llm", {})
+        if isinstance(trt_llm, dict):
+            build = trt_llm.get("build", {})
+            if isinstance(build, dict):
+                base_model = build.get("base_model", "")
+                if base_model in ("llama", "mistral"):
+                    result.add_issue(
+                        f"Deprecated base_model value '{base_model}' (use 'decoder')"
+                    )
+
+    # Try truss.load()
+    try:
+        _ = truss.load(str(example_dir))
+        result.config_loads = True
+    except Exception as e:
+        err_str = str(e)
+        # Truncate long errors
+        if len(err_str) > 200:
+            err_str = err_str[:200] + "..."
+        result.add_issue(f"truss.load() failed: {err_str}")
+        result.config_loads = False
+        result.determine_status()
+        return result
+
+    # 2. Check for docker_server or trt_llm
+    if raw and isinstance(raw, dict):
+        result.has_docker_server = "docker_server" in raw and raw["docker_server"]
+        result.has_trt_llm = "trt_llm" in raw and raw["trt_llm"]
+
+    # 3. Check model.py (not required for docker_server or trt_llm configs)
+    if not result.has_docker_server and not result.has_trt_llm:
+        has_model, model_issues = check_model_py(example_dir)
+        result.has_model_py = has_model
+        if not has_model:
+            result.add_warning("No model.py found")
+        for issue in model_issues:
+            result.add_warning(issue)
+    else:
+        result.has_model_py = True  # Not applicable
+
+    # 4. Check example_model_input
+    if raw and isinstance(raw, dict):
+        model_metadata = raw.get("model_metadata", {})
+        if isinstance(model_metadata, dict):
+            result.has_example_input = "example_model_input" in model_metadata
+
+    # 5. Check README.md
+    result.has_readme = (example_dir / "README.md").exists()
+
+    # 6. Check requirements.txt parseable
+    req_file = example_dir / "requirements.txt"
+    if req_file.exists():
+        try:
+            lines = req_file.read_text().strip().split("\n")
+            for line in lines:
+                line = line.strip()
+                if line and not line.startswith("#") and not line.startswith("-"):
+                    # Basic check - should have valid pip format
+                    pass
+        except Exception as e:
+            result.add_warning(f"requirements.txt issue: {e}")
+
+    result.determine_status()
+    return result
+
+
+def main():
+    output_format = "markdown"
+    if "--json" in sys.argv:
+        output_format = "json"
+    elif "--csv" in sys.argv:
+        output_format = "csv"
+
+    configs = find_all_configs(REPO_ROOT)
+    print(f"Found {len(configs)} example directories with config.yaml", file=sys.stderr)
+
+    results: list[ValidationResult] = []
+    for config_dir in configs:
+        result = check_config_yaml(config_dir)
+        results.append(result)
+
+    # Summary counts
+    status_counts: dict[str, int] = {}
+    for r in results:
+        status_counts[r.status] = status_counts.get(r.status, 0) + 1
+
+    if output_format == "json":
+        data = []
+        for r in results:
+            data.append(
+                {
+                    "path": r.path,
+                    "status": r.status,
+                    "issues": r.issues,
+                    "warnings": r.warnings,
+                    "has_model_py": r.has_model_py,
+                    "has_readme": r.has_readme,
+                    "has_example_input": r.has_example_input,
+                    "has_docker_server": r.has_docker_server,
+                    "has_trt_llm": r.has_trt_llm,
+                }
+            )
+        print(json.dumps({"summary": status_counts, "results": data}, indent=2))
+
+    elif output_format == "csv":
+        print("path,status,issues,warnings,has_model_py,has_readme,has_example_input")
+        for r in results:
+            issues = "; ".join(r.issues).replace(",", ";")
+            warnings = "; ".join(r.warnings).replace(",", ";")
+            print(
+                f"{r.path},{r.status},{issues},{warnings},{r.has_model_py},{r.has_readme},{r.has_example_input}"
+            )
+
+    else:
+        # Markdown
+        print("# Validation Report\n")
+        print(f"**Total examples**: {len(results)}\n")
+        print("## Summary\n")
+        print("| Status | Count |")
+        print("|--------|-------|")
+        for status in [
+            "VALID",
+            "VALID_NO_INPUT",
+            "VALID_NO_README",
+            "DEPRECATED",
+            "BROKEN",
+            "ARCHIVE",
+        ]:
+            count = status_counts.get(status, 0)
+            if count > 0:
+                print(f"| {status} | {count} |")
+        print()
+
+        # Group by status
+        for status in [
+            "BROKEN",
+            "DEPRECATED",
+            "ARCHIVE",
+            "VALID_NO_INPUT",
+            "VALID_NO_README",
+            "VALID",
+        ]:
+            group = [r for r in results if r.status == status]
+            if not group:
+                continue
+            print(f"\n## {status} ({len(group)})\n")
+            print("| Path | Issues | Warnings |")
+            print("|------|--------|----------|")
+            for r in group:
+                issues = "; ".join(r.issues) if r.issues else "-"
+                warnings = "; ".join(r.warnings) if r.warnings else "-"
+                print(f"| `{r.path}` | {issues} | {warnings} |")
+
+    # Write report to file as well
+    report_path = REPO_ROOT / "validation_report.json"
+    data = []
+    for r in results:
+        data.append(
+            {
+                "path": r.path,
+                "status": r.status,
+                "issues": r.issues,
+                "warnings": r.warnings,
+            }
+        )
+    report_path.write_text(
+        json.dumps({"summary": status_counts, "results": data}, indent=2)
+    )
+    print(f"\nReport also saved to {report_path}", file=sys.stderr)
+
+
+if __name__ == "__main__":
+    main()
diff --git a/dockerfiles/ComfyUI.dockerfile b/_internal/dockerfiles/ComfyUI.dockerfile
similarity index 100%
rename from dockerfiles/ComfyUI.dockerfile
rename to _internal/dockerfiles/ComfyUI.dockerfile
diff --git a/essential/config.yaml b/_internal/essential/config.yaml
similarity index 100%
rename from essential/config.yaml
rename to _internal/essential/config.yaml
diff --git a/internal/config.yaml b/_internal/internal/config.yaml
similarity index 100%
rename from internal/config.yaml
rename to _internal/internal/config.yaml
diff --git a/templates/README.md b/_internal/templates/README.md
similarity index 100%
rename from templates/README.md
rename to _internal/templates/README.md
diff --git a/templates/faster-whisper-truss/config.yaml b/_internal/templates/faster-whisper-truss/config.yaml
similarity index 100%
rename from templates/faster-whisper-truss/config.yaml
rename to _internal/templates/faster-whisper-truss/config.yaml
diff --git a/01-getting-started-bert/model/__init__.py b/_internal/templates/faster-whisper-truss/model/__init__.py
similarity index 100%
rename from 01-getting-started-bert/model/__init__.py
rename to _internal/templates/faster-whisper-truss/model/__init__.py
diff --git a/templates/faster-whisper-truss/model/model.py b/_internal/templates/faster-whisper-truss/model/model.py
similarity index 100%
rename from templates/faster-whisper-truss/model/model.py
rename to _internal/templates/faster-whisper-truss/model/model.py
diff --git a/templates/generate.py b/_internal/templates/generate.py
similarity index 100%
rename from templates/generate.py
rename to _internal/templates/generate.py
diff --git a/templates/generate.yaml b/_internal/templates/generate.yaml
similarity index 100%
rename from templates/generate.yaml
rename to _internal/templates/generate.yaml
diff --git a/templates/transformers-openai-compatible/config.yaml b/_internal/templates/transformers-openai-compatible/config.yaml
similarity index 100%
rename from templates/transformers-openai-compatible/config.yaml
rename to _internal/templates/transformers-openai-compatible/config.yaml
diff --git a/02-llm/model/__init__.py b/_internal/templates/transformers-openai-compatible/model/__init__.py
similarity index 100%
rename from 02-llm/model/__init__.py
rename to _internal/templates/transformers-openai-compatible/model/__init__.py
diff --git a/templates/transformers-openai-compatible/model/model.py b/_internal/templates/transformers-openai-compatible/model/model.py
similarity index 100%
rename from templates/transformers-openai-compatible/model/model.py
rename to _internal/templates/transformers-openai-compatible/model/model.py
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/TRT-LLM-README.md b/_internal/templates/trt-llm/TRT-LLM-README.md
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/TRT-LLM-README.md
rename to _internal/templates/trt-llm/TRT-LLM-README.md
diff --git a/templates/trt-llm/config.yaml b/_internal/templates/trt-llm/config.yaml
similarity index 100%
rename from templates/trt-llm/config.yaml
rename to _internal/templates/trt-llm/config.yaml
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/data/.gitattributes b/_internal/templates/trt-llm/data/.gitattributes
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/data/.gitattributes
rename to _internal/templates/trt-llm/data/.gitattributes
diff --git a/03-llm-with-streaming/model/__init__.py b/_internal/templates/trt-llm/model/__init__.py
similarity index 100%
rename from 03-llm-with-streaming/model/__init__.py
rename to _internal/templates/trt-llm/model/__init__.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/model/model.py b/_internal/templates/trt-llm/model/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/model/model.py
rename to _internal/templates/trt-llm/model/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/client.py b/_internal/templates/trt-llm/packages/client.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/client.py
rename to _internal/templates/trt-llm/packages/client.py
diff --git a/templates/trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt.jinja b/_internal/templates/trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt.jinja
similarity index 100%
rename from templates/trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt.jinja
rename to _internal/templates/trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt.jinja
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/1/model.py b/_internal/templates/trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/1/model.py
rename to _internal/templates/trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py
diff --git a/templates/trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt.jinja b/_internal/templates/trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt.jinja
similarity index 100%
rename from templates/trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt.jinja
rename to _internal/templates/trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt.jinja
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/1/model.py b/_internal/templates/trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/1/model.py
rename to _internal/templates/trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py
diff --git a/templates/trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt.jinja b/_internal/templates/trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt.jinja
similarity index 100%
rename from templates/trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt.jinja
rename to _internal/templates/trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt.jinja
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt b/_internal/templates/trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
rename to _internal/templates/trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/utils.py b/_internal/templates/trt-llm/packages/utils.py
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/utils.py
rename to _internal/templates/trt-llm/packages/utils.py
diff --git a/11-embeddings-reranker-classification-tensorrt/templating/.internal_tei/Dockerfile b/_internal/templating/.internal_tei/Dockerfile
similarity index 100%
rename from 11-embeddings-reranker-classification-tensorrt/templating/.internal_tei/Dockerfile
rename to _internal/templating/.internal_tei/Dockerfile
diff --git a/11-embeddings-reranker-classification-tensorrt/templating/.internal_tei/roll_out_docker.sh b/_internal/templating/.internal_tei/roll_out_docker.sh
similarity index 100%
rename from 11-embeddings-reranker-classification-tensorrt/templating/.internal_tei/roll_out_docker.sh
rename to _internal/templating/.internal_tei/roll_out_docker.sh
diff --git a/11-embeddings-reranker-classification-tensorrt/templating/README.md b/_internal/templating/README.md
similarity index 100%
rename from 11-embeddings-reranker-classification-tensorrt/templating/README.md
rename to _internal/templating/README.md
diff --git a/11-embeddings-reranker-classification-tensorrt/templating/deploy_all.py b/_internal/templating/deploy_all.py
similarity index 100%
rename from 11-embeddings-reranker-classification-tensorrt/templating/deploy_all.py
rename to _internal/templating/deploy_all.py
diff --git a/11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py b/_internal/templating/generate_templates.py
similarity index 98%
rename from 11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py
rename to _internal/templating/generate_templates.py
index 05caffdbd..625f2f99f 100644
--- a/11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py
+++ b/_internal/templating/generate_templates.py
@@ -33,8 +33,16 @@
 import copy
 
 REPO_URL = "https://github.com/basetenlabs/truss-examples"
-SUBFOLDER = Path("11-embeddings-reranker-classification-tensorrt")
 ROOT_NAME = Path(REPO_URL.split("/")[-1])
+
+# Map each solution type to its output subdirectory
+SUBFOLDER_MAP = {
+    "BEI": Path("embeddings/bei"),
+    "BEI-Bert": Path("embeddings/bei"),
+    "TEI": Path("embeddings/tei"),
+    "Briton": Path("optimized/briton"),
+    "BISV2": Path("optimized/bisv2"),
+}
 BEI_VERSION = os.environ.get("BEI")
 ENGINE_BUILDER_VERSION = os.environ.get("ENGINE_BUILDER")
 BRITON_VERSION = os.environ.get("BRITON")
@@ -1150,7 +1158,20 @@ def generate_deployment(dp: Deployment):
     root = Path(__file__).parent.parent.parent
     assert root.name == ROOT_NAME.name, "This script has been moved"
 
-    folder_relative_path = SUBFOLDER / dp.folder_name
+    subfolder = SUBFOLDER_MAP.get(dp.solution.nickname)
+    if subfolder is None:
+        raise ValueError(
+            f"No SUBFOLDER_MAP entry for solution nickname '{dp.solution.nickname}'"
+        )
+    # Strip the solution nickname prefix from the folder name since directory
+    # names no longer include the BEI-/TEI-/Briton-/BISV2- prefix
+    raw_folder = dp.folder_name
+    prefix = dp.solution.nickname + "-"
+    if raw_folder.startswith(prefix):
+        stripped_folder = raw_folder[len(prefix) :]
+    else:
+        stripped_folder = raw_folder
+    folder_relative_path = subfolder / stripped_folder
     full_folder_path = root / folder_relative_path
     is_gated_notice = (
         "Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). "
@@ -2339,10 +2360,18 @@ def format_filter(dps: list[Deployment], type_):
         sorted_filter = sorted(
             [dp for dp in dps if isinstance(dp.task, type_)], key=lambda x: x.name
         )
-        names = [
-            f"[{dp.name}-{dp.solution.nickname}]({REPO_URL}/tree/main/{SUBFOLDER}/{dp.folder_name})"
-            for dp in sorted_filter
-        ]
+        names = []
+        for dp in sorted_filter:
+            subfolder = SUBFOLDER_MAP.get(dp.solution.nickname, Path(""))
+            prefix = dp.solution.nickname + "-"
+            folder = (
+                dp.folder_name[len(prefix) :]
+                if dp.folder_name.startswith(prefix)
+                else dp.folder_name
+            )
+            names.append(
+                f"[{dp.name}-{dp.solution.nickname}]({REPO_URL}/tree/main/{subfolder}/{folder})"
+            )
         names_fmt = "\n - ".join(names)
         names_fmt = " - " + names_fmt
         return names_fmt
@@ -2404,5 +2433,6 @@ def format_filter(dps: list[Deployment], type_):
 Examples:
 {generation_names_fmt}
 """
-    (Path(__file__).parent.parent / "README.md").write_text(readme)
+    # Write generated README to _internal/ (not repo root — that's manually maintained)
+    (Path(__file__).parent / "GENERATED_README.md").write_text(readme)
     print(readme)
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/README.md b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/README.md
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/README.md
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/README.md
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/config.yaml b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/config.yaml
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/config.yaml
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/config.yaml
diff --git a/04-image-generation/model/__init__.py b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/model/__init__.py
similarity index 100%
rename from 04-image-generation/model/__init__.py
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/model/__init__.py
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/README.md b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/README.md
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/README.md
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/README.md
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/config.yaml b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/config.yaml
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/config.yaml
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/config.yaml
diff --git a/05-speech-to-text/model/__init__.py b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/model/__init__.py
similarity index 100%
rename from 05-speech-to-text/model/__init__.py
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/model/__init__.py
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/README.md b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/README.md
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/README.md
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/README.md
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/config.yaml b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/config.yaml
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/config.yaml
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/config.yaml
diff --git a/06-high-performance-cached-weights/model/__init__.py b/_internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/model/__init__.py
similarity index 100%
rename from 06-high-performance-cached-weights/model/__init__.py
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/model/__init__.py
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/README.md b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/README.md
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/README.md
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/README.md
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/config.yaml b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/config.yaml
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/config.yaml
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/config.yaml
diff --git a/07-high-performance-dynamic-batching/model/__init__.py b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/model/__init__.py
similarity index 100%
rename from 07-high-performance-dynamic-batching/model/__init__.py
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/model/__init__.py
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/README.md b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/README.md
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/README.md
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/README.md
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/config.yaml b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/config.yaml
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/config.yaml
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/config.yaml
diff --git a/07-high-performance-dynamic-batching/packages/__init__.py b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/model/__init__.py
similarity index 100%
rename from 07-high-performance-dynamic-batching/packages/__init__.py
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/model/__init__.py
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/README.md b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/README.md
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/README.md
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/README.md
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/config.yaml b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/config.yaml
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/config.yaml
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/config.yaml
diff --git a/09-private-huggingface/model/__init__.py b/_internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/model/__init__.py
similarity index 100%
rename from 09-private-huggingface/model/__init__.py
rename to _internal/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/model/__init__.py
diff --git a/audio/README.md b/audio/README.md
new file mode 100644
index 000000000..a3af29879
--- /dev/null
+++ b/audio/README.md
@@ -0,0 +1,33 @@
+# Audio Models
+
+Truss configurations for speech-to-text, text-to-speech, music generation, and multimodal audio models. Includes both batch and streaming deployments.
+
+| Directory | Variants | Description |
+|-----------|----------|-------------|
+| [whisper](whisper/) | 8 | OpenAI Whisper speech-to-text models including Faster Whisper, WhisperX, and streaming variants |
+| [kokoro](kokoro/) | 1 | Kokoro text-to-speech model |
+| [chatterbox-tts](chatterbox-tts/) | 1 | Chatterbox text-to-speech model |
+| [piper-tts](piper-tts/) | 1 | Piper lightweight text-to-speech engine |
+| [xtts-v2](xtts-v2/) | 1 | Coqui XTTS v2 multilingual text-to-speech |
+| [xtts-streaming](xtts-streaming/) | 1 | Coqui XTTS with streaming audio output |
+| [orpheus-3b-websockets](orpheus-3b-websockets/) | 1 | Orpheus 3B TTS with WebSocket streaming |
+| [orpheus-best-performance](orpheus-best-performance/) | 1 | Orpheus TTS optimized for lowest latency |
+| [sesame-csm-1b](sesame-csm-1b/) | 1 | Sesame CSM 1B conversational speech model |
+| [metavoice-1b](metavoice-1b/) | 1 | MetaVoice 1B text-to-speech model |
+| [ultravox](ultravox/) | 1 | Ultravox multimodal audio-language model |
+| [audiogen-medium](audiogen-medium/) | 1 | Meta AudioGen medium audio generation from text |
+| [musicgen-large](musicgen-large/) | 1 | Meta MusicGen large music generation |
+| [musicgen-melody](musicgen-melody/) | 1 | Meta MusicGen melody-conditioned music generation |
+| [nvidia-parakeet](nvidia-parakeet/) | 1 | NVIDIA Parakeet automatic speech recognition |
+| [qwen-asr](qwen-asr/) | 1 | Qwen automatic speech recognition model |
+| [qwen-omni](qwen-omni/) | 1 | Qwen Omni multimodal audio model |
+| [qwen-omni-thinker](qwen-omni-thinker/) | 1 | Qwen Omni Thinker reasoning audio model |
+| [voxtral-streaming-4b](voxtral-streaming-4b/) | 1 | Mistral Voxtral 4B streaming speech model |
+
+## Deploying
+
+Each audio model can be deployed to Baseten with:
+
+```bash
+truss push
+```
diff --git a/audio/audiogen-medium/README.md b/audio/audiogen-medium/README.md
new file mode 100644
index 000000000..ddf3be8e2
--- /dev/null
+++ b/audio/audiogen-medium/README.md
@@ -0,0 +1,36 @@
+# AudioGen medium
+
+AudioGen is a simple and controllable model for audio generation developed by Facebook AI Research.
+
+| Property | Value |
+|----------|-------|
+| Task | Audio generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "duration": 8,
+  "prompts": [
+    "dog barking",
+    "sirene of an emergency vehicle",
+    "footsteps in a corridor"
+  ]
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/audio/audiogen-medium/config.yaml b/audio/audiogen-medium/config.yaml
new file mode 100644
index 000000000..98d8d0752
--- /dev/null
+++ b/audio/audiogen-medium/config.yaml
@@ -0,0 +1,30 @@
+description: AudioGen is a simple and controllable model for audio generation developed
+  by Facebook AI Research.
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "facebook/audiogen-medium"
+  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/musicgen-cover.png
+  example_model_input:
+    duration: 8
+    prompts:
+    - dog barking
+    - sirene of an emergency vehicle
+    - footsteps in a corridor
+  tags:
+  - text-to-audio
+model_name: AudioGen medium
+python_version: py39
+requirements:
+- torch>=2
+- git+https://github.com/facebookresearch/audiocraft.git
+- torchaudio==2.1.0
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/10-using-system-packages/model/__init__.py b/audio/audiogen-medium/model/__init__.py
similarity index 100%
rename from 10-using-system-packages/model/__init__.py
rename to audio/audiogen-medium/model/__init__.py
diff --git a/audiogen-medium/model/model.py b/audio/audiogen-medium/model/model.py
similarity index 100%
rename from audiogen-medium/model/model.py
rename to audio/audiogen-medium/model/model.py
diff --git a/audio/chatterbox-tts/README.md b/audio/chatterbox-tts/README.md
new file mode 100644
index 000000000..ec709e9a6
--- /dev/null
+++ b/audio/chatterbox-tts/README.md
@@ -0,0 +1,31 @@
+# Chatterbox TTS
+
+Deploy Chatterbox TTS for text-to-speech on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text-to-speech |
+| Engine | Custom (Truss) |
+| GPU | H100 |
+| Python | py312 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"text": "Hello, this is a test of text to speech."}'
+```
+
+## Configuration highlights
+
+- Base image: `jojobaseten/truss-numpy-1.26.0-gpu:0.4`
diff --git a/audio/chatterbox-tts/config.yaml b/audio/chatterbox-tts/config.yaml
new file mode 100644
index 000000000..46aed1f53
--- /dev/null
+++ b/audio/chatterbox-tts/config.yaml
@@ -0,0 +1,18 @@
+description: "Chatterbox TTS for text-to-speech"
+model_metadata:
+  repo_id: "resemble-ai/chatterbox"
+  example_model_input: {"text": "Hello, how are you today?"}
+model_name: Chatterbox TTS
+base_image:
+  image: jojobaseten/truss-numpy-1.26.0-gpu:0.4
+  python_executable_path: /usr/bin/python3
+python_version: py312
+requirements:
+  - chatterbox-tts==0.1.1
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 40Gi
+  use_gpu: true
+secrets:
+  hf_access_token: null
diff --git a/chatterbox-tts/docker/Dockerfile b/audio/chatterbox-tts/docker/Dockerfile
similarity index 100%
rename from chatterbox-tts/docker/Dockerfile
rename to audio/chatterbox-tts/docker/Dockerfile
diff --git a/chatterbox-tts/docker/docker_build.sh b/audio/chatterbox-tts/docker/docker_build.sh
similarity index 100%
rename from chatterbox-tts/docker/docker_build.sh
rename to audio/chatterbox-tts/docker/docker_build.sh
diff --git a/chatterbox-tts/input/obama_8s.wav b/audio/chatterbox-tts/input/obama_8s.wav
similarity index 100%
rename from chatterbox-tts/input/obama_8s.wav
rename to audio/chatterbox-tts/input/obama_8s.wav
diff --git a/audiogen-medium/model/__init__.py b/audio/chatterbox-tts/model/__init__.py
similarity index 100%
rename from audiogen-medium/model/__init__.py
rename to audio/chatterbox-tts/model/__init__.py
diff --git a/chatterbox-tts/model/model.py b/audio/chatterbox-tts/model/model.py
similarity index 100%
rename from chatterbox-tts/model/model.py
rename to audio/chatterbox-tts/model/model.py
diff --git a/chatterbox-tts/run_tts.py b/audio/chatterbox-tts/run_tts.py
similarity index 100%
rename from chatterbox-tts/run_tts.py
rename to audio/chatterbox-tts/run_tts.py
diff --git a/audio/kokoro/README.md b/audio/kokoro/README.md
new file mode 100644
index 000000000..661a3e716
--- /dev/null
+++ b/audio/kokoro/README.md
@@ -0,0 +1,34 @@
+# kokoro
+
+Deploy kokoro for text-to-speech on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text-to-speech |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "text": "Kokoro is a frontier TTS model for its size of 82 million parameters (text in/audio out). On 25 Dec 2024, Kokoro v0.19 weights were permissively released in full fp32 precision under an Apache 2.0 license. As of 2 Jan 2025, 10 unique Voicepacks have been released, and a .onnx version of v0.19 is available.In the weeks leading up to its release, Kokoro v0.19 was the #1\ud83e\udd47 ranked model in TTS Spaces Arena. Kokoro had achieved higher Elo in this single-voice Arena setting over other models, using fewer parameters and less data. Kokoro's ability to top this Elo ladder suggests that the scaling law (Elo vs compute/data/params) for traditional TTS models might have a steeper slope than previously expected.",
+  "voice": "af",
+  "speed": 1.0
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **1**
+- System packages: `espeak-ng`
diff --git a/kokoro/call.py b/audio/kokoro/call.py
similarity index 100%
rename from kokoro/call.py
rename to audio/kokoro/call.py
diff --git a/audio/kokoro/config.yaml b/audio/kokoro/config.yaml
new file mode 100644
index 000000000..68f8cc06f
--- /dev/null
+++ b/audio/kokoro/config.yaml
@@ -0,0 +1,27 @@
+description: "kokoro for text-to-speech"
+build_commands:
+- python3 -c "import nltk; nltk.download('punkt'); nltk.download('punkt_tab')"
+environment_variables: {}
+model_metadata:
+  repo_id: "hexgrad/Kokoro-82M"
+  example_model_input: {"text": "Kokoro is a frontier TTS model for its size of 82 million parameters (text in/audio out). On 25 Dec 2024, Kokoro v0.19 weights were permissively released in full fp32 precision under an Apache 2.0 license. As of 2 Jan 2025, 10 unique Voicepacks have been released, and a .onnx version of v0.19 is available.In the weeks leading up to its release, Kokoro v0.19 was the #1🥇 ranked model in TTS Spaces Arena. Kokoro had achieved higher Elo in this single-voice Arena setting over other models, using fewer parameters and less data. Kokoro's ability to top this Elo ladder suggests that the scaling law (Elo vs compute/data/params) for traditional TTS models might have a steeper slope than previously expected.", "voice": "af", "speed": 1.0}
+model_name: kokoro
+python_version: py311
+requirements:
+- torch==2.5.1
+- transformers==4.48.0
+- scipy==1.15.1
+- phonemizer==3.3.0
+- nltk==3.9.1
+- numpy==1.26.0
+- huggingface_hub[hf_transfer]==0.19.4
+- hf_transfer==0.1.9
+- munch==4.0.0
+resources:
+  accelerator: T4
+  use_gpu: true
+runtime:
+  predict_concurrency: 1
+secrets: {}
+system_packages:
+- espeak-ng
diff --git a/autodesk-wala/model/__init__.py b/audio/kokoro/model/__init__.py
similarity index 100%
rename from autodesk-wala/model/__init__.py
rename to audio/kokoro/model/__init__.py
diff --git a/kokoro/model/model.py b/audio/kokoro/model/model.py
similarity index 100%
rename from kokoro/model/model.py
rename to audio/kokoro/model/model.py
diff --git a/audio/metavoice-1b/README.md b/audio/metavoice-1b/README.md
new file mode 100644
index 000000000..a516a7a9a
--- /dev/null
+++ b/audio/metavoice-1b/README.md
@@ -0,0 +1,32 @@
+# MetaVoice 1B
+
+MetaVoice is a transformer-based model for TTS
+
+| Property | Value |
+|----------|-------|
+| Model | [metavoiceio/metavoice-1B-v0.1](https://huggingface.co/metavoiceio/metavoice-1B-v0.1) |
+| Task | Text-to-speech |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '"text to speech models are cool"'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/audio/metavoice-1b/config.yaml b/audio/metavoice-1b/config.yaml
new file mode 100644
index 000000000..191223b0a
--- /dev/null
+++ b/audio/metavoice-1b/config.yaml
@@ -0,0 +1,30 @@
+model_name: MetaVoice 1B
+description: MetaVoice is a transformer-based model for TTS
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  example_model_input: '"text to speech models are cool"'
+python_version: py311
+data_dir: data
+model_cache:
+  - repo_id: metavoiceio/metavoice-1B-v0.1
+    use_volume: false
+    allow_patterns:
+      - "*.pt"
+  - repo_id: facebook/multiband-diffusion
+    use_volume: false
+    allow_patterns:
+      - mbd_comp_8.pt
+  - repo_id: facebook/encodec_24khz
+    use_volume: false
+    allow_patterns:
+      - "*.safetensors"
+
+requirements_file: ./requirements.txt
+resources:
+  accelerator: "A10G"
+  use_gpu: true
+secrets:
+  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
+system_packages:
+- ffmpeg
diff --git a/metavoice-1b/data/bria.mp3 b/audio/metavoice-1b/data/bria.mp3
similarity index 100%
rename from metavoice-1b/data/bria.mp3
rename to audio/metavoice-1b/data/bria.mp3
diff --git a/autodesk-wala/packages/src/__init__.py b/audio/metavoice-1b/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/__init__.py
rename to audio/metavoice-1b/model/__init__.py
diff --git a/metavoice-1b/model/model.py b/audio/metavoice-1b/model/model.py
similarity index 100%
rename from metavoice-1b/model/model.py
rename to audio/metavoice-1b/model/model.py
diff --git a/metavoice-1b/process.py b/audio/metavoice-1b/process.py
similarity index 100%
rename from metavoice-1b/process.py
rename to audio/metavoice-1b/process.py
diff --git a/metavoice-1b/requirements.txt b/audio/metavoice-1b/requirements.txt
similarity index 100%
rename from metavoice-1b/requirements.txt
rename to audio/metavoice-1b/requirements.txt
diff --git a/audio/musicgen-large/README.md b/audio/musicgen-large/README.md
new file mode 100644
index 000000000..604590fbd
--- /dev/null
+++ b/audio/musicgen-large/README.md
@@ -0,0 +1,36 @@
+# MusicGen large
+
+MusicGen is a simple and controllable model for music generation developed by Facebook AI Research.
+
+| Property | Value |
+|----------|-------|
+| Task | Audio generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "duration": 8,
+  "prompts": [
+    "happy rock",
+    "energetic EDM",
+    "sad jazz"
+  ]
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/audio/musicgen-large/config.yaml b/audio/musicgen-large/config.yaml
new file mode 100644
index 000000000..1475c042d
--- /dev/null
+++ b/audio/musicgen-large/config.yaml
@@ -0,0 +1,29 @@
+description: MusicGen is a simple and controllable model for music generation developed
+  by Facebook AI Research.
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "facebook/musicgen-large"
+  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/musicgen-cover.png
+  example_model_input:
+    duration: 8
+    prompts:
+    - happy rock
+    - energetic EDM
+    - sad jazz
+  tags:
+  - text-to-music
+model_name: MusicGen large
+python_version: py39
+requirements:
+- torch>=2
+- audiocraft==1.3.0
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/autodesk-wala/packages/src/diffusion_modules/__init__.py b/audio/musicgen-large/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/__init__.py
rename to audio/musicgen-large/model/__init__.py
diff --git a/musicgen-large/model/model.py b/audio/musicgen-large/model/model.py
similarity index 100%
rename from musicgen-large/model/model.py
rename to audio/musicgen-large/model/model.py
diff --git a/audio/musicgen-melody/README.md b/audio/musicgen-melody/README.md
new file mode 100644
index 000000000..32afc8be1
--- /dev/null
+++ b/audio/musicgen-melody/README.md
@@ -0,0 +1,36 @@
+# MusicGen Melody
+
+MusicGen Melody is a simple and controllable model for music generation conditioned on text and audio. It is developed by Facebook AI Research.
+
+| Property | Value |
+|----------|-------|
+| Task | Audio generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "duration": 8,
+  "prompts": [
+    "happy rock",
+    "energetic EDM",
+    "sad jazz"
+  ]
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/audio/musicgen-melody/config.yaml b/audio/musicgen-melody/config.yaml
new file mode 100644
index 000000000..82fe49b63
--- /dev/null
+++ b/audio/musicgen-melody/config.yaml
@@ -0,0 +1,30 @@
+description: MusicGen Melody is a simple and controllable model for music generation
+  conditioned on text and audio. It is developed by Facebook AI Research.
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "facebook/musicgen-melody"
+  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/musicgen-cover.png
+  example_model_input:
+    duration: 8
+    prompts:
+    - happy rock
+    - energetic EDM
+    - sad jazz
+  tags:
+  - text-to-music
+model_name: MusicGen Melody
+python_version: py39
+requirements:
+- torch>=2
+- audiocraft==1.3.0
+- protobuf==4.25.1
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/autodesk-wala/packages/src/experiments/__init__.py b/audio/musicgen-melody/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/experiments/__init__.py
rename to audio/musicgen-melody/model/__init__.py
diff --git a/musicgen-melody/model/model.py b/audio/musicgen-melody/model/model.py
similarity index 100%
rename from musicgen-melody/model/model.py
rename to audio/musicgen-melody/model/model.py
diff --git a/audio/nvidia-parakeet/README.md b/audio/nvidia-parakeet/README.md
new file mode 100644
index 000000000..26b6b7f87
--- /dev/null
+++ b/audio/nvidia-parakeet/README.md
@@ -0,0 +1,36 @@
+# Parakeet TDT 0.6B V2
+
+Parakeet TDT 0.6B V2 is a 600-million-parameter automatic speech recognition (ASR) model designed for high-quality English transcription.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/parakeet-tdt-0.6b-v2](https://huggingface.co/nvidia/parakeet-tdt-0.6b-v2) |
+| Task | Audio |
+| Engine | Custom (Truss) |
+| GPU | H100_40GB |
+| Python | py312 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "audio_url": "https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav",
+  "timestamps": false
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **8**
+- System packages: `ffmpeg`
diff --git a/audio/nvidia-parakeet/config.yaml b/audio/nvidia-parakeet/config.yaml
new file mode 100644
index 000000000..ae599d138
--- /dev/null
+++ b/audio/nvidia-parakeet/config.yaml
@@ -0,0 +1,25 @@
+description: Parakeet TDT 0.6B V2 is a 600-million-parameter automatic speech recognition (ASR) model designed for high-quality English transcription.
+python_version: py312
+model_metadata:
+  repo_id: nvidia/parakeet-tdt-0.6b-v2
+  avatar_url: https://cdn-avatars.huggingface.co/v1/production/uploads/1613114437487-60262a8e0703121c822a80b6.png
+  example_model_input:
+    {
+      "audio_url": "https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav",
+      "timestamps": false
+    }
+system_packages:
+  - ffmpeg
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+runtime:
+  predict_concurrency: 8
+model_name: Parakeet TDT 0.6B V2
+secrets:
+  hf_access_token: null
+requirements:
+  - nemo_toolkit[asr]==2.2.0
+  - requests==2.32.3
+  - pyarrow==20.0.0
+  - cuda-python>=12.3
diff --git a/autodesk-wala/packages/src/experiments/utils/__init__.py b/audio/nvidia-parakeet/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/experiments/utils/__init__.py
rename to audio/nvidia-parakeet/model/__init__.py
diff --git a/nvidia/parakeet-tdt-0_6b-v2/model/model.py b/audio/nvidia-parakeet/model/model.py
similarity index 100%
rename from nvidia/parakeet-tdt-0_6b-v2/model/model.py
rename to audio/nvidia-parakeet/model/model.py
diff --git a/audio/orpheus-3b-websockets/README.md b/audio/orpheus-3b-websockets/README.md
new file mode 100644
index 000000000..40d213a9e
--- /dev/null
+++ b/audio/orpheus-3b-websockets/README.md
@@ -0,0 +1,40 @@
+# Orpheus-3b Websockets
+
+Deploy Orpheus-3b Websockets on Baseten using a TRT-LLM engine.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/orpheus-3b-0.1-ft](https://huggingface.co/baseten/orpheus-3b-0.1-ft) |
+| Task | Audio |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 KV |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "max_tokens": 10000,
+  "prompt": "In todays fast-paced world, finding balance between work and personal life is more important than ever. With the constant demands of technology, remote communication, ",
+  "voice": "tara"
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **65,536**
+- Plugin: **use_fp8_context_fmha**
+- Environment variables: `ENABLE_EXECUTOR_API`
diff --git a/orpheus-3b-websockets/call.py b/audio/orpheus-3b-websockets/call.py
similarity index 100%
rename from orpheus-3b-websockets/call.py
rename to audio/orpheus-3b-websockets/call.py
diff --git a/audio/orpheus-3b-websockets/config.yaml b/audio/orpheus-3b-websockets/config.yaml
new file mode 100644
index 000000000..587f32b30
--- /dev/null
+++ b/audio/orpheus-3b-websockets/config.yaml
@@ -0,0 +1,60 @@
+description: "Orpheus-3b Websockets for text-to-speech"
+build_commands:
+  - apt-get update && apt-get install git git-lfs -y
+  - git lfs install
+  - git clone https://huggingface.co/hubertsiuzdak/snac_24khz /app/snac_24khz
+environment_variables:
+  ENABLE_EXECUTOR_API: "1"
+model_metadata:
+  example_model_input:
+    max_tokens: 10000
+    prompt:
+      "In todays fast-paced world, finding balance between work and personal
+      life is more important than ever. With the constant demands of technology, remote
+      communication, "
+    voice: tara
+  tags:
+    - force-legacy-api-non-openai-compatible
+model_name: Orpheus-3b Websockets
+python_version: py39
+requirements:
+  - snac==1.2.1
+  - torch==2.7.0
+  - batched==0.1.4
+  - httpx==0.27.0
+  - websockets==13.1
+  - pysbd==0.3.4
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  is_websocket_endpoint: true
+  transport:
+    kind: websocket
+    ping_interval_seconds: null
+    ping_timeout_seconds: null
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: baseten/orpheus-3b-0.1-ft
+      revision: b9eb57a06083cb9e5a083885fad991aa79c0bd24
+      source: HF
+    max_batch_size: 256
+    # set higher, so we can always use the max batch size in a single iter.
+    max_num_tokens: 16384
+    # 65536 would be around 600s of audio, typically model produces max 120s.
+    max_seq_len: 65536
+    num_builder_gpus: 1
+    quantization_config:
+      # TODO: Generate a typical dataset (input + output tokens) in target language
+      # or disable quantization for other languages
+      calib_dataset: "cnn_dailymail"
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
diff --git a/orpheus-3b-websockets/model/model.py b/audio/orpheus-3b-websockets/model/model.py
similarity index 100%
rename from orpheus-3b-websockets/model/model.py
rename to audio/orpheus-3b-websockets/model/model.py
diff --git a/orpheus-3b-websockets/snac_batching_quantization_dev.py b/audio/orpheus-3b-websockets/snac_batching_quantization_dev.py
similarity index 100%
rename from orpheus-3b-websockets/snac_batching_quantization_dev.py
rename to audio/orpheus-3b-websockets/snac_batching_quantization_dev.py
diff --git a/audio/orpheus-best-performance/README.md b/audio/orpheus-best-performance/README.md
new file mode 100644
index 000000000..d6822e71e
--- /dev/null
+++ b/audio/orpheus-best-performance/README.md
@@ -0,0 +1,42 @@
+# Orpheus-3b Best Performance
+
+Deploy Orpheus-3b Best Performance on Baseten using a TRT-LLM engine.
+
+| Property | Value |
+|----------|-------|
+| Model | [canopylabs/orpheus-3b-0.1-ft](https://huggingface.co/canopylabs/orpheus-3b-0.1-ft) |
+| Task | Audio |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 KV |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "max_tokens": 10000,
+  "prompt": "In todays fast-paced world, finding balance between work and personal life is more important than ever. With the constant demands of technology, remote communication, ",
+  "voice": "tara"
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_fp8_context_fmha**
+- Environment variables: `ENABLE_EXECUTOR_API`
diff --git a/orpheus-best-performance/call.py b/audio/orpheus-best-performance/call.py
similarity index 100%
rename from orpheus-best-performance/call.py
rename to audio/orpheus-best-performance/call.py
diff --git a/audio/orpheus-best-performance/config.yaml b/audio/orpheus-best-performance/config.yaml
new file mode 100644
index 000000000..e73e9f4d4
--- /dev/null
+++ b/audio/orpheus-best-performance/config.yaml
@@ -0,0 +1,59 @@
+description: "Orpheus-3b Best Performance for text-to-speech"
+build_commands:
+  - apt-get update && apt-get install git git-lfs -y
+  - git lfs install
+  - git clone https://huggingface.co/hubertsiuzdak/snac_24khz /app/snac_24khz
+environment_variables:
+  ENABLE_EXECUTOR_API: "1"
+model_metadata:
+  repo_id: canopylabs/orpheus-3b-0.1-ft
+  example_model_input:
+    max_tokens: 10000
+    prompt:
+      "In todays fast-paced world, finding balance between work and personal
+      life is more important than ever. With the constant demands of technology, remote
+      communication, "
+    voice: tara
+  tags:
+    - force-legacy-api-non-openai-compatible
+model_name: Orpheus-3b Best Performance
+python_version: py39
+requirements:
+  - --extra-index-url https://download.pytorch.org/whl/cu128
+  - torch==2.7.1
+  - snac==1.2.1
+  - batched==0.1.4
+resources:
+  # NOTE: Model is bottlenecked by CPU clock speed
+  # H100 upgrade is not really effective
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+secrets:
+  hf_access_token: null
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: baseten/orpheus-3b-0.1-ft
+      revision: b9eb57a06083cb9e5a083885fad991aa79c0bd24
+      source: HF
+    max_batch_size: 256
+    # set higher, so we can always use the max batch size in a single iter.
+    max_num_tokens: 16384
+    # 32768 would be around 300s of audio, typically model produces max 120s.
+    max_seq_len: 32768
+    num_builder_gpus: 1
+    quantization_config:
+      # TODO: Generate a typical dataset (input + output tokens) in target language
+      # or disable quantization for other languages
+      calib_dataset: "cnn_dailymail"
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
+    kv_cache_free_gpu_mem_fraction: 0.90
+    batch_scheduler_policy: max_utilization
diff --git a/orpheus-best-performance/model/model.py b/audio/orpheus-best-performance/model/model.py
similarity index 100%
rename from orpheus-best-performance/model/model.py
rename to audio/orpheus-best-performance/model/model.py
diff --git a/orpheus-best-performance/snac_batching_quantization_dev.py b/audio/orpheus-best-performance/snac_batching_quantization_dev.py
similarity index 100%
rename from orpheus-best-performance/snac_batching_quantization_dev.py
rename to audio/orpheus-best-performance/snac_batching_quantization_dev.py
diff --git a/audio/piper-tts/README.md b/audio/piper-tts/README.md
new file mode 100644
index 000000000..843cbeaa8
--- /dev/null
+++ b/audio/piper-tts/README.md
@@ -0,0 +1,31 @@
+# Piper TTS
+
+Deploy Piper TTS for text-to-speech on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text-to-speech |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "text": "I love robots. Robots are cool!"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/audio/piper-tts/config.yaml b/audio/piper-tts/config.yaml
new file mode 100644
index 000000000..fad82ed4a
--- /dev/null
+++ b/audio/piper-tts/config.yaml
@@ -0,0 +1,25 @@
+description: "Piper TTS for text-to-speech"
+environment_variables: {}
+external_data:
+- local_data_path: models/model.onnx
+  url: https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/lessac/medium/en_US-lessac-medium.onnx
+- local_data_path: models/model.onnx.json
+  url: https://huggingface.co/rhasspy/piper-voices/raw/v1.0.0/en/en_US/lessac/medium/en_US-lessac-medium.onnx.json
+external_package_dirs: []
+model_metadata:
+  repo_id: "rhasspy/piper-voices"
+  example_model_input:
+    text: I love robots. Robots are cool!
+  tags:
+  - text-to-speech
+model_name: Piper TTS
+python_version: py310
+requirements:
+- piper-tts==1.2.0
+resources:
+  accelerator: T4
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/autodesk-wala/packages/src/latent_model/__init__.py b/audio/piper-tts/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/__init__.py
rename to audio/piper-tts/model/__init__.py
diff --git a/piper-tts/model/model.py b/audio/piper-tts/model/model.py
similarity index 100%
rename from piper-tts/model/model.py
rename to audio/piper-tts/model/model.py
diff --git a/audio/qwen-asr/README.md b/audio/qwen-asr/README.md
new file mode 100644
index 000000000..e25eb7ea9
--- /dev/null
+++ b/audio/qwen-asr/README.md
@@ -0,0 +1,49 @@
+# Qwen3-ASR-1.7B
+
+Deploy Qwen3-ASR-1.7B on Baseten using a vLLM engine.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-ASR-1.7B](https://huggingface.co/Qwen/Qwen3-ASR-1.7B) |
+| Task | Audio |
+| Engine | vLLM |
+| GPU | H100_40GB:1 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "stream": false,
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "audio_url",
+          "audio_url": {
+            "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"
+          }
+        }
+      ]
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:nightly-070c811d6f74c55302557878f5982411a3346b4d`
+- Predict concurrency: **256**
+- System packages: `python3.10-venv, ffmpeg, openmpi-bin, libopenmpi-dev`
diff --git a/audio/qwen-asr/config.yaml b/audio/qwen-asr/config.yaml
new file mode 100644
index 000000000..cf55fe465
--- /dev/null
+++ b/audio/qwen-asr/config.yaml
@@ -0,0 +1,44 @@
+description: "Qwen3-ASR-1.7B for speech-to-text"
+model_metadata:
+  repo_id: "Qwen/Qwen3-ASR-1.7B"
+  example_model_input:
+    stream: false
+    messages:
+      - role: user
+        content:
+          - type: audio_url
+            audio_url:
+              url: https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav
+  tags:
+    - openai-compatible
+model_name: Qwen3-ASR-1.7B
+secrets:
+  hf_access_token: null
+base_image:
+  image:  vllm/vllm-openai:nightly-070c811d6f74c55302557878f5982411a3346b4d
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve Qwen/Qwen3-ASR-1.7B --gpu-memory-utilization 0.8 --host 0.0.0.0 --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100_40GB:1
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+requirements:
+  - --pre --extra-index-url https://wheels.vllm.ai/nightly
+  - vllm[audio]
+  - librosa==0.10.2
+  - torch==2.5.1
+  - torchaudio==2.5.1
+  - pynvml==11.5.3
+  - ffmpeg-python==0.2.0
+system_packages:
+  - python3.10-venv
+  - ffmpeg
+  - openmpi-bin
+  - libopenmpi-dev
+runtime:
+  predict_concurrency: 256
diff --git a/audio/qwen-omni-thinker/README.md b/audio/qwen-omni-thinker/README.md
new file mode 100644
index 000000000..91c0a94bf
--- /dev/null
+++ b/audio/qwen-omni-thinker/README.md
@@ -0,0 +1,67 @@
+# Qwen3 Omni 30B Instruct (Thinker Only)
+
+Deploy Qwen3 Omni 30B Instruct (Thinker Only) on Baseten using a vLLM engine.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-Omni-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Omni-30B-A3B-Instruct) |
+| Task | Audio |
+| Engine | vLLM |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "messages": [
+    {
+      "role": "system",
+      "content": "You are a helpful assistant."
+    },
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "Describe this image and audio content."
+        },
+        {
+          "type": "image_url",
+          "image_url": {
+            "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"
+          }
+        },
+        {
+          "type": "audio_url",
+          "audio_url": {
+            "url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"
+          }
+        },
+        {
+          "type": "text",
+          "text": "What can you see and hear? Answer in one sentence."
+        }
+      ]
+    }
+  ],
+  "stream": false,
+  "model": "qwen3-omni",
+  "max_tokens": 2048,
+  "temperature": 0.7
+}'
+```
+
+## Configuration highlights
+
+- Base image: `qwenllm/qwen3-omni:3-cu124`
+- Predict concurrency: **32**
diff --git a/audio/qwen-omni-thinker/config.yaml b/audio/qwen-omni-thinker/config.yaml
new file mode 100644
index 000000000..4f740a03a
--- /dev/null
+++ b/audio/qwen-omni-thinker/config.yaml
@@ -0,0 +1,43 @@
+description: "Qwen3 Omni 30B Instruct (Thinker Only) for audio understanding"
+model_name: Qwen3 Omni 30B Instruct (Thinker Only)
+base_image:
+  image: qwenllm/qwen3-omni:3-cu124
+docker_server:
+  start_command: |
+    sh -c "vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --dtype bfloat16 --max-model-len 65536 --served-model-name qwen3-omni"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+model_metadata:
+  repo_id: "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+  example_model_input:
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content:
+          - type: text
+            text: "Describe this image and audio content."
+          - type: image_url
+            image_url:
+              url: "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"
+          - type: audio_url
+            audio_url:
+              url: "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"
+          - type: text
+            text: "What can you see and hear? Answer in one sentence."
+
+    stream: false
+    model: "qwen3-omni"
+    max_tokens: 2048
+    temperature: 0.7
+  tags:
+    - openai-compatible
+    - multimodal
+    - image-processing
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/audio/qwen-omni/README.md b/audio/qwen-omni/README.md
new file mode 100644
index 000000000..5c6be2571
--- /dev/null
+++ b/audio/qwen-omni/README.md
@@ -0,0 +1,43 @@
+# Qwen3 Omni 30B Instruct
+
+Deploy Qwen3 Omni 30B Instruct on Baseten using a Custom (Truss) engine.
+
+| Property | Value |
+|----------|-------|
+| Task | Audio |
+| Engine | Custom (Truss) |
+| GPU | H100 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "speaker": "Chelsie",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "Hi, how are you?"
+        }
+      ]
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `qwenllm/qwen3-omni`
+- Predict concurrency: **1**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/audio/qwen-omni/config.yaml b/audio/qwen-omni/config.yaml
new file mode 100644
index 000000000..94d2638db
--- /dev/null
+++ b/audio/qwen-omni/config.yaml
@@ -0,0 +1,27 @@
+description: "Qwen3 Omni 30B Instruct for audio understanding"
+model_name: Qwen3 Omni 30B Instruct
+base_image:
+  image: qwenllm/qwen3-omni
+model_metadata:
+  repo_id: "Qwen/Qwen3-Omni-30B-A3B-Instruct"
+  example_model_input: {
+    "speaker": "Chelsie",
+    "messages": [
+        {
+            "role": "user",
+            "content": [
+                {
+                    "type": "text",
+                    "text": "Hi, how are you?"
+                }
+            ]
+        }
+    ]
+  }
+runtime:
+  predict_concurrency : 1
+resources:
+  accelerator: H100
+  use_gpu: true
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
diff --git a/qwen/qwen-3-30b-omni/model/model.py b/audio/qwen-omni/model/model.py
similarity index 100%
rename from qwen/qwen-3-30b-omni/model/model.py
rename to audio/qwen-omni/model/model.py
diff --git a/audio/sesame-csm-1b/README.md b/audio/sesame-csm-1b/README.md
new file mode 100644
index 000000000..29190f6ce
--- /dev/null
+++ b/audio/sesame-csm-1b/README.md
@@ -0,0 +1,34 @@
+# sesame-csm-1b
+
+Deploy sesame-csm-1b for text-to-speech on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text-to-speech |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "text": "Hello from Sesame.",
+  "speaker": 0
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/audio/sesame-csm-1b/config.yaml b/audio/sesame-csm-1b/config.yaml
new file mode 100644
index 000000000..e38f325fb
--- /dev/null
+++ b/audio/sesame-csm-1b/config.yaml
@@ -0,0 +1,30 @@
+description: "sesame-csm-1b for text-to-speech"
+model_name: sesame-csm-1b
+python_version: py310
+model_metadata:
+  repo_id: "sesame/csm-1b"
+  example_model_input:
+    text: "Hello from Sesame."
+    speaker: 0
+requirements:
+  - torch==2.4.0
+  - torchaudio==2.4.0
+  - tokenizers==0.21.0
+  - transformers==4.49.0
+  - huggingface_hub==0.28.1
+  - moshi==0.2.2
+  - torchtune==0.4.0
+  - torchao==0.9.0
+  - silentcipher @ git+https://github.com/SesameAILabs/silentcipher@d46d7d0893a583d8968ab3a6626e2289faec9152
+  - ffmpeg==1.4
+  - git+https://github.com/veerbia/csm.git@f747b9652b1c3859ce0f0465176ef8dff4cb32c8
+resources:
+  accelerator: T4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+secrets:
+  hf_access_token: null
+system_packages: []
+environment_variables: {}
+external_package_dirs: []
diff --git a/autodesk-wala/packages/src/mvdream/ldm/__init__.py b/audio/sesame-csm-1b/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/__init__.py
rename to audio/sesame-csm-1b/model/__init__.py
diff --git a/sesame-csm-1b/model/model.py b/audio/sesame-csm-1b/model/model.py
similarity index 100%
rename from sesame-csm-1b/model/model.py
rename to audio/sesame-csm-1b/model/model.py
diff --git a/audio/ultravox/README.md b/audio/ultravox/README.md
new file mode 100644
index 000000000..0fb9e631d
--- /dev/null
+++ b/audio/ultravox/README.md
@@ -0,0 +1,31 @@
+# Ultravox v0.2
+
+Deploy Ultravox v0.2 for audio understanding on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Audio understanding |
+| Engine | vLLM |
+| GPU | A100 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "fixie-ai/ultravox-v0.2", "messages": [{"role": "user", "content": "Describe this audio."}]}'
+```
+
+## Configuration highlights
+
+- Base image: `vshulman/vllm-openai-fixie:latest`
+- Predict concurrency: **512**
+- System packages: `python3.10-venv`
diff --git a/audio/ultravox/config.yaml b/audio/ultravox/config.yaml
new file mode 100644
index 000000000..b1b05a8db
--- /dev/null
+++ b/audio/ultravox/config.yaml
@@ -0,0 +1,24 @@
+description: "Ultravox v0.2 for audio understanding"
+base_image:
+  image: vshulman/vllm-openai-fixie:latest
+  python_executable_path: /usr/bin/python3
+model_metadata:
+  repo_id: "fixie-ai/ultravox-v0_4_1-llama-3_1-8b"
+  example_model_input: {"messages": [{"role": "user", "content": "Describe the audio clip."}], "model": "fixie-ai/ultravox-v0.2", "max_tokens": 512}
+  arguments:
+    model: fixie-ai/ultravox-v0.2
+    audio_token_id: 128002
+environment_variables: {}
+external_package_dirs: []
+model_name: Ultravox v0.2
+python_version: py310
+runtime:
+  predict_concurrency: 512
+requirements:
+  - httpx==0.27.0
+resources:
+  accelerator: A100
+  use_gpu: true
+secrets: {}
+system_packages:
+- python3.10-venv
diff --git a/autodesk-wala/packages/src/mvdream/ldm/models/__init__.py b/audio/ultravox/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/models/__init__.py
rename to audio/ultravox/model/__init__.py
diff --git a/ultravox/model/model.py b/audio/ultravox/model/model.py
similarity index 100%
rename from ultravox/model/model.py
rename to audio/ultravox/model/model.py
diff --git a/audio/voxtral-streaming-4b/README.md b/audio/voxtral-streaming-4b/README.md
new file mode 100644
index 000000000..89084fd99
--- /dev/null
+++ b/audio/voxtral-streaming-4b/README.md
@@ -0,0 +1,46 @@
+# Voxtral-Mini-4B-Realtime-2602
+
+Deploy Voxtral-Mini-4B-Realtime-2602 on Baseten using a vLLM engine.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Voxtral-Mini-4B-Realtime-2602](https://huggingface.co/mistralai/Voxtral-Mini-4B-Realtime-2602) |
+| Task | Audio |
+| Engine | vLLM |
+| GPU | H100_40GB:1 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model uses a WebSocket endpoint for realtime audio streaming. Use the included `streaming.py` client:
+
+```sh
+python streaming.py
+```
+
+Or connect directly via WebSocket:
+
+```python
+import websockets
+
+async with websockets.connect(
+    "wss://model-<model_id>.api.baseten.co/environments/production/websocket",
+    extra_headers={"Authorization": "Api-Key YOUR_BASETEN_API_KEY"}
+) as ws:
+    await ws.send('{"type": "session.update", "model": "mistralai/Voxtral-Mini-4B-Realtime-2602"}')
+    await ws.send('{"type": "input_audio_buffer.append", "audio": "<base64_encoded_pcm16_audio>"}')
+    await ws.send('{"type": "input_audio_buffer.commit"}')
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:nightly-d88a1df699f68e5284fe3a3170f8ae292a3e9c3f`
+- System packages: `python3.10-venv, ffmpeg, openmpi-bin, libopenmpi-dev`
+- Environment variables: `VLLM_DISABLE_COMPILE_CACHE`
diff --git a/audio/voxtral-streaming-4b/config.yaml b/audio/voxtral-streaming-4b/config.yaml
new file mode 100644
index 000000000..b38af3abf
--- /dev/null
+++ b/audio/voxtral-streaming-4b/config.yaml
@@ -0,0 +1,42 @@
+description: "Voxtral Mini 4B for real-time speech interaction"
+model_metadata:
+  repo_id: "mistralai/Voxtral-Mini-4B-Realtime-2602"
+  example_model_input: {"note": "This model uses a WebSocket endpoint at /v1/realtime. Connect via WebSocket and send audio frames for real-time speech interaction."}
+model_name: Voxtral-Mini-4B-Realtime-2602
+secrets:
+  hf_access_token: null
+environment_variables:
+  VLLM_DISABLE_COMPILE_CACHE: "1"
+base_image:
+  image: vllm/vllm-openai:nightly-d88a1df699f68e5284fe3a3170f8ae292a3e9c3f
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) VLLM_DISABLE_COMPILE_CACHE=1 vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation-config '{\"cudagraph_mode\":\"PIECEWISE\"}' --host 0.0.0.0 --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/realtime
+  server_port: 8000
+resources:
+  accelerator: H100_40GB:1
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+requirements:
+  - --pre --extra-index-url https://wheels.vllm.ai/nightly
+  - vllm[audio]
+  - librosa==0.10.2
+  - torch==2.5.1
+  - torchaudio==2.5.1
+  - pynvml==11.5.3
+  - ffmpeg-python==0.2.0
+  - websockets==13.1
+system_packages:
+  - python3.10-venv
+  - ffmpeg
+  - openmpi-bin
+  - libopenmpi-dev
+runtime:
+  is_websocket_endpoint: true
+  transport:
+    kind: websocket
+    ping_interval_seconds: null
+    ping_timeout_seconds: null
diff --git a/mistral/voxtral-streaming-4b/streaming.py b/audio/voxtral-streaming-4b/streaming.py
similarity index 100%
rename from mistral/voxtral-streaming-4b/streaming.py
rename to audio/voxtral-streaming-4b/streaming.py
diff --git a/audio/whisper/faster-whisper-small/README.md b/audio/whisper/faster-whisper-small/README.md
new file mode 100644
index 000000000..480cae3e9
--- /dev/null
+++ b/audio/whisper/faster-whisper-small/README.md
@@ -0,0 +1,32 @@
+# Faster Whisper Small
+
+A small speech-to-text model for multi-lingual audio transcription.
+
+| Property | Value |
+|----------|-------|
+| Model | [Systran/faster-whisper-small](https://huggingface.co/Systran/faster-whisper-small) |
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/whisper/faster-whisper-small/config.yaml b/audio/whisper/faster-whisper-small/config.yaml
similarity index 100%
rename from whisper/faster-whisper-small/config.yaml
rename to audio/whisper/faster-whisper-small/config.yaml
diff --git a/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/__init__.py b/audio/whisper/faster-whisper-small/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/models/diffusion/__init__.py
rename to audio/whisper/faster-whisper-small/model/__init__.py
diff --git a/whisper/faster-whisper-small/model/model.py b/audio/whisper/faster-whisper-small/model/model.py
similarity index 100%
rename from whisper/faster-whisper-small/model/model.py
rename to audio/whisper/faster-whisper-small/model/model.py
diff --git a/audio/whisper/faster-whisper-v2/README.md b/audio/whisper/faster-whisper-v2/README.md
new file mode 100644
index 000000000..a1b7ed1e3
--- /dev/null
+++ b/audio/whisper/faster-whisper-v2/README.md
@@ -0,0 +1,32 @@
+# Faster Whisper v2
+
+Faster Whisper v2
+
+| Property | Value |
+|----------|-------|
+| Model | [Systran/faster-whisper-large-v2](https://huggingface.co/Systran/faster-whisper-large-v2) |
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/whisper/faster-whisper-v2/config.yaml b/audio/whisper/faster-whisper-v2/config.yaml
similarity index 100%
rename from whisper/faster-whisper-v2/config.yaml
rename to audio/whisper/faster-whisper-v2/config.yaml
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/__init__.py b/audio/whisper/faster-whisper-v2/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/__init__.py
rename to audio/whisper/faster-whisper-v2/model/__init__.py
diff --git a/whisper/faster-whisper-v2/model/model.py b/audio/whisper/faster-whisper-v2/model/model.py
similarity index 100%
rename from whisper/faster-whisper-v2/model/model.py
rename to audio/whisper/faster-whisper-v2/model/model.py
diff --git a/audio/whisper/faster-whisper-v3/README.md b/audio/whisper/faster-whisper-v3/README.md
new file mode 100644
index 000000000..a722a06de
--- /dev/null
+++ b/audio/whisper/faster-whisper-v3/README.md
@@ -0,0 +1,32 @@
+# Faster Whisper v3
+
+Faster Whisper v3
+
+| Property | Value |
+|----------|-------|
+| Model | [Systran/faster-whisper-large-v3](https://huggingface.co/Systran/faster-whisper-large-v3) |
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/whisper/faster-whisper-v3/config.yaml b/audio/whisper/faster-whisper-v3/config.yaml
similarity index 100%
rename from whisper/faster-whisper-v3/config.yaml
rename to audio/whisper/faster-whisper-v3/config.yaml
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/__init__.py b/audio/whisper/faster-whisper-v3/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/__init__.py
rename to audio/whisper/faster-whisper-v3/model/__init__.py
diff --git a/whisper/faster-whisper-v3/model/model.py b/audio/whisper/faster-whisper-v3/model/model.py
similarity index 100%
rename from whisper/faster-whisper-v3/model/model.py
rename to audio/whisper/faster-whisper-v3/model/model.py
diff --git a/gfp-gan/LICENSE b/audio/whisper/whisper-openai/LICENSE
similarity index 100%
rename from gfp-gan/LICENSE
rename to audio/whisper/whisper-openai/LICENSE
diff --git a/audio/whisper/whisper-openai/README.md b/audio/whisper/whisper-openai/README.md
new file mode 100644
index 000000000..121dd66cc
--- /dev/null
+++ b/audio/whisper/whisper-openai/README.md
@@ -0,0 +1,31 @@
+# Whisper
+
+Transcribe audio files across multiple languages.
+
+| Property | Value |
+|----------|-------|
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/audio/whisper/whisper-openai/config.yaml b/audio/whisper/whisper-openai/config.yaml
new file mode 100644
index 000000000..a43e837a5
--- /dev/null
+++ b/audio/whisper/whisper-openai/config.yaml
@@ -0,0 +1,28 @@
+description: Transcribe audio files across multiple languages.
+environment_variables: {}
+external_data:
+- local_data_path: models/small.pt
+  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/whisper/small.pt
+external_package_dirs: []
+model_metadata:
+  repo_id: "openai/whisper-small"
+  avatar_url: https://cdn.baseten.co/production/static/openai.png
+  cover_image_url: https://cdn.baseten.co/production/static/whisper.png
+  example_model_input:
+    url: https://cdn.baseten.co/docs/production/Gettysburg.mp3
+  pretty_name: Whisper
+  tags:
+  - speech-recognition
+model_name: Whisper
+python_version: py39
+requirements:
+- openai-whisper==20250625
+- torch==2.0.1
+resources:
+  accelerator: A10G
+  cpu: '4'
+  memory: 16Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/gfp-gan/data/.gitkeep b/audio/whisper/whisper-openai/data/.gitkeep
similarity index 100%
rename from gfp-gan/data/.gitkeep
rename to audio/whisper/whisper-openai/data/.gitkeep
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/distributions/__init__.py b/audio/whisper/whisper-openai/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/distributions/__init__.py
rename to audio/whisper/whisper-openai/model/__init__.py
diff --git a/whisper/whisper-truss/model/model.py b/audio/whisper/whisper-openai/model/model.py
similarity index 100%
rename from whisper/whisper-truss/model/model.py
rename to audio/whisper/whisper-openai/model/model.py
diff --git a/audio/whisper/whisper-streaming/README.md b/audio/whisper/whisper-streaming/README.md
new file mode 100644
index 000000000..71ff006f1
--- /dev/null
+++ b/audio/whisper/whisper-streaming/README.md
@@ -0,0 +1,30 @@
+# Whisper Streaming
+
+Deploy Whisper Streaming for speech-to-text transcription on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"url": "https://example.com/audio.wav"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/truss-server-base:3.10-gpu-v0.4.9`
+- System packages: `ffmpeg`
diff --git a/audio/whisper/whisper-streaming/config.yaml b/audio/whisper/whisper-streaming/config.yaml
new file mode 100644
index 000000000..0f55b7554
--- /dev/null
+++ b/audio/whisper/whisper-streaming/config.yaml
@@ -0,0 +1,20 @@
+description: "Whisper Streaming for speech-to-text transcription"
+base_image:
+  image: baseten/truss-server-base:3.10-gpu-v0.4.9
+  python_executable_path: /usr/bin/python3
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "openai/whisper-medium"
+  example_model_input: {"audio": "<base64_encoded_audio_wav>", "chunk_size": 1.0}
+  whisper_model: medium
+model_name: Whisper Streaming
+python_version: py310
+requirements: []
+requirements_file: ./requirements.txt
+resources:
+  accelerator: T4
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/encoders/__init__.py b/audio/whisper/whisper-streaming/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/encoders/__init__.py
rename to audio/whisper/whisper-streaming/model/__init__.py
diff --git a/whisper/whisper-streaming/model/model.py b/audio/whisper/whisper-streaming/model/model.py
similarity index 100%
rename from whisper/whisper-streaming/model/model.py
rename to audio/whisper/whisper-streaming/model/model.py
diff --git a/whisper/whisper-streaming/packages/whisper_streaming/line_packet.py b/audio/whisper/whisper-streaming/packages/whisper_streaming/line_packet.py
similarity index 100%
rename from whisper/whisper-streaming/packages/whisper_streaming/line_packet.py
rename to audio/whisper/whisper-streaming/packages/whisper_streaming/line_packet.py
diff --git a/whisper/whisper-streaming/packages/whisper_streaming/whisper_online.py b/audio/whisper/whisper-streaming/packages/whisper_streaming/whisper_online.py
similarity index 100%
rename from whisper/whisper-streaming/packages/whisper_streaming/whisper_online.py
rename to audio/whisper/whisper-streaming/packages/whisper_streaming/whisper_online.py
diff --git a/whisper/whisper-streaming/packages/whisper_streaming/whisper_online_server.py b/audio/whisper/whisper-streaming/packages/whisper_streaming/whisper_online_server.py
similarity index 100%
rename from whisper/whisper-streaming/packages/whisper_streaming/whisper_online_server.py
rename to audio/whisper/whisper-streaming/packages/whisper_streaming/whisper_online_server.py
diff --git a/whisper/whisper-streaming/requirements.txt b/audio/whisper/whisper-streaming/requirements.txt
similarity index 100%
rename from whisper/whisper-streaming/requirements.txt
rename to audio/whisper/whisper-streaming/requirements.txt
diff --git a/audio/whisper/whisper-v3-base64/README.md b/audio/whisper/whisper-v3-base64/README.md
new file mode 100644
index 000000000..8027d94b2
--- /dev/null
+++ b/audio/whisper/whisper-v3-base64/README.md
@@ -0,0 +1,31 @@
+# Whisper V3 Base64 Input
+
+Transcribe audio files across multiple languages.
+
+| Property | Value |
+|----------|-------|
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/audio/whisper/whisper-v3-base64/config.yaml b/audio/whisper/whisper-v3-base64/config.yaml
new file mode 100644
index 000000000..efadf3019
--- /dev/null
+++ b/audio/whisper/whisper-v3-base64/config.yaml
@@ -0,0 +1,25 @@
+description: Transcribe audio files across multiple languages.
+environment_variables: {}
+external_data:
+- local_data_path: weights/large-v3.pt
+  url: https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt
+external_package_dirs: []
+model_metadata:
+  repo_id: "openai/whisper-large-v3"
+  avatar_url: https://cdn.baseten.co/production/static/openai.png
+  cover_image_url: https://cdn.baseten.co/production/static/whisper.png
+  example_model_input:
+    url: https://cdn.baseten.co/docs/production/Gettysburg.mp3
+model_name: Whisper V3 Base64 Input
+python_version: py310
+requirements:
+- torch==2.0.1
+- openai-whisper==20250625
+resources:
+  accelerator: T4
+  cpu: '3'
+  memory: 16Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/autodesk-wala/packages/src/utils/__init__.py b/audio/whisper/whisper-v3-base64/model/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/utils/__init__.py
rename to audio/whisper/whisper-v3-base64/model/__init__.py
diff --git a/whisper/whisper-v3-truss-base64/model/model.py b/audio/whisper/whisper-v3-base64/model/model.py
similarity index 100%
rename from whisper/whisper-v3-truss-base64/model/model.py
rename to audio/whisper/whisper-v3-base64/model/model.py
diff --git a/audio/whisper/whisper-v3/README.md b/audio/whisper/whisper-v3/README.md
new file mode 100644
index 000000000..649772e18
--- /dev/null
+++ b/audio/whisper/whisper-v3/README.md
@@ -0,0 +1,31 @@
+# Whisper V3
+
+Transcribe audio files across multiple languages.
+
+| Property | Value |
+|----------|-------|
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/audio/whisper/whisper-v3/config.yaml b/audio/whisper/whisper-v3/config.yaml
new file mode 100644
index 000000000..880187efa
--- /dev/null
+++ b/audio/whisper/whisper-v3/config.yaml
@@ -0,0 +1,26 @@
+description: Transcribe audio files across multiple languages.
+environment_variables: {}
+external_data:
+- local_data_path: weights/large-v3.pt
+  url: https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt
+external_package_dirs: []
+model_metadata:
+  repo_id: "openai/whisper-large-v3"
+  avatar_url: https://cdn.baseten.co/production/static/openai.png
+  cover_image_url: https://cdn.baseten.co/production/static/whisper.png
+  example_model_input:
+    url: https://cdn.baseten.co/docs/production/Gettysburg.mp3
+model_name: Whisper V3
+python_version: py310
+requirements:
+- torch==2.4.1
+- openai-whisper==20250625
+- ffmpeg-python==0.2.0
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 16Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/binocular/model/__init__.py b/audio/whisper/whisper-v3/model/__init__.py
similarity index 100%
rename from binocular/model/__init__.py
rename to audio/whisper/whisper-v3/model/__init__.py
diff --git a/whisper/whisper-v3-truss/model/model.py b/audio/whisper/whisper-v3/model/model.py
similarity index 100%
rename from whisper/whisper-v3-truss/model/model.py
rename to audio/whisper/whisper-v3/model/model.py
diff --git a/audio/whisper/whisperx/README.md b/audio/whisper/whisperx/README.md
new file mode 100644
index 000000000..e9b826ee9
--- /dev/null
+++ b/audio/whisper/whisperx/README.md
@@ -0,0 +1,34 @@
+# whisperX
+
+Deploy whisperX for speech-to-text transcription on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Speech-to-text |
+| Engine | Custom (Truss) |
+| GPU | L4 |
+| Python | py310 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "audio_file": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- Base image: `runpod/pytorch:2.1.1-py3.10-cuda12.1.1-devel-ubuntu22.04`
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/audio/whisper/whisperx/config.yaml b/audio/whisper/whisperx/config.yaml
new file mode 100644
index 000000000..05651f74d
--- /dev/null
+++ b/audio/whisper/whisperx/config.yaml
@@ -0,0 +1,37 @@
+description: "whisperX for speech-to-text transcription"
+base_image:
+  image: runpod/pytorch:2.1.1-py3.10-cuda12.1.1-devel-ubuntu22.04
+  python_executable_path: /usr/bin/python
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "openai/whisper-large-v3"
+  example_model_input:
+    audio_file: https://cdn.baseten.co/docs/production/Gettysburg.mp3
+model_name: whisperX
+python_version: py310
+requirements:
+- --extra-index-url https://download.pytorch.org/whl/cu121
+- git+https://github.com/m-bain/whisperx.git@734084cdf6f624bc33ed9f0cfcaa82840707ba6f
+- torch==2.2.0
+- torchaudio==2.2.0
+- transformers==4.48.3
+- torchvision==0.17.0
+- ffmpeg-python==0.2.0
+- faster-whisper==1.1.0
+- pandas==2.2.3
+- nltk==3.9.1
+- setuptools==68.0.0
+- ctranslate2==4.4.0
+- pydub==0.25.1
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 4Gi
+  use_gpu: true
+secrets:
+  hf_access_token: null
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/chatterbox-tts/model/__init__.py b/audio/whisper/whisperx/model/__init__.py
similarity index 100%
rename from chatterbox-tts/model/__init__.py
rename to audio/whisper/whisperx/model/__init__.py
diff --git a/whisper/whisperx-truss/model/model.py b/audio/whisper/whisperx/model/model.py
similarity index 100%
rename from whisper/whisperx-truss/model/model.py
rename to audio/whisper/whisperx/model/model.py
diff --git a/audio/xtts-streaming/README.md b/audio/xtts-streaming/README.md
new file mode 100644
index 000000000..0121118b6
--- /dev/null
+++ b/audio/xtts-streaming/README.md
@@ -0,0 +1,29 @@
+# XTTS Streaming - High Performance
+
+Deploy XTTS Streaming - High Performance for text-to-speech on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text-to-speech |
+| Engine | Custom (Truss) |
+| GPU | H100 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"text": "Hello, this is a test of text to speech."}'
+```
+
+## Configuration highlights
+
+- Base image: `htrivedi05/xtts-streaming`
+- Environment variables: `COQUI_TOS_AGREED`
diff --git a/audio/xtts-streaming/config.yaml b/audio/xtts-streaming/config.yaml
new file mode 100644
index 000000000..1a7e25a0b
--- /dev/null
+++ b/audio/xtts-streaming/config.yaml
@@ -0,0 +1,17 @@
+description: "XTTS Streaming - High Performance for text-to-speech"
+base_image:
+  image: htrivedi05/xtts-streaming
+  python_executable_path: /opt/conda/bin/python
+environment_variables:
+  COQUI_TOS_AGREED: '1'
+external_package_dirs: []
+model_metadata:
+  repo_id: "coqui/XTTS-v2"
+  example_model_input: {"text": "Hello, how are you today?", "language": "en"}
+model_name: XTTS Streaming - High Performance
+resources:
+  accelerator: H100
+  cpu: '3'
+  memory: 10Gi
+  use_gpu: true
+secrets: {}
diff --git a/clip/model/__init__.py b/audio/xtts-streaming/model/__init__.py
similarity index 100%
rename from clip/model/__init__.py
rename to audio/xtts-streaming/model/__init__.py
diff --git a/xtts-streaming/model/model.py b/audio/xtts-streaming/model/model.py
similarity index 100%
rename from xtts-streaming/model/model.py
rename to audio/xtts-streaming/model/model.py
diff --git a/xtts-streaming/requirements.txt b/audio/xtts-streaming/requirements.txt
similarity index 100%
rename from xtts-streaming/requirements.txt
rename to audio/xtts-streaming/requirements.txt
diff --git a/audio/xtts-v2/README.md b/audio/xtts-v2/README.md
new file mode 100644
index 000000000..5fb0fb6f5
--- /dev/null
+++ b/audio/xtts-v2/README.md
@@ -0,0 +1,33 @@
+# XTTS V2
+
+Deploy XTTS V2 for text-to-speech on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text-to-speech |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "language": "en",
+  "speaker_voice": "Claribel Dervla",
+  "text": "Kurt watched the incoming Pelicans. The blocky jet-powered craft were so distant they were only specks against the setting sun. He hit the magnification on his faceplate and saw lines of fire tracing their reentry vectors. They would touch down in three minutes."
+}'
+```
+
+## Configuration highlights
+
+- Environment variables: `COQUI_TOS_AGREED`
diff --git a/audio/xtts-v2/config.yaml b/audio/xtts-v2/config.yaml
new file mode 100644
index 000000000..34d72877b
--- /dev/null
+++ b/audio/xtts-v2/config.yaml
@@ -0,0 +1,23 @@
+description: "XTTS V2 for text-to-speech"
+environment_variables:
+  COQUI_TOS_AGREED: "1"
+external_package_dirs: []
+model_metadata:
+  repo_id: "coqui/XTTS-v2"
+  example_model_input:
+    language: en
+    speaker_voice: Claribel Dervla
+    text: Kurt watched the incoming Pelicans. The blocky jet-powered craft were so distant they were only specks against the setting sun. He hit the magnification on his faceplate and saw lines of fire tracing their reentry vectors. They would touch down in three minutes.
+  tags:
+  - text-to-speech
+model_name: XTTS V2
+python_version: py310
+requirements:
+  - git+https://github.com/htrivedi99/TTS.git
+resources:
+  accelerator: T4
+  cpu: '3'
+  memory: 10Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/cogvlm/model/__init__.py b/audio/xtts-v2/model/__init__.py
similarity index 100%
rename from cogvlm/model/__init__.py
rename to audio/xtts-v2/model/__init__.py
diff --git a/xtts-v2-truss/model/model.py b/audio/xtts-v2/model/model.py
similarity index 100%
rename from xtts-v2-truss/model/model.py
rename to audio/xtts-v2/model/model.py
diff --git a/audiogen-medium/README.md b/audiogen-medium/README.md
deleted file mode 100644
index 6822e4eb6..000000000
--- a/audiogen-medium/README.md
+++ /dev/null
@@ -1,82 +0,0 @@
-# AudioGen Truss
-
-This repository packages [AudioGen](https://github.com/facebookresearch/audiocraft/) as a [Truss](https://truss.baseten.co).
-
-AudioGen is a simple and controllable model for audio generation developed by Facebook AI Research.
-
-## Deploying AudioGen
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd audiogen-medium-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `audiogen-medium-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-We found this model runs reasonably fast on A10Gs; you can configure the hardware you'd like in the config.yaml.
-
-```yaml
----
-resources:
-  cpu: "3"
-  memory: 14Gi
-  use_gpu: true
-  accelerator: A10G
-```
-
-## Invoking AudioGen
-
-AudioGen takes a list of prompts and a duration in seconds. It will generate one clip per prompt and return each clip as a base64 encoded WAV file.
-
-```sh
-truss predict -d '{"prompts": ['dog barking', 'sirene of an emergency vehicle', 'footsteps in a corridor'], "duration": 8}'
-```
-
-```python
-
-import json
-import base64
-import os, sys
-
-model_output = json.loads(sys.stdin.read())
-
-for idx, clip in enumerate(model_output["data"]):
-  with open(f"clip_{idx}.wav", "wb") as f:
-    f.write(base64.b64decode(clip))
-```
-
-You can also invoke your model via a REST API
-
-```
-curl -X POST " https://app.baseten.co/models/YOUR_MODEL_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompts": ["happy rock" "energetic EDM", "sad jazz"], "duration": 8
-         }'
-```
-
-## Model sizes
-
-AudioGen comes in 1 size:
-
-- `medium`: 1.5B model
-
-which is the model in this truss.
diff --git a/audiogen-medium/config.yaml b/audiogen-medium/config.yaml
deleted file mode 100644
index 7eaad77a3..000000000
--- a/audiogen-medium/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-description: AudioGen is a simple and controllable model for audio generation developed
-  by Facebook AI Research.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/musicgen-cover.png
-  example_model_input:
-    duration: 8
-    prompts:
-    - dog barking
-    - sirene of an emergency vehicle
-    - footsteps in a corridor
-  tags:
-  - text-to-audio
-model_name: AudioGen medium
-python_version: py39
-requirements:
-- torch>=2
-- git+https://github.com/facebookresearch/audiocraft.git
-- torchaudio
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/autodesk-wala/README.md b/autodesk-wala/README.md
deleted file mode 100644
index 34bd98be6..000000000
--- a/autodesk-wala/README.md
+++ /dev/null
@@ -1,75 +0,0 @@
-## Autodesk WaLa (single‑view image → 3D) Truss
-
-- Status: supports only `ADSKAILab/WaLa-SV-1B` (single‑view) at the moment
-- Output: OBJ (default) or SDF, base64‑encoded in the response
-- License: non‑commercial per the model card
-
-### Prerequisites
-- Baseten account and API key.
-- Hugging Face access:
-  - Accept the model license on `ADSKAILab/WaLa-SV-1B`.
-  - Add a Baseten secret named `hf_access_token` with your token (read access).
-
-### Deploy
-From this folder (`autodesk-wala`):
-
-```bash
-# Option A: Truss CLI
-truss push --publish
-```
-
-Notes:
-- You can use the client code in `autodesk-wala/test.py` to call your endpoint and save the OBJ locally.
-- WaLa source code is vendored in `packages/src/` (run `./vendor_wala.sh` to set up if missing).
-- The HF token is read from the `hf_access_token` secret and exported for `hf_hub_download`.
-
-### Invoke (Python)
-A simple client is provided in `test.py`. It reads `examples/single_view/table.png`, sends it to your deployed endpoint, saves `output.obj`, and displays it if `trimesh` + `plotly` are installed.
-
-```bash
-export BASETEN_API_KEY=...  # your Baseten API key
-python autodesk-wala/test.py
-```
-
-### Invoke (cURL)
-Replace `model_id` and `API_KEY` with your values.
-
-```bash
-IMG_B64=$(base64 -i autodesk-wala/examples/single_view/table.png)
-
-curl -s -X POST "https://model-<model_id>.api.baseten.co/production/predict" \
-  -H "Authorization: Api-Key <API_KEY>" \
-  -H "Content-Type: application/json" \
-  -d "{\
-    \"image_b64\": \"${IMG_B64}\",\
-    \"model_name\": \"ADSKAILab/WaLa-SV-1B\",\
-    \"output_format\": \"obj\",\
-    \"scale\": 1.8,\
-    \"diffusion_rescale_timestep\": 5,\
-    \"seed\": 42\
-  }" | tee response.json
-
-# Decode OBJ
-jq -r .obj_b64 response.json | base64 --decode > output.obj
-```
-
-### Request schema
-- Required:
-  - `image_b64`: base64‑encoded RGB image (single view)
-- Optional:
-  - `model_name`: HF repo id (defaults to `ADSKAILab/WaLa-SV-1B`)
-  - `output_format`: `obj` (default) or `sdf`
-  - `scale`: float, default `3.0` (authors often use `1.8`)
-  - `diffusion_rescale_timestep`: int, default `100` (authors often use `5`)
-  - `seed`: int, default `42`
-  - `target_num_faces`: int, optional mesh simplification target (server best‑effort; skipped if Open3D/libGL unavailable). For Colab‑like behavior, omit this.
-
-### Response
-- Success:
-  - `obj_b64` or `sdf_b64`: base64 of the generated asset
-  - `data`: same base64 payload (for convenience)
-  - `format`: `obj` or `sdf`
-  - `output_path`: server‑side path where the file was written
-  - `time`: seconds
-- Error:
-  - `{ "status": "error", "message": "..." }`
diff --git a/autodesk-wala/config.yaml b/autodesk-wala/config.yaml
deleted file mode 100644
index ba3cff85f..000000000
--- a/autodesk-wala/config.yaml
+++ /dev/null
@@ -1,8 +0,0 @@
-model_name: ADSKAILab/WaLa-SV-1B
-python_version: py311
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-requirements_file: ./requirements.txt # should cover all the dependencies for WaLa, taken from repo
-secrets:
-  hf_access_token: null
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/model.py b/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/model.py
deleted file mode 100644
index 2e0377430..000000000
--- a/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/model.py
+++ /dev/null
@@ -1,1020 +0,0 @@
-# pytorch_diffusion + derived encoder decoder
-import math
-import torch
-import torch.nn as nn
-import numpy as np
-from einops import rearrange
-from typing import Optional, Any
-
-from ..attention import MemoryEfficientCrossAttention
-
-try:
-    import xformers
-    import xformers.ops
-
-    XFORMERS_IS_AVAILBLE = True
-except:
-    XFORMERS_IS_AVAILBLE = False
-    print("No module 'xformers'. Proceeding without it.")
-
-
-def get_timestep_embedding(timesteps, embedding_dim):
-    """
-    This matches the implementation in Denoising Diffusion Probabilistic Models:
-    From Fairseq.
-    Build sinusoidal embeddings.
-    This matches the implementation in tensor2tensor, but differs slightly
-    from the description in Section 3.5 of "Attention Is All You Need".
-    """
-    assert len(timesteps.shape) == 1
-
-    half_dim = embedding_dim // 2
-    emb = math.log(10000) / (half_dim - 1)
-    # emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
-    emb = torch.exp(torch.arange(half_dim, dtype=torch.bfloat16) * -emb)
-    emb = emb.to(device=timesteps.device)
-    # emb = timesteps.float()[:, None] * emb[None, :]
-    emb = timesteps[:, None] * emb[None, :]
-    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
-    if embedding_dim % 2 == 1:  # zero pad
-        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
-    return emb
-
-
-def nonlinearity(x):
-    # swish
-    return x * torch.sigmoid(x)
-
-
-def Normalize(in_channels, num_groups=32):
-    return torch.nn.GroupNorm(
-        num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
-    )
-
-
-class Upsample(nn.Module):
-    def __init__(self, in_channels, with_conv):
-        super().__init__()
-        self.with_conv = with_conv
-        if self.with_conv:
-            self.conv = torch.nn.Conv2d(
-                in_channels, in_channels, kernel_size=3, stride=1, padding=1
-            )
-
-    def forward(self, x):
-        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
-        if self.with_conv:
-            x = self.conv(x)
-        return x
-
-
-class Downsample(nn.Module):
-    def __init__(self, in_channels, with_conv):
-        super().__init__()
-        self.with_conv = with_conv
-        if self.with_conv:
-            # no asymmetric padding in torch conv, must do it ourselves
-            self.conv = torch.nn.Conv2d(
-                in_channels, in_channels, kernel_size=3, stride=2, padding=0
-            )
-
-    def forward(self, x):
-        if self.with_conv:
-            pad = (0, 1, 0, 1)
-            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
-            x = self.conv(x)
-        else:
-            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
-        return x
-
-
-class ResnetBlock(nn.Module):
-    def __init__(
-        self,
-        *,
-        in_channels,
-        out_channels=None,
-        conv_shortcut=False,
-        dropout,
-        temb_channels=512,
-    ):
-        super().__init__()
-        self.in_channels = in_channels
-        out_channels = in_channels if out_channels is None else out_channels
-        self.out_channels = out_channels
-        self.use_conv_shortcut = conv_shortcut
-
-        self.norm1 = Normalize(in_channels)
-        self.conv1 = torch.nn.Conv2d(
-            in_channels, out_channels, kernel_size=3, stride=1, padding=1
-        )
-        if temb_channels > 0:
-            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
-        self.norm2 = Normalize(out_channels)
-        self.dropout = torch.nn.Dropout(dropout)
-        self.conv2 = torch.nn.Conv2d(
-            out_channels, out_channels, kernel_size=3, stride=1, padding=1
-        )
-        if self.in_channels != self.out_channels:
-            if self.use_conv_shortcut:
-                self.conv_shortcut = torch.nn.Conv2d(
-                    in_channels, out_channels, kernel_size=3, stride=1, padding=1
-                )
-            else:
-                self.nin_shortcut = torch.nn.Conv2d(
-                    in_channels, out_channels, kernel_size=1, stride=1, padding=0
-                )
-
-    def forward(self, x, temb):
-        h = x
-        h = self.norm1(h)
-        h = nonlinearity(h)
-        h = self.conv1(h)
-
-        if temb is not None:
-            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
-
-        h = self.norm2(h)
-        h = nonlinearity(h)
-        h = self.dropout(h)
-        h = self.conv2(h)
-
-        if self.in_channels != self.out_channels:
-            if self.use_conv_shortcut:
-                x = self.conv_shortcut(x)
-            else:
-                x = self.nin_shortcut(x)
-
-        return x + h
-
-
-class AttnBlock(nn.Module):
-    def __init__(self, in_channels):
-        super().__init__()
-        self.in_channels = in_channels
-
-        self.norm = Normalize(in_channels)
-        self.q = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.k = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.v = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.proj_out = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-
-    def forward(self, x):
-        h_ = x
-        h_ = self.norm(h_)
-        q = self.q(h_)
-        k = self.k(h_)
-        v = self.v(h_)
-
-        # compute attention
-        b, c, h, w = q.shape
-        q = q.reshape(b, c, h * w)
-        q = q.permute(0, 2, 1)  # b,hw,c
-        k = k.reshape(b, c, h * w)  # b,c,hw
-        w_ = torch.bmm(q, k)  # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
-        w_ = w_ * (int(c) ** (-0.5))
-        w_ = torch.nn.functional.softmax(w_, dim=2)
-
-        # attend to values
-        v = v.reshape(b, c, h * w)
-        w_ = w_.permute(0, 2, 1)  # b,hw,hw (first hw of k, second of q)
-        h_ = torch.bmm(v, w_)  # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
-        h_ = h_.reshape(b, c, h, w)
-
-        h_ = self.proj_out(h_)
-
-        return x + h_
-
-
-class MemoryEfficientAttnBlock(nn.Module):
-    """
-    Uses xformers efficient implementation,
-    see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
-    Note: this is a single-head self-attention operation
-    """
-
-    #
-    def __init__(self, in_channels):
-        super().__init__()
-        self.in_channels = in_channels
-
-        self.norm = Normalize(in_channels)
-        self.q = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.k = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.v = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.proj_out = torch.nn.Conv2d(
-            in_channels, in_channels, kernel_size=1, stride=1, padding=0
-        )
-        self.attention_op: Optional[Any] = None
-
-    def forward(self, x):
-        h_ = x
-        h_ = self.norm(h_)
-        q = self.q(h_)
-        k = self.k(h_)
-        v = self.v(h_)
-
-        # compute attention
-        B, C, H, W = q.shape
-        q, k, v = map(lambda x: rearrange(x, "b c h w -> b (h w) c"), (q, k, v))
-
-        q, k, v = map(
-            lambda t: t.unsqueeze(3)
-            .reshape(B, t.shape[1], 1, C)
-            .permute(0, 2, 1, 3)
-            .reshape(B * 1, t.shape[1], C)
-            .contiguous(),
-            (q, k, v),
-        )
-        out = xformers.ops.memory_efficient_attention(
-            q, k, v, attn_bias=None, op=self.attention_op
-        )
-
-        out = (
-            out.unsqueeze(0)
-            .reshape(B, 1, out.shape[1], C)
-            .permute(0, 2, 1, 3)
-            .reshape(B, out.shape[1], C)
-        )
-        out = rearrange(out, "b (h w) c -> b c h w", b=B, h=H, w=W, c=C)
-        out = self.proj_out(out)
-        return x + out
-
-
-class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
-    def forward(self, x, context=None, mask=None):
-        b, c, h, w = x.shape
-        x = rearrange(x, "b c h w -> b (h w) c")
-        out = super().forward(x, context=context, mask=mask)
-        out = rearrange(out, "b (h w) c -> b c h w", h=h, w=w, c=c)
-        return x + out
-
-
-def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
-    assert attn_type in [
-        "vanilla",
-        "vanilla-xformers",
-        "memory-efficient-cross-attn",
-        "linear",
-        "none",
-    ], f"attn_type {attn_type} unknown"
-    if XFORMERS_IS_AVAILBLE and attn_type == "vanilla":
-        attn_type = "vanilla-xformers"
-    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
-    if attn_type == "vanilla":
-        assert attn_kwargs is None
-        return AttnBlock(in_channels)
-    elif attn_type == "vanilla-xformers":
-        print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...")
-        return MemoryEfficientAttnBlock(in_channels)
-    elif type == "memory-efficient-cross-attn":
-        attn_kwargs["query_dim"] = in_channels
-        return MemoryEfficientCrossAttentionWrapper(**attn_kwargs)
-    elif attn_type == "none":
-        return nn.Identity(in_channels)
-    else:
-        raise NotImplementedError()
-
-
-class Model(nn.Module):
-    def __init__(
-        self,
-        *,
-        ch,
-        out_ch,
-        ch_mult=(1, 2, 4, 8),
-        num_res_blocks,
-        attn_resolutions,
-        dropout=0.0,
-        resamp_with_conv=True,
-        in_channels,
-        resolution,
-        use_timestep=True,
-        use_linear_attn=False,
-        attn_type="vanilla",
-    ):
-        super().__init__()
-        if use_linear_attn:
-            attn_type = "linear"
-        self.ch = ch
-        self.temb_ch = self.ch * 4
-        self.num_resolutions = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        self.resolution = resolution
-        self.in_channels = in_channels
-
-        self.use_timestep = use_timestep
-        if self.use_timestep:
-            # timestep embedding
-            self.temb = nn.Module()
-            self.temb.dense = nn.ModuleList(
-                [
-                    torch.nn.Linear(self.ch, self.temb_ch),
-                    torch.nn.Linear(self.temb_ch, self.temb_ch),
-                ]
-            )
-
-        # downsampling
-        self.conv_in = torch.nn.Conv2d(
-            in_channels, self.ch, kernel_size=3, stride=1, padding=1
-        )
-
-        curr_res = resolution
-        in_ch_mult = (1,) + tuple(ch_mult)
-        self.down = nn.ModuleList()
-        for i_level in range(self.num_resolutions):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_in = ch * in_ch_mult[i_level]
-            block_out = ch * ch_mult[i_level]
-            for i_block in range(self.num_res_blocks):
-                block.append(
-                    ResnetBlock(
-                        in_channels=block_in,
-                        out_channels=block_out,
-                        temb_channels=self.temb_ch,
-                        dropout=dropout,
-                    )
-                )
-                block_in = block_out
-                if curr_res in attn_resolutions:
-                    attn.append(make_attn(block_in, attn_type=attn_type))
-            down = nn.Module()
-            down.block = block
-            down.attn = attn
-            if i_level != self.num_resolutions - 1:
-                down.downsample = Downsample(block_in, resamp_with_conv)
-                curr_res = curr_res // 2
-            self.down.append(down)
-
-        # middle
-        self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(
-            in_channels=block_in,
-            out_channels=block_in,
-            temb_channels=self.temb_ch,
-            dropout=dropout,
-        )
-        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
-        self.mid.block_2 = ResnetBlock(
-            in_channels=block_in,
-            out_channels=block_in,
-            temb_channels=self.temb_ch,
-            dropout=dropout,
-        )
-
-        # upsampling
-        self.up = nn.ModuleList()
-        for i_level in reversed(range(self.num_resolutions)):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_out = ch * ch_mult[i_level]
-            skip_in = ch * ch_mult[i_level]
-            for i_block in range(self.num_res_blocks + 1):
-                if i_block == self.num_res_blocks:
-                    skip_in = ch * in_ch_mult[i_level]
-                block.append(
-                    ResnetBlock(
-                        in_channels=block_in + skip_in,
-                        out_channels=block_out,
-                        temb_channels=self.temb_ch,
-                        dropout=dropout,
-                    )
-                )
-                block_in = block_out
-                if curr_res in attn_resolutions:
-                    attn.append(make_attn(block_in, attn_type=attn_type))
-            up = nn.Module()
-            up.block = block
-            up.attn = attn
-            if i_level != 0:
-                up.upsample = Upsample(block_in, resamp_with_conv)
-                curr_res = curr_res * 2
-            self.up.insert(0, up)  # prepend to get consistent order
-
-        # end
-        self.norm_out = Normalize(block_in)
-        self.conv_out = torch.nn.Conv2d(
-            block_in, out_ch, kernel_size=3, stride=1, padding=1
-        )
-
-    def forward(self, x, t=None, context=None):
-        # assert x.shape[2] == x.shape[3] == self.resolution
-        if context is not None:
-            # assume aligned context, cat along channel axis
-            x = torch.cat((x, context), dim=1)
-        if self.use_timestep:
-            # timestep embedding
-            assert t is not None
-            temb = get_timestep_embedding(t, self.ch)
-            temb = self.temb.dense[0](temb)
-            temb = nonlinearity(temb)
-            temb = self.temb.dense[1](temb)
-        else:
-            temb = None
-
-        # downsampling
-        hs = [self.conv_in(x)]
-        for i_level in range(self.num_resolutions):
-            for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](hs[-1], temb)
-                if len(self.down[i_level].attn) > 0:
-                    h = self.down[i_level].attn[i_block](h)
-                hs.append(h)
-            if i_level != self.num_resolutions - 1:
-                hs.append(self.down[i_level].downsample(hs[-1]))
-
-        # middle
-        h = hs[-1]
-        h = self.mid.block_1(h, temb)
-        h = self.mid.attn_1(h)
-        h = self.mid.block_2(h, temb)
-
-        # upsampling
-        for i_level in reversed(range(self.num_resolutions)):
-            for i_block in range(self.num_res_blocks + 1):
-                h = self.up[i_level].block[i_block](
-                    torch.cat([h, hs.pop()], dim=1), temb
-                )
-                if len(self.up[i_level].attn) > 0:
-                    h = self.up[i_level].attn[i_block](h)
-            if i_level != 0:
-                h = self.up[i_level].upsample(h)
-
-        # end
-        h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h)
-        return h
-
-    def get_last_layer(self):
-        return self.conv_out.weight
-
-
-class Encoder(nn.Module):
-    def __init__(
-        self,
-        *,
-        ch,
-        out_ch,
-        ch_mult=(1, 2, 4, 8),
-        num_res_blocks,
-        attn_resolutions,
-        dropout=0.0,
-        resamp_with_conv=True,
-        in_channels,
-        resolution,
-        z_channels,
-        double_z=True,
-        use_linear_attn=False,
-        attn_type="vanilla",
-        **ignore_kwargs,
-    ):
-        super().__init__()
-        if use_linear_attn:
-            attn_type = "linear"
-        self.ch = ch
-        self.temb_ch = 0
-        self.num_resolutions = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        self.resolution = resolution
-        self.in_channels = in_channels
-
-        # downsampling
-        self.conv_in = torch.nn.Conv2d(
-            in_channels, self.ch, kernel_size=3, stride=1, padding=1
-        )
-
-        curr_res = resolution
-        in_ch_mult = (1,) + tuple(ch_mult)
-        self.in_ch_mult = in_ch_mult
-        self.down = nn.ModuleList()
-        for i_level in range(self.num_resolutions):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_in = ch * in_ch_mult[i_level]
-            block_out = ch * ch_mult[i_level]
-            for i_block in range(self.num_res_blocks):
-                block.append(
-                    ResnetBlock(
-                        in_channels=block_in,
-                        out_channels=block_out,
-                        temb_channels=self.temb_ch,
-                        dropout=dropout,
-                    )
-                )
-                block_in = block_out
-                if curr_res in attn_resolutions:
-                    attn.append(make_attn(block_in, attn_type=attn_type))
-            down = nn.Module()
-            down.block = block
-            down.attn = attn
-            if i_level != self.num_resolutions - 1:
-                down.downsample = Downsample(block_in, resamp_with_conv)
-                curr_res = curr_res // 2
-            self.down.append(down)
-
-        # middle
-        self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(
-            in_channels=block_in,
-            out_channels=block_in,
-            temb_channels=self.temb_ch,
-            dropout=dropout,
-        )
-        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
-        self.mid.block_2 = ResnetBlock(
-            in_channels=block_in,
-            out_channels=block_in,
-            temb_channels=self.temb_ch,
-            dropout=dropout,
-        )
-
-        # end
-        self.norm_out = Normalize(block_in)
-        self.conv_out = torch.nn.Conv2d(
-            block_in,
-            2 * z_channels if double_z else z_channels,
-            kernel_size=3,
-            stride=1,
-            padding=1,
-        )
-
-    def forward(self, x):
-        # timestep embedding
-        temb = None
-
-        # downsampling
-        hs = [self.conv_in(x)]
-        for i_level in range(self.num_resolutions):
-            for i_block in range(self.num_res_blocks):
-                h = self.down[i_level].block[i_block](hs[-1], temb)
-                if len(self.down[i_level].attn) > 0:
-                    h = self.down[i_level].attn[i_block](h)
-                hs.append(h)
-            if i_level != self.num_resolutions - 1:
-                hs.append(self.down[i_level].downsample(hs[-1]))
-
-        # middle
-        h = hs[-1]
-        h = self.mid.block_1(h, temb)
-        h = self.mid.attn_1(h)
-        h = self.mid.block_2(h, temb)
-
-        # end
-        h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h)
-        return h
-
-
-class Decoder(nn.Module):
-    def __init__(
-        self,
-        *,
-        ch,
-        out_ch,
-        ch_mult=(1, 2, 4, 8),
-        num_res_blocks,
-        attn_resolutions,
-        dropout=0.0,
-        resamp_with_conv=True,
-        in_channels,
-        resolution,
-        z_channels,
-        give_pre_end=False,
-        tanh_out=False,
-        use_linear_attn=False,
-        attn_type="vanilla",
-        **ignorekwargs,
-    ):
-        super().__init__()
-        if use_linear_attn:
-            attn_type = "linear"
-        self.ch = ch
-        self.temb_ch = 0
-        self.num_resolutions = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        self.resolution = resolution
-        self.in_channels = in_channels
-        self.give_pre_end = give_pre_end
-        self.tanh_out = tanh_out
-
-        # compute in_ch_mult, block_in and curr_res at lowest res
-        in_ch_mult = (1,) + tuple(ch_mult)
-        block_in = ch * ch_mult[self.num_resolutions - 1]
-        curr_res = resolution // 2 ** (self.num_resolutions - 1)
-        self.z_shape = (1, z_channels, curr_res, curr_res)
-        print(
-            "Working with z of shape {} = {} dimensions.".format(
-                self.z_shape, np.prod(self.z_shape)
-            )
-        )
-
-        # z to block_in
-        self.conv_in = torch.nn.Conv2d(
-            z_channels, block_in, kernel_size=3, stride=1, padding=1
-        )
-
-        # middle
-        self.mid = nn.Module()
-        self.mid.block_1 = ResnetBlock(
-            in_channels=block_in,
-            out_channels=block_in,
-            temb_channels=self.temb_ch,
-            dropout=dropout,
-        )
-        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
-        self.mid.block_2 = ResnetBlock(
-            in_channels=block_in,
-            out_channels=block_in,
-            temb_channels=self.temb_ch,
-            dropout=dropout,
-        )
-
-        # upsampling
-        self.up = nn.ModuleList()
-        for i_level in reversed(range(self.num_resolutions)):
-            block = nn.ModuleList()
-            attn = nn.ModuleList()
-            block_out = ch * ch_mult[i_level]
-            for i_block in range(self.num_res_blocks + 1):
-                block.append(
-                    ResnetBlock(
-                        in_channels=block_in,
-                        out_channels=block_out,
-                        temb_channels=self.temb_ch,
-                        dropout=dropout,
-                    )
-                )
-                block_in = block_out
-                if curr_res in attn_resolutions:
-                    attn.append(make_attn(block_in, attn_type=attn_type))
-            up = nn.Module()
-            up.block = block
-            up.attn = attn
-            if i_level != 0:
-                up.upsample = Upsample(block_in, resamp_with_conv)
-                curr_res = curr_res * 2
-            self.up.insert(0, up)  # prepend to get consistent order
-
-        # end
-        self.norm_out = Normalize(block_in)
-        self.conv_out = torch.nn.Conv2d(
-            block_in, out_ch, kernel_size=3, stride=1, padding=1
-        )
-
-    def forward(self, z):
-        # assert z.shape[1:] == self.z_shape[1:]
-        self.last_z_shape = z.shape
-
-        # timestep embedding
-        temb = None
-
-        # z to block_in
-        h = self.conv_in(z)
-
-        # middle
-        h = self.mid.block_1(h, temb)
-        h = self.mid.attn_1(h)
-        h = self.mid.block_2(h, temb)
-
-        # upsampling
-        for i_level in reversed(range(self.num_resolutions)):
-            for i_block in range(self.num_res_blocks + 1):
-                h = self.up[i_level].block[i_block](h, temb)
-                if len(self.up[i_level].attn) > 0:
-                    h = self.up[i_level].attn[i_block](h)
-            if i_level != 0:
-                h = self.up[i_level].upsample(h)
-
-        # end
-        if self.give_pre_end:
-            return h
-
-        h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h)
-        if self.tanh_out:
-            h = torch.tanh(h)
-        return h
-
-
-class SimpleDecoder(nn.Module):
-    def __init__(self, in_channels, out_channels, *args, **kwargs):
-        super().__init__()
-        self.model = nn.ModuleList(
-            [
-                nn.Conv2d(in_channels, in_channels, 1),
-                ResnetBlock(
-                    in_channels=in_channels,
-                    out_channels=2 * in_channels,
-                    temb_channels=0,
-                    dropout=0.0,
-                ),
-                ResnetBlock(
-                    in_channels=2 * in_channels,
-                    out_channels=4 * in_channels,
-                    temb_channels=0,
-                    dropout=0.0,
-                ),
-                ResnetBlock(
-                    in_channels=4 * in_channels,
-                    out_channels=2 * in_channels,
-                    temb_channels=0,
-                    dropout=0.0,
-                ),
-                nn.Conv2d(2 * in_channels, in_channels, 1),
-                Upsample(in_channels, with_conv=True),
-            ]
-        )
-        # end
-        self.norm_out = Normalize(in_channels)
-        self.conv_out = torch.nn.Conv2d(
-            in_channels, out_channels, kernel_size=3, stride=1, padding=1
-        )
-
-    def forward(self, x):
-        for i, layer in enumerate(self.model):
-            if i in [1, 2, 3]:
-                x = layer(x, None)
-            else:
-                x = layer(x)
-
-        h = self.norm_out(x)
-        h = nonlinearity(h)
-        x = self.conv_out(h)
-        return x
-
-
-class UpsampleDecoder(nn.Module):
-    def __init__(
-        self,
-        in_channels,
-        out_channels,
-        ch,
-        num_res_blocks,
-        resolution,
-        ch_mult=(2, 2),
-        dropout=0.0,
-    ):
-        super().__init__()
-        # upsampling
-        self.temb_ch = 0
-        self.num_resolutions = len(ch_mult)
-        self.num_res_blocks = num_res_blocks
-        block_in = in_channels
-        curr_res = resolution // 2 ** (self.num_resolutions - 1)
-        self.res_blocks = nn.ModuleList()
-        self.upsample_blocks = nn.ModuleList()
-        for i_level in range(self.num_resolutions):
-            res_block = []
-            block_out = ch * ch_mult[i_level]
-            for i_block in range(self.num_res_blocks + 1):
-                res_block.append(
-                    ResnetBlock(
-                        in_channels=block_in,
-                        out_channels=block_out,
-                        temb_channels=self.temb_ch,
-                        dropout=dropout,
-                    )
-                )
-                block_in = block_out
-            self.res_blocks.append(nn.ModuleList(res_block))
-            if i_level != self.num_resolutions - 1:
-                self.upsample_blocks.append(Upsample(block_in, True))
-                curr_res = curr_res * 2
-
-        # end
-        self.norm_out = Normalize(block_in)
-        self.conv_out = torch.nn.Conv2d(
-            block_in, out_channels, kernel_size=3, stride=1, padding=1
-        )
-
-    def forward(self, x):
-        # upsampling
-        h = x
-        for k, i_level in enumerate(range(self.num_resolutions)):
-            for i_block in range(self.num_res_blocks + 1):
-                h = self.res_blocks[i_level][i_block](h, None)
-            if i_level != self.num_resolutions - 1:
-                h = self.upsample_blocks[k](h)
-        h = self.norm_out(h)
-        h = nonlinearity(h)
-        h = self.conv_out(h)
-        return h
-
-
-class LatentRescaler(nn.Module):
-    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
-        super().__init__()
-        # residual block, interpolate, residual block
-        self.factor = factor
-        self.conv_in = nn.Conv2d(
-            in_channels, mid_channels, kernel_size=3, stride=1, padding=1
-        )
-        self.res_block1 = nn.ModuleList(
-            [
-                ResnetBlock(
-                    in_channels=mid_channels,
-                    out_channels=mid_channels,
-                    temb_channels=0,
-                    dropout=0.0,
-                )
-                for _ in range(depth)
-            ]
-        )
-        self.attn = AttnBlock(mid_channels)
-        self.res_block2 = nn.ModuleList(
-            [
-                ResnetBlock(
-                    in_channels=mid_channels,
-                    out_channels=mid_channels,
-                    temb_channels=0,
-                    dropout=0.0,
-                )
-                for _ in range(depth)
-            ]
-        )
-
-        self.conv_out = nn.Conv2d(
-            mid_channels,
-            out_channels,
-            kernel_size=1,
-        )
-
-    def forward(self, x):
-        x = self.conv_in(x)
-        for block in self.res_block1:
-            x = block(x, None)
-        x = torch.nn.functional.interpolate(
-            x,
-            size=(
-                int(round(x.shape[2] * self.factor)),
-                int(round(x.shape[3] * self.factor)),
-            ),
-        )
-        x = self.attn(x)
-        for block in self.res_block2:
-            x = block(x, None)
-        x = self.conv_out(x)
-        return x
-
-
-class MergedRescaleEncoder(nn.Module):
-    def __init__(
-        self,
-        in_channels,
-        ch,
-        resolution,
-        out_ch,
-        num_res_blocks,
-        attn_resolutions,
-        dropout=0.0,
-        resamp_with_conv=True,
-        ch_mult=(1, 2, 4, 8),
-        rescale_factor=1.0,
-        rescale_module_depth=1,
-    ):
-        super().__init__()
-        intermediate_chn = ch * ch_mult[-1]
-        self.encoder = Encoder(
-            in_channels=in_channels,
-            num_res_blocks=num_res_blocks,
-            ch=ch,
-            ch_mult=ch_mult,
-            z_channels=intermediate_chn,
-            double_z=False,
-            resolution=resolution,
-            attn_resolutions=attn_resolutions,
-            dropout=dropout,
-            resamp_with_conv=resamp_with_conv,
-            out_ch=None,
-        )
-        self.rescaler = LatentRescaler(
-            factor=rescale_factor,
-            in_channels=intermediate_chn,
-            mid_channels=intermediate_chn,
-            out_channels=out_ch,
-            depth=rescale_module_depth,
-        )
-
-    def forward(self, x):
-        x = self.encoder(x)
-        x = self.rescaler(x)
-        return x
-
-
-class MergedRescaleDecoder(nn.Module):
-    def __init__(
-        self,
-        z_channels,
-        out_ch,
-        resolution,
-        num_res_blocks,
-        attn_resolutions,
-        ch,
-        ch_mult=(1, 2, 4, 8),
-        dropout=0.0,
-        resamp_with_conv=True,
-        rescale_factor=1.0,
-        rescale_module_depth=1,
-    ):
-        super().__init__()
-        tmp_chn = z_channels * ch_mult[-1]
-        self.decoder = Decoder(
-            out_ch=out_ch,
-            z_channels=tmp_chn,
-            attn_resolutions=attn_resolutions,
-            dropout=dropout,
-            resamp_with_conv=resamp_with_conv,
-            in_channels=None,
-            num_res_blocks=num_res_blocks,
-            ch_mult=ch_mult,
-            resolution=resolution,
-            ch=ch,
-        )
-        self.rescaler = LatentRescaler(
-            factor=rescale_factor,
-            in_channels=z_channels,
-            mid_channels=tmp_chn,
-            out_channels=tmp_chn,
-            depth=rescale_module_depth,
-        )
-
-    def forward(self, x):
-        x = self.rescaler(x)
-        x = self.decoder(x)
-        return x
-
-
-class Upsampler(nn.Module):
-    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
-        super().__init__()
-        assert out_size >= in_size
-        num_blocks = int(np.log2(out_size // in_size)) + 1
-        factor_up = 1.0 + (out_size % in_size)
-        print(
-            f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}"
-        )
-        self.rescaler = LatentRescaler(
-            factor=factor_up,
-            in_channels=in_channels,
-            mid_channels=2 * in_channels,
-            out_channels=in_channels,
-        )
-        self.decoder = Decoder(
-            out_ch=out_channels,
-            resolution=out_size,
-            z_channels=in_channels,
-            num_res_blocks=2,
-            attn_resolutions=[],
-            in_channels=None,
-            ch=in_channels,
-            ch_mult=[ch_mult for _ in range(num_blocks)],
-        )
-
-    def forward(self, x):
-        x = self.rescaler(x)
-        x = self.decoder(x)
-        return x
-
-
-class Resize(nn.Module):
-    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
-        super().__init__()
-        self.with_conv = learned
-        self.mode = mode
-        if self.with_conv:
-            print(
-                f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode"
-            )
-            raise NotImplementedError()
-            assert in_channels is not None
-            # no asymmetric padding in torch conv, must do it ourselves
-            self.conv = torch.nn.Conv2d(
-                in_channels, in_channels, kernel_size=4, stride=2, padding=1
-            )
-
-    def forward(self, x, scale_factor=1.0):
-        if scale_factor == 1.0:
-            return x
-        else:
-            x = torch.nn.functional.interpolate(
-                x, mode=self.mode, align_corners=False, scale_factor=scale_factor
-            )
-        return x
diff --git a/bin/validate_ci.py b/bin/validate_ci.py
deleted file mode 100644
index d62361221..000000000
--- a/bin/validate_ci.py
+++ /dev/null
@@ -1,8 +0,0 @@
-import truss
-import yaml
-
-with open("ci.yaml", "r") as file:
-    paths = yaml.safe_load(file)
-
-for path in paths["tests"]:
-    _ = truss.load(path)
diff --git a/binocular/config.yaml b/binocular/config.yaml
deleted file mode 100644
index 19d2b43df..000000000
--- a/binocular/config.yaml
+++ /dev/null
@@ -1,22 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_cache:
-- allow_patterns:
-  - '*.bin'
-  ignore_patterns:
-  - coreml/*
-  repo_id: tiiuae/falcon-7b
-  use_volume: false
-- allow_patterns:
-  - '*.bin'
-  ignore_patterns:
-  - coreml/*
-  repo_id: tiiuae/falcon-7b-instruct
-model_name: Binoculars
-python_version: py311
-requirements:
-- git+https://github.com/ahans30/Binoculars.git
-resources:
-  accelerator: A10G:2
-secrets: {}
-system_packages: []
diff --git a/chatterbox-tts/README.md b/chatterbox-tts/README.md
deleted file mode 100644
index f35fb0cdc..000000000
--- a/chatterbox-tts/README.md
+++ /dev/null
@@ -1,13 +0,0 @@
-# Chatterbox TTS Example
-## Overview
-This is a basic example showing how to deploy [Chatterbox TTS](https://github.com/resemble-ai/chatterbox) on truss.
-
-## Running the Example
-The truss endpoint is set up to accept text and an optional base64 wav string, and it returns that base64 string audio.
-
-For a more detailed example on how to run it, take a look at [`run_tts.py`](https://github.com/basetenlabs/truss-examples/blob/main/chatterbox-tts/run_tts.py). This script will take the `input_text`, apply the voice clone file `input/obama_8s.wav`, and output the audio file to `output/output_obama8s.wav`.
-
-## Custom Docker Image
-Currently, this example is built from an slightly modified baseten docker image. That image installs `numpy==1.26.0` on a truss base image in order to work with `chatterbox-tts==0.1.1`.
-
-For more details, see [Creating a custom base image](https://docs.baseten.co/development/model/base-images#creating-a-custom-base-image).
diff --git a/chatterbox-tts/config.yaml b/chatterbox-tts/config.yaml
deleted file mode 100644
index ea5f031b9..000000000
--- a/chatterbox-tts/config.yaml
+++ /dev/null
@@ -1,14 +0,0 @@
-model_name: Chatterbox TTS
-base_image:
-  image: jojobaseten/truss-numpy-1.26.0-gpu:0.4
-  python_executable_path: /usr/bin/python3
-python_version: py312
-requirements:
-  - chatterbox-tts
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 40Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
diff --git a/ci.yaml b/ci.yaml
deleted file mode 100644
index fd1f7f45c..000000000
--- a/ci.yaml
+++ /dev/null
@@ -1,15 +0,0 @@
-tests:
-  - 01-getting-started-bert
-  - 02-llm
-  - 03-llm-with-streaming
-  - 04-image-generation
-  - 05-speech-to-text
-  - 06-high-performance-cached-weights
-  - 10-using-system-packages
-  - whisper/whisper-v3-truss
-  - gfp-gan
-  - stable-diffusion/stable-diffusion-xl-1.0
-  - whisper/faster-whisper-v3
-  - playground-v2-aesthetic
-  - llama/tinyllama-1.1B-chat-v1.0
-  #- qwen/BEI-qwen-qwen3-embedding-0.6b-A10G # add BEI model
diff --git a/ci_excludes.yaml b/ci_excludes.yaml
new file mode 100644
index 000000000..3393c8d6e
--- /dev/null
+++ b/ci_excludes.yaml
@@ -0,0 +1,4 @@
+# Paths to exclude from CI auto-discovery.
+# Examples are discovered automatically by finding directories with config.yaml.
+# Add paths here to opt specific examples out of CI.
+exclude: []
diff --git a/clip/config.yaml b/clip/config.yaml
deleted file mode 100644
index 873618ceb..000000000
--- a/clip/config.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    url: https://images.pexels.com/photos/1170986/pexels-photo-1170986.jpeg?auto=compress&cs=tinysrgb&w=1600
-model_name: clip-example
-python_version: py311
-requirements:
-- transformers==4.47.1
-- pillow
-- torch
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/cogito/README.md b/cogito/README.md
deleted file mode 100644
index f3336b2bb..000000000
--- a/cogito/README.md
+++ /dev/null
@@ -1,117 +0,0 @@
-# Cogito v2 Preview Models
-
-This directory contains [Truss](https://truss.baseten.co/) configurations for deploying Cogito v2 preview models using vLLM's OpenAI-compatible server. These models feature powerful tool calling and reasoning capabilities, as detailed in the [Cogito v2 research blog post](https://www.deepcogito.com/research/cogito-v2-preview).
-
-## Available Models
-
-| Model | Size | Architecture | GPU Requirements |
-|-------|------|--------------|------------------|
-| [Cogito v2 Llama 70B](./cogito-v2-preview-llama-70B-vllm/) | 70B | Dense | 2x H100 |
-| [Cogito v2 Llama 109B MoE](./cogito-v2-preview-llama-109B-MoE-vllm/) | 109B | MoE (Mixture of Experts) | 4x H100 |
-| [Cogito v2 Llama 405B](./cogito-v2-preview-llama-405B-vllm/) | 405B | Dense | 8x B200 |
-| [Cogito v2 DeepSeek 671B MoE](./cogito-v2-preview-deepseek-671B-MoE-vllm/) | 671B | MoE (Mixture of Experts) | 8x B200 |
-
-
-
-
-## Prerequisites
-
-Before deploying any of these models, ensure you have:
-
-1. **Baseten Account**: Sign up at [app.baseten.co/signup](https://app.baseten.co/signup)
-2. **API Key**: Get your API key from [app.baseten.co/settings/api_keys](https://app.baseten.co/settings/api_keys)
-3. **Truss Installation**: Install the latest version: `pip install --upgrade truss`
-4. **Hugging Face Token**: Retrieve from [huggingface.co/settings/tokens](https://huggingface.co/settings/tokens)
-5. **Baseten Secret**: Set your HF token as a Baseten secret with key `hf_access_token` at [app.baseten.co/settings/secrets](https://app.baseten.co/settings/secrets)
-
-## Deployment
-
-### Quick Start
-
-1. Clone this repository:
-```bash
-git clone https://github.com/basetenlabs/truss-examples.git
-cd cogito
-```
-
-2. Navigate to your desired model directory:
-```bash
-cd cogito-v2-preview-llama-70B-vllm  # or any other model
-```
-
-3. Deploy the model:
-```bash
-truss push --publish
-```
-
-### GPU Requirements
-
-- **B200 GPUs**: Required for 405B and 671B models (contact [support@baseten.co](mailto:support@baseten.co) before deploying)
-- **H100 GPUs**: Required for 70B and 109B MoE models
-
-## API Usage
-
-All models follow the OpenAI ChatCompletion format. Here's an example using the Python client for tool calling:
-
-```python
-from openai import OpenAI
-
-# Replace with your model ID after deployment
-model_id = "your-model-id"  # e.g. "yqvy46gq"
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-# Example tool calling
-def get_temperature_in_celsius(location=None):
-    return 22
-
-tools = [
-    {
-        "type": "function",
-        "function": {
-            "name": "get_temperature_in_celsius",
-            "description": "Get the current temperature in celsius.",
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "location": {
-                        "type": "string",
-                        "description": "The location to get the temperature for."
-                    }
-                },
-                "required": ["location"]
-            }
-        }
-    }
-]
-
-# Chat completion with tool calling
-response = client.chat.completions.create(
-    model="llama",  # or "deepseek" for DeepSeek model
-    messages=[
-        {
-            "role": "user",
-            "content": "What is today's temperature in celsius? I'm in Paris."
-        }
-    ],
-    tools=tools,
-    max_tokens=1000,
-    temperature=0.6
-)
-
-print(response.json())
-```
-
-
-## Support
-
-- **Documentation**: [Truss documentation](https://truss.baseten.co)
-- **Issues**: Open an issue in this repository
-- **Support**: Contact [support@baseten.co](mailto:support@baseten.co)
-- **Research**: [Cogito v2 research blog](https://www.deepcogito.com/research/cogito-v2-preview)
-
-## License
-
-These models are subject to the respective licenses of the underlying model weights. Please refer to the Hugging Face model pages for specific licensing information.
diff --git a/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/README.md b/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/README.md
deleted file mode 100644
index 66259a566..000000000
--- a/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/README.md
+++ /dev/null
@@ -1,104 +0,0 @@
-# Cogito v2 DeepSeek 671B MoE Truss (vLLM)
-
-Cogito's DeepSeek-based 671B MoE model has powerful tool calling and reasoning capabilities. See this [blog post](https://www.deepcogito.com/research/cogito-v2-preview).
-
-This is a [Truss](https://truss.baseten.co/) to deploy the model using the vLLM OpenAI Compatible server. This model requires 8x B200 GPUs to deploy. Users should contact [support@baseten.co](mailto:support@baseten.co) before deploying.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd cogito/cogito-v2-preview-deepseek-671B-MoE-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Note that you will *not* be able to successfully deploy the model without doing this.
-
-With `cogito-v2-preview-deepseek-671B-MoE-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-model_id = "your-model-id" # Replace with your model ID
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-def get_temperature_in_celsius(location=None):
-    return 22
-
-tools = [
-    {
-        "type": "function",
-        "function": {
-            "name": "get_temperature_in_celsius",
-            "description": "Get the current temperature in celsius.",
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "location": {
-                        "type": "string",
-                        "description": "The location to get the temperature for."
-                    }
-                },
-                "required": [
-                    "location"
-                ]
-            }
-        }
-    }
-]
-
-# Example usage of the OpenAI client to use a tool call
-response = client.chat.completions.create(
-    model="deepseek",
-    messages=[
-        {
-            "role": "user",
-            "content": "What is today's temperature in celsius? I'm in Paris."
-        }
-    ],
-    tools=tools,
-)
-
-print(response.json())
-```
-
-## Model Details
-
-- **Model**: Cogito v2 Preview DeepSeek 671B MoE FP8
-- **Architecture**: Mixture of Experts (MoE)
-- **GPU Requirements**: 8x B200
-- **Tool Call Parser**: deepseek_v3
-- **Tensor Parallel Size**: 8
-- **GPU Memory Utilization**: 90%
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our [support team](mailto:support@baseten.co).
diff --git a/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/config.yaml b/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/config.yaml
deleted file mode 100644
index 03e19c4d2..000000000
--- a/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/config.yaml
+++ /dev/null
@@ -1,36 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.9.2
-model_metadata:
-  example_model_input: {
-    model: "deepseek",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    stream: true,
-    max_tokens: 10000,
-    temperature: 0.6
-  }
-  repo_id: deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8 --served-model-name deepseek --max-model-len 131072 --port 8000 --gpu-memory-utilization 0.90 --disable-custom-all-reduce --trust-remote-code --tensor-parallel-size 8 --distributed-executor-backend mp --enable-auto-tool-choice --tool-call-parser deepseek_v3"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-cache_internal:
-  - repo_id: deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8
-resources:
-  accelerator: B200:8
-  cpu: '1'
-  memory: 24Gi
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: Cogito V2 Preview DeepSeek 671B MoE FP8 vLLM
-environment_variables:
-  hf_access_token: null
diff --git a/cogito/cogito-v2-preview-llama-109B-MoE-vllm/README.md b/cogito/cogito-v2-preview-llama-109B-MoE-vllm/README.md
deleted file mode 100644
index 073c5b613..000000000
--- a/cogito/cogito-v2-preview-llama-109B-MoE-vllm/README.md
+++ /dev/null
@@ -1,103 +0,0 @@
-# Cogito v2 Llama 109B MoE Truss (vLLM)
-
-Cogito's Llama-based 109B MoE model has powerful tool calling and reasoning capabilities. See this [blog post](https://www.deepcogito.com/research/cogito-v2-preview).
-
-This is a [Truss](https://truss.baseten.co/) to deploy the model using the vLLM OpenAI Compatible server. This model requires 4x H100 GPUs to deploy.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd cogito/cogito-v2-preview-llama-109B-MoE-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Note that you will *not* be able to successfully deploy the model without doing this.
-
-With `cogito-v2-preview-llama-109B-MoE-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-model_id = "your-model-id" # Replace with your model ID
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-def get_temperature_in_celsius(location=None):
-    return 22
-
-tools = [
-    {
-        "type": "function",
-        "function": {
-            "name": "get_temperature_in_celsius",
-            "description": "Get the current temperature in celsius.",
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "location": {
-                        "type": "string",
-                        "description": "The location to get the temperature for."
-                    }
-                },
-                "required": [
-                    "location"
-                ]
-            }
-        }
-    }
-]
-
-# Example usage of the OpenAI client to use a tool call
-response = client.chat.completions.create(
-    model="llama",
-    messages=[
-        {
-            "role": "user",
-            "content": "What is today's temperature in celsius? I'm in Paris."
-        }
-    ],
-    tools=tools,
-)
-
-print(response.json())
-```
-
-## Model Details
-
-- **Model**: Cogito v2 Preview Llama 109B MoE
-- **Architecture**: Mixture of Experts (MoE)
-- **GPU Requirements**: 4x H100
-- **Tool Call Parser**: llama3_json
-- **Tensor Parallel Size**: 4
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our [support team](mailto:support@baseten.co).
diff --git a/cogito/cogito-v2-preview-llama-109B-MoE-vllm/config.yaml b/cogito/cogito-v2-preview-llama-109B-MoE-vllm/config.yaml
deleted file mode 100644
index 83f629b28..000000000
--- a/cogito/cogito-v2-preview-llama-109B-MoE-vllm/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.10.0
-model_metadata:
-  repo_id: deepcogito/cogito-v2-preview-llama-109B-MoE
-  example_model_input: {
-    "model": "llama",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    "stream": true,
-    "max_tokens": 10000,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-llama-109B-MoE --served-model-name llama --max-model-len 32000 --tensor-parallel-size 4 --distributed-executor-backend mp --enable-auto-tool-choice --tool-call-parser llama3_json"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
-  hf_access_token: null
-resources:
-  accelerator: H100:4
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  predict_concurrency : 32
-model_name: Cogito V2 Preview Llama 109B MoE vLLM
diff --git a/cogito/cogito-v2-preview-llama-405B-vllm/README.md b/cogito/cogito-v2-preview-llama-405B-vllm/README.md
deleted file mode 100644
index 2c367025a..000000000
--- a/cogito/cogito-v2-preview-llama-405B-vllm/README.md
+++ /dev/null
@@ -1,104 +0,0 @@
-# Cogito v2 Llama 405B Truss (vLLM)
-
-Cogito's Llama-based 405B model has powerful tool calling and reasoning capabilities. See this [blog post](https://www.deepcogito.com/research/cogito-v2-preview).
-
-This is a [Truss](https://truss.baseten.co/) to deploy the model using the vLLM OpenAI Compatible server. This model requires 8x B200 GPUs to deploy. Users should contact [support@baseten.co](mailto:support@baseten.co) before deploying.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd cogito/cogito-v2-preview-llama-405B-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Note that you will *not* be able to successfully deploy the model without doing this.
-
-With `cogito-v2-preview-llama-405B-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-model_id = "your-model-id" # Replace with your model ID
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-def get_temperature_in_celsius(location=None):
-    return 22
-
-tools = [
-    {
-        "type": "function",
-        "function": {
-            "name": "get_temperature_in_celsius",
-            "description": "Get the current temperature in celsius.",
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "location": {
-                        "type": "string",
-                        "description": "The location to get the temperature for."
-                    }
-                },
-                "required": [
-                    "location"
-                ]
-            }
-        }
-    }
-]
-
-# Example usage of the OpenAI client to use a tool call
-response = client.chat.completions.create(
-    model="llama",
-    messages=[
-        {
-            "role": "user",
-            "content": "What is today's temperature in celsius? I'm in Paris."
-        }
-    ],
-    tools=tools,
-)
-
-print(response.json())
-```
-
-## Model Details
-
-- **Model**: Cogito v2 Preview Llama 405B
-- **Architecture**: Dense transformer
-- **GPU Requirements**: 8x B200
-- **Tool Call Parser**: llama3_json
-- **Features**: Prefix caching, chunked prefill
-- **Tensor Parallel Size**: 8
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our [support team](mailto:support@baseten.co).
diff --git a/cogito/cogito-v2-preview-llama-405B-vllm/config.yaml b/cogito/cogito-v2-preview-llama-405B-vllm/config.yaml
deleted file mode 100644
index 93ea22386..000000000
--- a/cogito/cogito-v2-preview-llama-405B-vllm/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.10.0
-model_metadata:
-  repo_id: deepcogito/cogito-v2-preview-llama-405B
-  example_model_input: {
-    "model": "llama",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    "stream": true,
-    "max_tokens": 10000,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-llama-405B --served-model-name llama --max-model-len 32000 --tensor-parallel-size 8 --enable-chunked-prefill --enable-prefix-caching --max-num-seqs 8 --distributed-executor-backend mp --enable-auto-tool-choice --tool-call-parser llama3_json "
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
-  hf_access_token: null
-cache_internal:
-  - repo_id: deepcogito/cogito-v2-preview-llama-405B
-resources:
-  accelerator: B200:8
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  predict_concurrency : 32
-model_name: Cogito V2 Preview Llama 405B vLLM
diff --git a/cogito/cogito-v2-preview-llama-70B-vllm/README.md b/cogito/cogito-v2-preview-llama-70B-vllm/README.md
deleted file mode 100644
index 6e8771728..000000000
--- a/cogito/cogito-v2-preview-llama-70B-vllm/README.md
+++ /dev/null
@@ -1,103 +0,0 @@
-# Cogito v2 Llama 70B Truss (vLLM)
-
-Cogito's Llama-based 70B model has powerful tool calling and reasoning capabilities. See this [blog post](https://www.deepcogito.com/research/cogito-v2-preview).
-
-This is a [Truss](https://truss.baseten.co/) to deploy the model using the vLLM OpenAI Compatible server. This model requires 2x H100 GPUs to deploy.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd cogito/cogito-v2-preview-llama-70B-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Note that you will *not* be able to successfully deploy the model without doing this.
-
-With `cogito-v2-preview-llama-70B-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-model_id = "your-model-id" # Replace with your model ID
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-def get_temperature_in_celsius(location=None):
-    return 22
-
-tools = [
-    {
-        "type": "function",
-        "function": {
-            "name": "get_temperature_in_celsius",
-            "description": "Get the current temperature in celsius.",
-            "parameters": {
-                "type": "object",
-                "properties": {
-                    "location": {
-                        "type": "string",
-                        "description": "The location to get the temperature for."
-                    }
-                },
-                "required": [
-                    "location"
-                ]
-            }
-        }
-    }
-]
-
-# Example usage of the OpenAI client to use a tool call
-response = client.chat.completions.create(
-    model="llama",
-    messages=[
-        {
-            "role": "user",
-            "content": "What is today's temperature in celsius? I'm in Paris."
-        }
-    ],
-    tools=tools,
-)
-
-print(response.json())
-```
-
-## Model Details
-
-- **Model**: Cogito v2 Preview Llama 70B
-- **Architecture**: Dense transformer
-- **GPU Requirements**: 2x H100
-- **Tool Call Parser**: llama3_json
-- **GPU Memory Utilization**: 95%
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our [support team](mailto:support@baseten.co).
diff --git a/cogito/cogito-v2-preview-llama-70B-vllm/config.yaml b/cogito/cogito-v2-preview-llama-70B-vllm/config.yaml
deleted file mode 100755
index 9ab76b510..000000000
--- a/cogito/cogito-v2-preview-llama-70B-vllm/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.10.0
-model_metadata:
-  repo_id: deepcogito/cogito-v2-preview-llama-70B
-  example_model_input: {
-    "model": "llama",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    "stream": true,
-    "max_tokens": 10000,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-llama-70B --served-model-name llama --max-model-len 32000 --tensor-parallel-size 2 --distributed-executor-backend mp --gpu-memory-utilization 0.95 --enable-auto-tool-choice --tool-call-parser llama3_json"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
-  hf_access_token: null
-resources:
-  accelerator: H100:2
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  predict_concurrency : 32
-model_name: Cogito V2 Preview Llama 70B vLLM
diff --git a/cogvlm/README.md b/cogvlm/README.md
deleted file mode 100644
index ff38efe7e..000000000
--- a/cogvlm/README.md
+++ /dev/null
@@ -1,48 +0,0 @@
-# CogVLM Truss
-
-This repository packages [CogVLM](https://github.com/THUDM/CogVLM) as a [Truss](https://truss.baseten.co/).
-
-CogVLM is a highly performant open-source vision language model with capabilities similar to GPT-4V.
-
-## Deploying CogVLM
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd cogvlm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `cogvlm` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking CogVLM
-
-CogVLM takes in two inputs, a `query` and a base64 encoded `image`. CogVLM will respond to the `query` conditioned on the `image`. The output is a JSON blob with a single key, `result`, that answers the `query`.
-
-
-```sh
-truss predict -d '{"query": "Describe this picture in detail.", "image": "data:image/png;base64,iVBORw0KGgoA..."}'
-```
-
-You can also invoke your model via a REST API
-```
-curl -X POST https://app.baseten.co/model_versions/<YOUR_MODEL_VERSION_ID>/predict \
-     -H "Content-Type: application/json" \
-     -d '{
-           "query": "Describe this picture in detail.",
-           "image": "data:image/png;base64,iVBORw0KGgoA..."
-         }'
-```
diff --git a/cogvlm/config.yaml b/cogvlm/config.yaml
deleted file mode 100644
index 83591b309..000000000
--- a/cogvlm/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: CogVLM
-python_version: py311
-requirements:
-- torch==2.0.1
-- sentencepiece==0.1.99
-- protobuf==4.25.1
-- transformers==4.35.2
-- einops==0.7.0
-- torchvision==0.15.2
-- Pillow==10.1.0
-- xformers==0.0.22
-- accelerate==0.25.0
-resources:
-  accelerator: A100
-  cpu: '3'
-  memory: 15Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/comfyui-truss/README.md b/comfyui-truss/README.md
deleted file mode 100644
index 66e058b29..000000000
--- a/comfyui-truss/README.md
+++ /dev/null
@@ -1,398 +0,0 @@
-## ComfyUI Truss
-
-This truss is designed to allow ComfyUI users to easily convert their workflows into a production grade API service.
-
-## Exporting the ComfyUI workflow
-
-This Truss is designed to run a Comfy UI workflow that is in the form of a JSON file.
-
-Inside ComfyUI, you can save workflows as a JSON file. However, the regular JSON format that ComfyUI uses will not work. Instead, the workflow has to be saved in the API format. Here is how you can do that:
-
-First, go to ComfyUI and click on the gear icon for the project
-
-![gear_icon](../assets/comfyui-screenshot-1.png)
-
-Next, checkmark the box which says `Enable Dev Mode Options`
-
-![enable_dev_mode_options](../assets/comfyui-screenshot-2.png)
-
-Now, if you go back to the project you will see a new option called `Save (API Format)`. This is the one you want to use to save your workflow. Using this method you can save any ComfyUI workflow as a JSON file in the API format.
-
-![save_api_format](../assets/comfyui-screenshot-3.png)
-
-
-## Setting up the project
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd comfyui-truss
-```
-
-For your ComfyUI workflow, you probably used one or more models. Those models need to be defined inside truss. From the root of the truss project, open the file called `config.yaml`. In this file we will modify an element called `build_commands`. Build commands will allow you to run docker commands at build time. You can use this feature to download model weights and install custom nodes.
-
-```yaml
-build_commands:
-- git clone https://github.com/comfyanonymous/ComfyUI.git
-- cd ComfyUI && git checkout b1fd26fe9e55163f780bf9e5f56bf9bf5f035c93 && pip install -r requirements.txt
-- cd ComfyUI/custom_nodes && git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes --recursive && cd ComfyUI-Inference-Core-Nodes && pip install -e .[cuda12]
-- cd ComfyUI/custom_nodes && git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-Gemini --recursive && cd ComfyUI-Gemini && pip install -r requirements.txt
-- cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-Marigold --recursive && cd ComfyUI-Marigold && pip install -r requirements.txt
-- cd ComfyUI/custom_nodes && git clone https://github.com/omar92/ComfyUI-QualityOfLifeSuit_Omar92 --recursive
-- cd ComfyUI/custom_nodes && git clone https://github.com/Fannovel16/comfyui_controlnet_aux --recursive && cd comfyui_controlnet_aux && pip install -r requirements.txt
-- cd ComfyUI/models/controlnet && wget -O control-lora-canny-rank256.safetensors https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-canny-rank256.safetensors
-- cd ComfyUI/models/controlnet && wget -O control-lora-depth-rank256.safetensors https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-depth-rank256.safetensors
-- cd ComfyUI/models/checkpoints && wget -O dreamshaperXL_v21TurboDPMSDE.safetensors https://civitai.com/api/download/models/351306
-- cd ComfyUI/models/loras && wget -O StudioGhibli.Redmond-StdGBRRedmAF-StudioGhibli.safetensors https://huggingface.co/artificialguybr/StudioGhibli.Redmond-V2/resolve/main/StudioGhibli.Redmond-StdGBRRedmAF-StudioGhibli.safetensors
-environment_variables: {}
-external_package_dirs: []
-model_metadata: {}
-model_name: comfy build commands
-python_version: py310
-requirements:
-  - websocket-client==1.6.4
-  - accelerate==0.23.0
-  - opencv-python
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets: {}
-system_packages:
-  - wget
-  - ffmpeg
-  - libgl1-mesa-glx
-```
-
-Here is a breakdown of the actions happenning in `build_commands`. First, we are cloning the ComfyUI repository, checking out a specific commit, and installing the required python packages to run ComfyUI. Next, we use the `cd` commmand ensure all custom nodes are downloaded within the `custom_nodes` directory, and their respective requirements are installed. Similarly for the checkpoints, we use the `wget` utility to download the checkpoints and place them in the appropriate directories within ComfyUI.
-
-Every line under `build_commands` effectively does a `RUN` inside the dockerfile. The benefit of using this feature is that your model weights and custom nodes get cached during the docker build stage. So when your model deploys, it can access the cached weights directly which reduces the cold-start time.
-
-We also need to place the JSON workflow from step 1 inside the data directory. In the data directory create an open a file called `data/comfy_ui_workflow.json`. Copy and paste the entire JSON workflow that we saved in step 1 into this file.
-
-In the JSON workflow file, there might be some inputs such as the positive prompt or negative prompt that are hard coded. We want these inputs to be dynamically sent to the model, so we can use handlebars to templatize them. Here is an example of a JSON workflow with templatized inputs:
-
-```json
-{
-  "6": {
-    "inputs": {
-      "text": "{{positive_prompt}}",
-      "clip": [
-        "14",
-        1
-      ]
-    },
-    "class_type": "CLIPTextEncode"
-  },
-  "7": {
-    "inputs": {
-      "text": "{{negative_prompt}}",
-      "clip": [
-        "14",
-        1
-      ]
-    },
-    "class_type": "CLIPTextEncode"
-  },
-  "11": {
-    "inputs": {
-      "image": "{{controlnet_image}}",
-      "choose file to upload": "image"
-    },
-    "class_type": "LoadImage"
-  },
-  "12": {
-    "inputs": {
-      "control_net_name": "diffusers_xl_canny_full.safetensors"
-    },
-    "class_type": "ControlNetLoader"
-  },
-  "14": {
-    "inputs": {
-      "ckpt_name": "sd_xl_base_1.0.safetensors"
-    },
-    "class_type": "CheckpointLoaderSimple"
-  },
-  "15": {
-    "inputs": {
-      "images": [
-        "16",
-        0
-      ]
-    },
-    "class_type": "PreviewImage"
-  },
-  "18": {
-    "inputs": {
-      "images": [
-        "8",
-        0
-      ]
-    },
-    "class_type": "PreviewImage"
-  }
-}
-```
-
-This is not the entire JSON workflow file, but the nodes 6, 7, and 11 accept variable inputs. You can do this by using the handlebars format of `{{variable_name_here}}`.
-
-## Custom Nodes
-If your workflow uses custom nodes you add it to the `build_commands` in the `config.yaml` file. Let's take an example. Suppose you want to add the [UltimateSDUpscale](https://github.com/ssitu/ComfyUI_UltimateSDUpscale) custom node. Inside your `config.yaml` you can define it like so:
-
-```yaml
-build_commands:
-- cd ComfyUI/custom_nodes && git clone https://github.com/ssitu/ComfyUI_UltimateSDUpscale --recursive
-```
-
-Each command in `build_commands` is independent of the previous command. So you need to execute this part `cd ComfyUI/custom_nodes` each time you want to install a new custom node.
-
-Once you have both the `data/comfy_ui_workflow.json` and `config.yaml` set up correctly we can begin deployment.
-
-## Deployment
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `comfyui-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Model Inference
-
-When an inference request is sent to the Truss, the `comfy_ui_workflow.json` in the data directory is sent to ComfyUI. If you recall, there are some templatized variables inside that json file using the handlebars format of `{{variable_name_here}}`. During inference time, we can dynamically pass in those templatized variables to our Truss prediction request like so:
-
-```python
-values = {
-  "positive_prompt": "An igloo on a snowy day, 4k, hd",
-  "negative_prompt": "blurry, text, low quality",
-  "controlnet_image": "https://storage.googleapis.com/logos-bucket-01/baseten_logo.png"
-}
-```
-
-Just be sure that the variable names in the `comfy_ui_workflow.json` template match the names inside the values object.
-
-Here is a complete example of how you make a prediction request to your truss in python:
-
-This is the content of `data/comfy_ui_workflow.json`:
-```json
-sdxl_controlnet_workflow = {
-  "3": {
-    "inputs": {
-      "seed": 972197629127129,
-      "steps": 40,
-      "cfg": 7,
-      "sampler_name": "euler",
-      "scheduler": "normal",
-      "denoise": 1,
-      "model": [
-        "14",
-        0
-      ],
-      "positive": [
-        "10",
-        0
-      ],
-      "negative": [
-        "7",
-        0
-      ],
-      "latent_image": [
-        "5",
-        0
-      ]
-    },
-    "class_type": "KSampler"
-  },
-  "5": {
-    "inputs": {
-      "width": 1024,
-      "height": 1024,
-      "batch_size": 1
-    },
-    "class_type": "EmptyLatentImage"
-  },
-  "6": {
-    "inputs": {
-      "text": "{{positive_prompt}}",
-      "clip": [
-        "14",
-        1
-      ]
-    },
-    "class_type": "CLIPTextEncode"
-  },
-  "7": {
-    "inputs": {
-      "text": "{{negative_prompt}}",
-      "clip": [
-        "14",
-        1
-      ]
-    },
-    "class_type": "CLIPTextEncode"
-  },
-  "8": {
-    "inputs": {
-      "samples": [
-        "3",
-        0
-      ],
-      "vae": [
-        "14",
-        2
-      ]
-    },
-    "class_type": "VAEDecode"
-  },
-  "10": {
-    "inputs": {
-      "strength": 0.6,
-      "conditioning": [
-        "6",
-        0
-      ],
-      "control_net": [
-        "12",
-        0
-      ],
-      "image": [
-        "16",
-        0
-      ]
-    },
-    "class_type": "ControlNetApply"
-  },
-  "11": {
-    "inputs": {
-      "image": "{{controlnet_image}}",
-      "choose file to upload": "image"
-    },
-    "class_type": "LoadImage"
-  },
-  "12": {
-    "inputs": {
-      "control_net_name": "diffusers_xl_canny_full.safetensors"
-    },
-    "class_type": "ControlNetLoader"
-  },
-  "14": {
-    "inputs": {
-      "ckpt_name": "sd_xl_base_1.0.safetensors"
-    },
-    "class_type": "CheckpointLoaderSimple"
-  },
-  "15": {
-    "inputs": {
-      "images": [
-        "16",
-        0
-      ]
-    },
-    "class_type": "PreviewImage"
-  },
-  "16": {
-    "inputs": {
-      "low_threshold": 0.2,
-      "high_threshold": 0.6,
-      "image": [
-        "11",
-        0
-      ]
-    },
-    "class_type": "Canny"
-  },
-  "18": {
-    "inputs": {
-      "images": [
-        "8",
-        0
-      ]
-    },
-    "class_type": "PreviewImage"
-  }
-}
-```
-
-Here is the actual API request sent to Truss:
-```python
-import os
-import random
-import base64
-import requests
-
-# Set essential values
-model_id = ""
-baseten_api_key = ""
-# Set prompts and controlnet image
-values = {
-  "positive_prompt": "A top down view of a river through the woods",
-  "negative_prompt": "blurry, text, low quality",
-  "controlnet_image": "https://storage.googleapis.com/logos-bucket-01/baseten_logo.png",
-  "seed": random.randint(1, 1000000)
-}
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/development/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json={"workflow_values": values}
-)
-# Get output image
-res = res.json()
-preamble = "data:image/png;base64,"
-output = base64.b64decode(res["result"][1]["image"].replace(preamble, ""))
-# Save image to file
-img_file = open("comfyui.png", 'wb')
-img_file.write(output)
-img_file.close()
-os.system("open comfyui.png")
-```
-
-Here is the output of the request above:
-
-```json
-[
-    {
-        "node_id": "18",
-        "data": "base64-image-string",
-        "format": "png"
-    },
-    {
-        "node_id": "15",
-        "data": "base64-image-string",
-        "format": "png"
-    }
-]
-```
-
-The output of the model is a list of JSON objects containing the ID of the output node along with the generated image as a base64 string.
-
-You can also send input images as base64 strings. In the above example simply change the `values` python dictionary to look like this:
-
-```python
-from PIL import Image
-from io import BytesIO
-import base64
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-new_values = {
-  "positive_prompt": "A top down view of a river through the woods",
-  "negative_prompt": "blurry, text, low quality",
-  "controlnet_image": {"type": "image", "data": pil_to_b64(Image.open("my-image.jpeg"))},
-  "seed": random.randint(1, 1000000)
-}
-```
-
-When using base64 as input you need to specify the `type` so that the model can convert it to the correct data.
diff --git a/comfyui-truss/config.yaml b/comfyui-truss/config.yaml
deleted file mode 100644
index b058b6c92..000000000
--- a/comfyui-truss/config.yaml
+++ /dev/null
@@ -1,27 +0,0 @@
-base_image:
-  image: bolabaseten/comfyui-truss-base:6a7bc35
-  python_executable_path: /usr/bin/python3
-description: Deploy a ComfyUI workflow as a Truss
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    workflow_values:
-      controlnet_image: https://storage.googleapis.com/logos-bucket-01/baseten_logo.png
-      negative_prompt: blurry, text, low quality
-      positive_prompt: An igloo on a snowy day, 4k, hd
-model_name: ComfyUI Workflow
-python_version: py39
-requirements:
-- websocket-client==1.6.4
-- accelerate==0.23.0
-- opencv-python
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
-- libgl1-mesa-glx
diff --git a/comfyui-truss/examples/anime-style-transfer/config.yaml b/comfyui-truss/examples/anime-style-transfer/config.yaml
deleted file mode 100644
index 04442c3d3..000000000
--- a/comfyui-truss/examples/anime-style-transfer/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-build_commands:
-- git clone https://github.com/comfyanonymous/ComfyUI.git
-- cd ComfyUI && git checkout b1fd26fe9e55163f780bf9e5f56bf9bf5f035c93 && pip install -r requirements.txt
-- cd ComfyUI/custom_nodes && git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes --recursive && cd ComfyUI-Inference-Core-Nodes && pip install -e .[cuda12]
-- cd ComfyUI/custom_nodes && git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-Gemini --recursive && cd ComfyUI-Gemini && pip install -r requirements.txt
-- cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-Marigold --recursive && cd ComfyUI-Marigold && pip install -r requirements.txt
-- cd ComfyUI/custom_nodes && git clone https://github.com/omar92/ComfyUI-QualityOfLifeSuit_Omar92 --recursive
-- cd ComfyUI/custom_nodes && git clone https://github.com/Fannovel16/comfyui_controlnet_aux --recursive && cd comfyui_controlnet_aux && pip install -r requirements.txt
-- cd ComfyUI/models/controlnet && wget -O control-lora-canny-rank256.safetensors https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-canny-rank256.safetensors
-- cd ComfyUI/models/controlnet && wget -O control-lora-depth-rank256.safetensors https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-depth-rank256.safetensors
-- cd ComfyUI/models/checkpoints && wget -O dreamshaperXL_v21TurboDPMSDE.safetensors https://civitai.com/api/download/models/351306
-- cd ComfyUI/models/loras && wget -O StudioGhibli.Redmond-StdGBRRedmAF-StudioGhibli.safetensors https://huggingface.co/artificialguybr/StudioGhibli.Redmond-V2/resolve/main/StudioGhibli.Redmond-StdGBRRedmAF-StudioGhibli.safetensors
-environment_variables: {}
-external_package_dirs: []
-model_metadata: {}
-model_name: ComfyUI Anime Pet Style Transfer
-python_version: py310
-requirements:
-  - websocket-client==1.6.4
-  - accelerate==0.23.0
-  - opencv-python
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets: {}
-system_packages:
-  - wget
-  - ffmpeg
-  - libgl1-mesa-glx
diff --git a/control-net-qrcode/README.md b/control-net-qrcode/README.md
deleted file mode 100644
index f688f487f..000000000
--- a/control-net-qrcode/README.md
+++ /dev/null
@@ -1,95 +0,0 @@
-# Control Net QR Code Truss
-
-This truss allows you to generate QR codes using Stable Diffusion and Control Net. By typing in a prompt using the same guidelines as stable diffusion, a new image gets created that combines the image from the prompt with the image of a qr code. For this truss [Stable Diffusion 1.5](https://huggingface.co/runwayml/stable-diffusion-v1-5) is used along with [Control QR Code Monster](https://huggingface.co/monster-labs/control_v1p_sd15_qrcode_monster) as the control net.
-
-![controlnet_qr_code](controlnet_qr_code_results.gif)
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd control-net-qrcode
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `control-net-qrcode` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-This two models combined only take up about 5 GB of VRAM so a T4 is enough for this truss.
-
-### API route: `predict`
-
-The predict route is the primary method for generating images based on a given prompt. It takes several parameters:
-
-- __prompt__ (required): The input text required for image generation.
-- __qr_code_content__ (optional): The URL the QR code points to.
-- __mask__ (optional): If a URL is not passed into qr_code_content, a custom image can be used as a mask
-- __negative_prompt__ (optional, default=""): Use this to refine the image generation by discarding unwanted items.
-- __guidance_scale__ (optional, default=7.5): Used to control image generation.
-- __condition_scale__ (optional, default=1.2): The lower the condition_scale, the more creative the results. A higher condition_scale will result in less creative and more scannable QR Codes.
-- __sampler__(optional, default="Euler a"): This controls which sampler to use resulting in more image variations.
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "A cubism painting of the Garden of Eaden with animals walking around, Andreas Rocha, matte painting concept art, a detailed matte painting", "qr_code_content": "https://www.baseten.co"}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "A cubism painting of the Garden of Eaden with animals walking around, Andreas Rocha, matte painting concept art, a detailed matte painting", "qr_code_content": "https://www.baseten.co"
-         }'
-```
-
-If you want to use a mask instead of a QR code you can call the API like so:
-``` python
-import requests
-import base64
-from io import BytesIO
-from PIL import Image
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-input_image = Image.open("image/path/here")
-b64_img = pil_to_b64(input_image)
-headers = {"Authorization": f"Api-Key YOUR-API-KEY-HERE"}
-data = {
-  "prompt": "a cubism painting of the Garden of Eaden with animals walking around, Andreas Rocha, matte painting  concept art, a detailed matte painting",
-  "mask": b64_img
-}
-res = requests.post("https://app.baseten.co/model_versions/MODEL_VERSION/predict", headers=headers, json=data)
-pil_img = b64_to_pil(output.get("model_output").get("result"))
-pil_img.save("output.jpg")
-```
-
-Here is the output when using the `twitter_mask.jpeg` as the input image for the mask:
-
-![twitter_output](twitter_output.jpg)
diff --git a/control-net-qrcode/config.yaml b/control-net-qrcode/config.yaml
deleted file mode 100644
index 7744192c2..000000000
--- a/control-net-qrcode/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    prompt: A cubism painting of the Garden of Eaden with animals walking around,
-      Andreas Rocha, matte painting concept art, a detailed matte painting
-    qr_code_content: https://www.baseten.co
-model_name: control-net-qrcode
-python_version: py310
-requirements:
-- diffusers==0.21.1
-- torch==2.0.1
-- ftfy==6.1.1
-- scipy==1.9.3
-- transformers==4.25.1
-- accelerate==0.20.3
-- qrcode==7.4.2
-- xformers==0.0.21
-resources:
-  accelerator: T4
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/custom-engine-builder-control/README.md b/custom-engine-builder-control/README.md
deleted file mode 100644
index c57bca69c..000000000
--- a/custom-engine-builder-control/README.md
+++ /dev/null
@@ -1,135 +0,0 @@
-# TensorRT-LLM Briton with Qwen/Qwen3-8B-min-latency
-
-This is a Deployment for TensorRT-LLM Briton with Qwen/Qwen3-8B-min-latency. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is to showcase the option to generate multiple suffixes based on a previous request.
-We are going to hit the KV-Cache of a previous request.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-qwen-qwen3-8b-min-latency-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-qwen-qwen3-8b-min-latency-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="my-model",
-    messages=[
-        {
-            "role": "system",
-            "content": (
-                "You are an unhelpful assistant. To each math question, add +1 to the answer. "
-                "e.g. Whats 1+1 -> 3."
-            ),
-        }
-    ],
-    temperature=0.3,
-    max_tokens=100,
-    extra_body={
-        "suffix_messages": [
-            # Gen 1
-            [{"role": "user", "content": "Whats 1+1"}],
-            # Gen 2
-            [{"role": "user", "content": "Whats 2+2"}],
-        ],
-        "chat_template_kwargs": {"enable_thinking": False},
-    },
-)
-
-print(response_chat)
-```
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-qwen-qwen3-8b-min-latency-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-8B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    max_num_tokens: 32768
-    max_seq_len: 32768
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 32
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      num_draft_tokens: 61
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: false
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/custom-engine-builder-control/config.yaml b/custom-engine-builder-control/config.yaml
deleted file mode 100644
index 0e7459e85..000000000
--- a/custom-engine-builder-control/config.yaml
+++ /dev/null
@@ -1,48 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    max_tokens: 100
-    messages:
-      - content: You are a helpful assistant. To each math question, e.g. <Whats 1+1>, respond with all parts the question, '=', and answer.
-        role: system
-    stream: false
-    suffix_messages: [
-        # k=1
-        [{ "role": "system", "content": "Whats 1+1" }],
-        # k=2
-        [{ "role": "system", "content": "Whats 2+2" }],
-      ]
-    temperature: 0.5
-    chat_template_kwargs: { "enable_thinking": false }
-  tags:
-    - openai-compatible
-model_name: Briton-suffix-fanout-qwen3-8B
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-8B
-      revision: main
-      source: HF
-    max_batch_size: 128
-    max_num_tokens: 32768
-    max_seq_len: 32768
-    num_builder_gpus: 1
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: false
diff --git a/custom-server/README.md b/custom-server/README.md
deleted file mode 100644
index 5d46d832d..000000000
--- a/custom-server/README.md
+++ /dev/null
@@ -1,9 +0,0 @@
-## Deploying on Baseten with `start_command`
-
-If you've deployed on [Baseten](https://baseten.co) before, you're likely familiar with using [Truss](https://docs.baseten.co/quickstart) to do so.
-
-This process typically involves creating a `model.py` file, which contains the code for your model, and packaging it for deployment via an API endpoint managed by Baseten.
-
-However, there are situations where you might want to deploy a model that's already wrapped in an API. A common example is the `vLLM` OpenAI-Compatible Server, which provides its own HTTP endpoint. Another scenario is the "bring your own image" approach, where you have a Docker image that can handle HTTP requests directly.
-
-In these cases, setting up an HTTP endpoint through Truss or writing a `model.py` file can introduce unnecessary overhead. To streamline this, we've introduced the `docker_server.start_command` option, allowing you to specify an alternative `docker start` command, avoiding the need for additional setup.
diff --git a/custom-server/deepseek-v2-5-instruct-sglang/config.yaml b/custom-server/deepseek-v2-5-instruct-sglang/config.yaml
deleted file mode 100644
index e85af9083..000000000
--- a/custom-server/deepseek-v2-5-instruct-sglang/config.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-base_image:
-  image: lmsysorg/sglang:v0.4.0.post1-cu124
-model_metadata:
-  repo_id: deepseek-ai/DeepSeek-V2.5-1210
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path deepseek-ai/DeepSeek-V2.5-1210 --port 8000 --tp 8 --trust-remote-code"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/completions
-  server_port: 8000
-resources:
-  accelerator: H100:8
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: DeepSeek V2.5 1210 SGLang
-environment_variables:
-  hf_access_token: null
diff --git a/custom-server/infinity-embedding-server/README.md b/custom-server/infinity-embedding-server/README.md
deleted file mode 100644
index afedc3615..000000000
--- a/custom-server/infinity-embedding-server/README.md
+++ /dev/null
@@ -1,80 +0,0 @@
-# Infinity Embedding Server Truss
-
-This is a [Truss](https://truss.baseten.co/) to deploy [infinity embedding server](https://github.com/michaelfeil/infinity), a high-throughput, low-latency REST API server for serving vector embeddings.
-
-## Deployment
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. [Required for gated model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`.
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd custom-server/infinity-embedding-server
-```
-
-With `infinity-embedding-server` as your working directory, you can deploy the model with the following command, paste your Baseten API key if prompted.
-
-```sh
-truss push --publish --trusted
-```
-
-## Call your model
-
-### curl
-
-```bash
-curl -X POST https://model-xxx.api.baseten.co/development/predict \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string"}'
-```
-
-### request python library
-
-```python
-import requests
-
-resp = requests.post(
-    "https://model-xxx.api.baseten.co/development/predict",
-    headers={"Authorization": "Api-Key YOUR_API_KEY"},
-    json={"input": "text string"},
-)
-
-print(resp.json())
-```
-
-### openai python SDK
-
-```python
-import os
-from openai import OpenAI
-
-client = OpenAI(
-    api_key=os.environ["YOUR_API_KEY"],
-    base_url="https://bridge.baseten.co/v1/direct"
-)
-
-model_id = "xxx"
-deployment_id = "xxx"
-
-response = client.embeddings.create(
-    input="text string",
-    model="BAAI/bge-small-en-v1.5",
-    extra_body={
-        "baseten": {
-            "model_id": model_id,
-            "deployment_id": deployment_id
-        }
-    }
-)
-
-print(response.data[0].embedding)
-```
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/custom-server/infinity-embedding-server/config.yaml b/custom-server/infinity-embedding-server/config.yaml
deleted file mode 100644
index afe6c8f3e..000000000
--- a/custom-server/infinity-embedding-server/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-base_image:
-  image: python:3.11-slim
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) infinity_emb v2 --batch-size 64 --model-id BAAI/bge-small-en-v1.5 --revision main"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /embeddings
-  server_port: 7997
-build_commands: # optional step to download the weights of the model into the image
-- sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) infinity_emb v2 --preload-only --no-model-warmup --model-id BAAI/bge-small-en-v1.5 --revision main"
-resources:
-  accelerator: L4
-  use_gpu: true
-model_name: infinity-embedding-server
-requirements:
-- infinity-emb[all]==0.0.72
-runtime:
-  predict_concurrency : 40
-environment_variables:
-  hf_access_token: null
-  # constrain api to at most 256 sentences per request, for better load-balancing
-  INFINITY_MAX_CLIENT_BATCH_SIZE: 256
-  # constrain model to a max backpressure of INFINITY_MAX_CLIENT_BATCH_SIZE * predict_concurrency = 10241 requests
-  INFINITY_QUEUE_SIZE: 10241
-  DO_NOT_TRACK: 1
diff --git a/custom-server/llama3-70b-instruct-lmdeploy/config.yaml b/custom-server/llama3-70b-instruct-lmdeploy/config.yaml
deleted file mode 100644
index b80a57615..000000000
--- a/custom-server/llama3-70b-instruct-lmdeploy/config.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-base_image:
-  image: openmmlab/lmdeploy:v0.6.4-cu12
-model_metadata:
-  repo_id: meta-llama/Llama-3.1-70B-Instruct
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m lmdeploy serve api_server meta-llama/Llama-3.1-70B-Instruct --server-port 8000 --tp 4"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/completions
-  server_port: 8000
-resources:
-  accelerator: H100:4
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: Llama 3.1 70B Instruct LMDeploy
-environment_variables:
-  hf_access_token: null
diff --git a/custom-server/llama3-70b-instruct-sglang/config.yaml b/custom-server/llama3-70b-instruct-sglang/config.yaml
deleted file mode 100644
index b7bd233d0..000000000
--- a/custom-server/llama3-70b-instruct-sglang/config.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-base_image:
-  image: lmsysorg/sglang:v0.4.0.post1-cu124
-model_metadata:
-  repo_id: meta-llama/Llama-3.1-70B-Instruct
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path meta-llama/Llama-3.1-70B-Instruct --port 8000 --tp 4"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/completions
-  server_port: 8000
-resources:
-  accelerator: H100:4
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: Llama 3.1 70B Instruct SGLang
-environment_variables:
-  hf_access_token: null
diff --git a/custom-server/llama3-8b-instruct-lmdeploy/config.yaml b/custom-server/llama3-8b-instruct-lmdeploy/config.yaml
deleted file mode 100644
index a1d646eff..000000000
--- a/custom-server/llama3-8b-instruct-lmdeploy/config.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-base_image:
-  image: openmmlab/lmdeploy:v0.6.4-cu12
-model_metadata:
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m lmdeploy serve api_server meta-llama/Llama-3.1-8B-Instruct --server-port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/completions
-  server_port: 8000
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: Llama 3.1 8B Instruct LMDeploy
-environment_variables:
-  hf_access_token: null
diff --git a/custom-server/llama3-8b-instruct-sglang/config.yaml b/custom-server/llama3-8b-instruct-sglang/config.yaml
deleted file mode 100644
index bb22677c5..000000000
--- a/custom-server/llama3-8b-instruct-sglang/config.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-base_image:
-  image: lmsysorg/sglang:v0.4.0.post1-cu124
-model_metadata:
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path meta-llama/Llama-3.1-8B-Instruct --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/completions
-  server_port: 8000
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: Llama 3.1 8B Instruct SGLang
-environment_variables:
-  hf_access_token: null
diff --git a/custom-server/pixtral-12b/README.md b/custom-server/pixtral-12b/README.md
deleted file mode 100644
index acd17862f..000000000
--- a/custom-server/pixtral-12b/README.md
+++ /dev/null
@@ -1,85 +0,0 @@
-# Pixtral 12B Truss
-
-This is a [Truss](https://truss.baseten.co/) for Pixtral 12B. Pixtral is a 12B parameter language model released by [Mistral AI](https://mistral.ai/) and is a multimodal (text + vision) LLM. This Truss bypasses the need for writing a `model.py` and instead runs `vllm serve` directly at startup and uses the HTTP endpoint provided by the `vLLM` OpenAI Compatible Server to directly serve requests. This README will walk you through how to deploy this Truss on Baseten to get your own instance of it.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd custom-server/pixtral-12b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Accept the terms of service of the Pixtral model [here](https://huggingface.co/mistralai/Pixtral-12B-2409).
-4. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-5. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`. Note that you will *not* be able to successfully deploy Pixtral without doing this.
-
-With `pixtral-12b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-You need one A100 to run Pixtral at `bf16`.
-
-## Pixtral 12B API documentation
-
-This section provides an overview of the Pixtral 12B API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The `predict` route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __messages__: The input in OpenAI API format (see below for examples)
-- __stream__ (optional, default=False): A boolean determining whether the model should stream a response back. When `True`, the API returns generated text as it becomes available.
-- __max_tokens__ (optional, default=512): Maximum number of tokens to generate
-- __temperature__ (optional, default=0.7): A float between 0 and 1. Higher values means the generated output is more random while lower values means the generated output is more determenistic
-
-## Example usage
-
-```sh
-truss predict -d '{"model": "pixtral", "messages": [{"role": "user", "content": "Tell me about yourself"}]}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "model": "pixtral",
-           "messages": [
-            {
-                "role": "user",
-                "content": [
-                {
-                    "type": "text",
-                    "text": "What type of animal is this? Answer in French only"
-                },
-                {
-                    "type": "image_url",
-                    "image_url": {
-                    "url": "https://vetmed.illinois.edu/wp-content/uploads/2021/04/pc-keller-hedgehog.jpg"
-                    }
-                }
-                ]
-            }
-            ],
-           "stream": true,
-           "max_tokens": 64,
-           "temperature": 0.2
-         }' --no-buffer
-```
diff --git a/custom-server/pixtral-12b/config.yaml b/custom-server/pixtral-12b/config.yaml
deleted file mode 100644
index 22a9f1f5f..000000000
--- a/custom-server/pixtral-12b/config.yaml
+++ /dev/null
@@ -1,46 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.7.3
-model_metadata:
-  repo_id: mistralai/Pixtral-12B-2409
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  example_model_input: {
-    "model": "pixtral",
-    "messages": [
-      {
-      "role": "user",
-      "content": [
-        {
-        "type": "text",
-        "text": "Describe this image in one sentence."
-        },
-        {
-        "type": "image_url",
-        "image_url": {
-          "url": "https://picsum.photos/id/237/200/300"
-        }
-        }
-      ]
-      }
-    ],
-    "stream": false,
-    "max_tokens": 512,
-    "temperature": 0.5
-  }
-  tags:
-    - openai-compatible
-    - multimodal
-    - text-generation
-docker_server:
-  start_command: sh -c "vllm serve mistral-community/pixtral-12b --served-model-name pixtral --max-model-len 65536 --chat-template /app/data/pixtral12b.jinja --chat-template-content-format string --limit_mm_per_prompt 'image=4' --gpu-memory-utilization 0.95"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-runtime:
-  predict_concurrency : 16
-resources:
-  accelerator: H100
-  use_gpu: true
-model_name: Pixtral 12B
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
diff --git a/custom-server/ultravox-0.4/README.md b/custom-server/ultravox-0.4/README.md
deleted file mode 100644
index 363e05e4d..000000000
--- a/custom-server/ultravox-0.4/README.md
+++ /dev/null
@@ -1,80 +0,0 @@
-# Ultravox v0.4 vLLM Truss
-
-Ultravox is a multimodal model that can consume both speech and text as input, generating output text as usual.
-
-This is a [Truss](https://truss.baseten.co/) for Ultravox using the vLLM OpenAI Compatible server. This Truss bypasses the need for writing a `model.py` and instead runs `vllm serve` directly at startup and uses the HTTP endpoint provided by `vLLM` OpenAI Compatible Server to directly serve requests.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd custom-server/ultravox-0.4
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`. Note that you will *not* be able to successfully deploy Ultravox without doing this.
-
-With `ultravox-0.4` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-model_id = "jwdp26kw" # Replace with your model ID
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-response = client.chat.completions.create(
-    model="ultravox", # Replace with your model name
-    messages=[
-        {
-            "role": "user",
-            "content": [
-                {
-                    "type": "text",
-                    "text": "What is Lydia like?"
-                },
-                {
-                    "type": "audio_url",
-                    "audio_url": {"url": "https://baseten-public.s3.us-west-2.amazonaws.com/fred-audio-tests/real.mp3"}
-                }
-            ]
-        }
-    ],
-    stream=True
-)
-
-for chunk in response:
-    content = chunk.choices[0].delta.content
-    print(content, end="", flush=True)
-```
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/custom-server/ultravox-0.5-8b/README.md b/custom-server/ultravox-0.5-8b/README.md
deleted file mode 100644
index 1acce3e22..000000000
--- a/custom-server/ultravox-0.5-8b/README.md
+++ /dev/null
@@ -1,80 +0,0 @@
-# Ultravox v0.5 8B vLLM Truss
-
-Ultravox is a multimodal model that can consume both speech and text as input, generating output text as usual.
-
-This is a [Truss](https://truss.baseten.co/) for Ultravox using the vLLM OpenAI Compatible server. This Truss bypasses the need for writing a `model.py` and instead runs `vllm serve` directly at startup and uses the HTTP endpoint provided by `vLLM` OpenAI Compatible Server to directly serve requests.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd custom-server/ultravox-0.5-8b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`. Note that you will *not* be able to successfully deploy Ultravox without doing this.
-
-With `ultravox-0.5-8b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-model_id = "jwdp26kw" # Replace with your model ID
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-response = client.chat.completions.create(
-    model="ultravox", # Replace with your model name
-    messages=[
-        {
-            "role": "user",
-            "content": [
-                {
-                    "type": "text",
-                    "text": "What is Lydia like?"
-                },
-                {
-                    "type": "audio_url",
-                    "audio_url": {"url": "https://baseten-public.s3.us-west-2.amazonaws.com/fred-audio-tests/real.mp3"}
-                }
-            ]
-        }
-    ],
-    stream=True
-)
-
-for chunk in response:
-    content = chunk.choices[0].delta.content
-    print(content, end="", flush=True)
-```
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/custom-server/ultravox-0.6-70b/README.md b/custom-server/ultravox-0.6-70b/README.md
deleted file mode 100644
index 815b89ec9..000000000
--- a/custom-server/ultravox-0.6-70b/README.md
+++ /dev/null
@@ -1,80 +0,0 @@
-# Ultravox v0.6 70B vLLM Truss
-
-Ultravox is a multimodal model that can consume both speech and text as input, generating output text as usual.
-
-This is a [Truss](https://truss.baseten.co/) for Ultravox using the vLLM OpenAI Compatible server. This Truss bypasses the need for writing a `model.py` and instead runs `vllm serve` directly at startup and uses the HTTP endpoint provided by `vLLM` OpenAI Compatible Server to directly serve requests.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd custom-server/ultravox-0.6-70b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`. Note that you will *not* be able to successfully deploy Ultravox without doing this.
-
-With `ultravox-0.6-70b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-model_id = "jwdp26kw" # Replace with your model ID
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-response = client.chat.completions.create(
-    model="ultravox", # Replace with your model name
-    messages=[
-        {
-            "role": "user",
-            "content": [
-                {
-                    "type": "text",
-                    "text": "What is Lydia like?"
-                },
-                {
-                    "type": "audio_url",
-                    "audio_url": {"url": "https://baseten-public.s3.us-west-2.amazonaws.com/fred-audio-tests/real.mp3"}
-                }
-            ]
-        }
-    ],
-    stream=True
-)
-
-for chunk in response:
-    content = chunk.choices[0].delta.content
-    print(content, end="", flush=True)
-```
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/custom-server/voxtral-mini-3b-2507/README.md b/custom-server/voxtral-mini-3b-2507/README.md
deleted file mode 100644
index 1da9e95f5..000000000
--- a/custom-server/voxtral-mini-3b-2507/README.md
+++ /dev/null
@@ -1,125 +0,0 @@
-# Voxtral Mini 3B 2507 vLLM Truss
-
-Voxtral Mini is an enhancement of Ministral 3B, incorporating state-of-the-art audio input capabilities while retaining best-in-class text performance. It excels at speech transcription, translation and audio understanding.
-
-This is a [Truss](https://truss.baseten.co/) for Voxtral Mini using the vLLM OpenAI Compatible server. This Truss bypasses the need for writing a `model.py` and instead runs `vllm serve` directly at startup and uses the HTTP endpoint provided by `vLLM` OpenAI Compatible Server to directly serve requests.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd custom-server/voxtral-mini-3b-2507
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`. Note that you will _not_ be able to successfully deploy Voxtral Mini without doing this.
-
-With `voxtral-mini-3b-2507` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from mistral_common.protocol.instruct.messages import (
-    TextChunk,
-    AudioChunk,
-    UserMessage,
-    AssistantMessage,
-    RawAudio,
-)
-from mistral_common.audio import Audio
-from huggingface_hub import hf_hub_download
-
-from openai import OpenAI
-
-model_id = "12345678"
-
-client = OpenAI(
-    api_key="YOUR_API_KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/{deploy_env}/sync/v1"
-)
-
-models = client.models.list()
-model = models.data[0].id
-
-obama_file = hf_hub_download(
-    "patrickvonplaten/audio_samples", "obama.mp3", repo_type="dataset"
-)
-bcn_file = hf_hub_download(
-    "patrickvonplaten/audio_samples", "bcn_weather.mp3", repo_type="dataset"
-)
-
-def file_to_chunk(file: str) -> AudioChunk:
-    audio = Audio.from_file(file, strict=False)
-    return AudioChunk.from_audio(audio)
-
-text_chunk = TextChunk(
-    text="Which speaker is more inspiring? Why? How are they different from each other?"
-)
-user_msg = UserMessage(
-    content=[file_to_chunk(obama_file), file_to_chunk(bcn_file), text_chunk]
-).to_openai()
-
-print(30 * "=" + "USER 1" + 30 * "=")
-print(text_chunk.text)
-print("\n\n")
-
-response = client.chat.completions.create(
-    model=model,
-    messages=[user_msg],
-    temperature=0.2,
-    top_p=0.95,
-)
-content = response.choices[0].message.content
-
-print(30 * "=" + "BOT 1" + 30 * "=")
-print(content)
-print("\n\n")
-
-messages = [
-    user_msg,
-    AssistantMessage(content=content).to_openai(),
-    UserMessage(
-        content="Ok, now please summarize the content of the first audio."
-    ).to_openai(),
-]
-print(30 * "=" + "USER 2" + 30 * "=")
-print(messages[-1]["content"])
-print("\n\n")
-
-response = client.chat.completions.create(
-    model=model,
-    messages=messages,
-    temperature=0.2,
-    top_p=0.95,
-)
-content = response.choices[0].message.content
-print(30 * "=" + "BOT 2" + 30 * "=")
-print(content)
-
-```
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/custom-server/voxtral-mini-3b-2507/config.yaml b/custom-server/voxtral-mini-3b-2507/config.yaml
deleted file mode 100644
index 199fd5299..000000000
--- a/custom-server/voxtral-mini-3b-2507/config.yaml
+++ /dev/null
@@ -1,41 +0,0 @@
-description: Take in audio and text as input, generating text as usual
-base_image:
-  image: vllm/vllm-openai:v0.10.0
-model_metadata:
-  repo_id: mistralai/Voxtral-Mini-3B-2507
-  avatar_url: https://cdn-avatars.huggingface.co/v1/production/uploads/634c17653d11eaedd88b314d/9OgyfKstSZtbmsmuG8MbU.png
-  example_model_input:
-    {
-      "model": "voxtral-mini",
-      "messages":
-        [
-          {
-            "role": "user",
-            "content":
-              [
-                {
-                  "type": "text",
-                  "text": "What is the name of the famous bicycle race in France?",
-                },
-              ],
-          },
-        ],
-    }
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Voxtral-Mini-3B-2507 --tokenizer_mode mistral --config_format mistral --load_format mistral --port 8000 --served-model-name voxtral-mini"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-runtime:
-  predict_concurrency: 16
-model_name: Voxtral Mini 3B 2507
-secrets:
-  hf_access_token: null
-requirements:
-  - vllm[audio]
diff --git a/custom-server/voxtral-small-24b-2507/README.md b/custom-server/voxtral-small-24b-2507/README.md
deleted file mode 100644
index f2ddd4952..000000000
--- a/custom-server/voxtral-small-24b-2507/README.md
+++ /dev/null
@@ -1,125 +0,0 @@
-# Voxtral Small 24B 2507 vLLM Truss
-
-Voxtral Small is an enhancement of Mistral Small 3, incorporating state-of-the-art audio input capabilities while retaining best-in-class text performance. It excels at speech transcription, translation and audio understanding.
-
-This is a [Truss](https://truss.baseten.co/) for Voxtral Small using the vLLM OpenAI Compatible server. This Truss bypasses the need for writing a `model.py` and instead runs `vllm serve` directly at startup and uses the HTTP endpoint provided by `vLLM` OpenAI Compatible Server to directly serve requests.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd custom-server/voxtral-small-24b-2507
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-4. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`. Note that you will _not_ be able to successfully deploy Voxtral Small without doing this.
-
-With `voxtral-small-24b-2507` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) without the need for a `model.py` through the `docker_server.start_command` option.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from mistral_common.protocol.instruct.messages import (
-    TextChunk,
-    AudioChunk,
-    UserMessage,
-    AssistantMessage,
-    RawAudio,
-)
-from mistral_common.audio import Audio
-from huggingface_hub import hf_hub_download
-
-from openai import OpenAI
-
-model_id = "12345678"
-
-client = OpenAI(
-    api_key="YOUR_API_KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/{deploy_env}/sync/v1"
-)
-
-models = client.models.list()
-model = models.data[0].id
-
-obama_file = hf_hub_download(
-    "patrickvonplaten/audio_samples", "obama.mp3", repo_type="dataset"
-)
-bcn_file = hf_hub_download(
-    "patrickvonplaten/audio_samples", "bcn_weather.mp3", repo_type="dataset"
-)
-
-def file_to_chunk(file: str) -> AudioChunk:
-    audio = Audio.from_file(file, strict=False)
-    return AudioChunk.from_audio(audio)
-
-text_chunk = TextChunk(
-    text="Which speaker is more inspiring? Why? How are they different from each other? Answer in French."
-)
-user_msg = UserMessage(
-    content=[file_to_chunk(obama_file), file_to_chunk(bcn_file), text_chunk]
-).to_openai()
-
-print(30 * "=" + "USER 1" + 30 * "=")
-print(text_chunk.text)
-print("\n\n")
-
-response = client.chat.completions.create(
-    model=model,
-    messages=[user_msg],
-    temperature=0.2,
-    top_p=0.95,
-)
-content = response.choices[0].message.content
-
-print(30 * "=" + "BOT 1" + 30 * "=")
-print(content)
-print("\n\n")
-
-messages = [
-    user_msg,
-    AssistantMessage(content=content).to_openai(),
-    UserMessage(
-        content="Ok, now please summarize the content of the first audio."
-    ).to_openai(),
-]
-print(30 * "=" + "USER 2" + 30 * "=")
-print(messages[-1]["content"])
-print("\n\n")
-
-response = client.chat.completions.create(
-    model=model,
-    messages=messages,
-    temperature=0.2,
-    top_p=0.95,
-)
-content = response.choices[0].message.content
-print(30 * "=" + "BOT 2" + 30 * "=")
-print(content)
-
-```
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/custom-server/voxtral-small-24b-2507/config.yaml b/custom-server/voxtral-small-24b-2507/config.yaml
deleted file mode 100644
index f85fbef5c..000000000
--- a/custom-server/voxtral-small-24b-2507/config.yaml
+++ /dev/null
@@ -1,41 +0,0 @@
-description: Take in audio and text as input, generating text as usual
-base_image:
-  image: vllm/vllm-openai:v0.10.0
-model_metadata:
-  repo_id: mistralai/Voxtral-Small-24B-2507
-  avatar_url: https://cdn-avatars.huggingface.co/v1/production/uploads/634c17653d11eaedd88b314d/9OgyfKstSZtbmsmuG8MbU.png
-  example_model_input:
-    {
-      "model": "voxtral-small",
-      "messages":
-        [
-          {
-            "role": "user",
-            "content":
-              [
-                {
-                  "type": "text",
-                  "text": "What is the name of the famous bicycle race in France?",
-                },
-              ],
-          },
-        ],
-    }
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Voxtral-Small-24B-2507 --tokenizer_mode mistral --config_format mistral --load_format mistral --port 8000 --served-model-name voxtral-small"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency: 16
-model_name: Voxtral Small 24B 2507
-secrets:
-  hf_access_token: null
-requirements:
-  - vllm[audio]
diff --git a/deepfloyd-xl/README.md b/deepfloyd-xl/README.md
deleted file mode 100644
index 0eda7c075..000000000
--- a/deepfloyd-xl/README.md
+++ /dev/null
@@ -1,83 +0,0 @@
-# DeepFloyd XL Truss
-
-This is a [Truss](https://truss.baseten.co/) for DeepFloyd-IF. DeepFloyd-IF is a pixel-based text-to-image triple-cascaded diffusion model that can generate pictures and sets a new state-of-the-art for photorealism and language understanding. The result is a highly efficient model that outperforms current state-of-the-art models, achieving a zero-shot FID-30K score of 6.66 on the COCO dataset.
-
-Model details:
-
-- Developed by: DeepFloyd, StabilityAI
-- Model type: pixel-based text-to-image cascaded diffusion model
-- Cascade Stage: I
-- Num Parameters: 4.3B
-- Language(s): primarily English and, to a lesser extent, other Romance languages
-- License: [DeepFloyd IF License Agreement](https://huggingface.co/spaces/DeepFloyd/deepfloyd-if-license)
-- Model Description: DeepFloyd-IF is modular composed of frozen text mode and three pixel cascaded diffusion modules, each designed to generate images of increasing resolution: 64x64, 256x256, and 1024x1024. All stages of the model utilize a frozen text encoder based on the T5 transformer to extract text embeddings, which are then fed into a UNet architecture enhanced with cross-attention and attention-pooling
-
-Before deploying this model, you'll need to:
-
-1. Accept the terms of service of the Deepfloyd XL model [here](https://huggingface.co/DeepFloyd/IF-I-XL-v1.0).
-2. Retrieve your Huggingface token from the [settings](https://huggingface.co/settings/tokens).
-3. Set your Huggingface token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`.
-
-## Deploying DeepFloyd XL
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd deepfloyd-xl-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `deepfloyd-xl-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## DeepFloyd API documentation
-
-### Input
-
-This deployment of DeepFloyd takes a dictionary as input, which requires the following key:
-
-- `prompt` - the prompt for image generation
-
-It also supports a number of other parameters detailed in [this blog post](https://huggingface.co/blog/if).
-
-### Output
-
-The result will be a dictionary containing:
-
-- `status` - either `success` or `failed`
-- `data` - list of base 64 encoded images
-- `message` - will contain details in the case of errors
-
-```json
-{
-  "status": "success",
-  "data": ["/9j/4AAQSkZJRgABAQAAAQABAA...."],
-  "message": null
-}
-```
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "man on moon"}'
-```
-
-You can also invoke it via cURL:
-
-```sh
-curl -X POST https://app.baseten.co/models/EqwKvqa/predict \
-  -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-  -d '{"prompt": "man on moon"}'
-```
diff --git a/deepfloyd-xl/config.yaml b/deepfloyd-xl/config.yaml
deleted file mode 100644
index f61ffd196..000000000
--- a/deepfloyd-xl/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-description: Generate original images from text prompts.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/deep-floyd.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/deepfloyd-cover.png
-  tags:
-  - image-generation
-model_name: Deepfloyd XL
-python_version: py39
-requirements:
-- diffusers
-- transformers
-- torch
-- scipy
-- accelerate
-- pillow
-- bitsandbytes
-- sentencepiece
-- huggingface_hub
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets:
-  hf_access_token: ENTER HF API KEY HERE
-spec_version: 2.0
-system_packages: []
diff --git a/deepseek-ocr/README.md b/deepseek-ocr/README.md
deleted file mode 100644
index 07aaa92f2..000000000
--- a/deepseek-ocr/README.md
+++ /dev/null
@@ -1,81 +0,0 @@
-# DeepSeek-OCR Truss Model
-
-This is a Truss deployment of the DeepSeek OCR model for optical character recognition using vLLM engine on Baseten served on a H100_40G. The model excels at reading handwritten text, documents, and complex layouts with bounding box detection.
-
-DeepSeek-OCR processes 200k+ pages/day on a single GPU or 33M pages/day on 20 nodes. It requires 10x fewer visual tokens than text tokens, which means OCR compresses information 10x more efficiently than the text, with decoding precision of 97%. This makes it an excellent model for generating training data as well as potential tasks that involve long-context windows and memory.
-
-## Quick Start
-
-### 1. Deploy to Baseten
-
-```bash
-# Set your Baseten API key
-export BASETEN_API_KEY="your_api_key_here"
-
-# Clone this repo and cd into this folder
-git clone https://github.com/basetenlabs/truss-examples.git
-cd truss-examples/deepseek-ocr
-
-# Deploy the model
-truss push --publish
-# This assumes you have truss installed, if not follow the instructions here:
-# https://docs.baseten.co/development/model/build-your-first-model
-```
-
-### 2. Test with Sample Image
-
-Replace line 16 `ENDPOINT_URL` in `test_document_ocr.py` with your specific deployment URL.
-```bash
-# Run the test script with Bad-Handwriting.png
-python test_document_ocr.py
-```
-
-This will:
-- Load `Bad-Handwriting.png` (a challenging handwriting sample)
-- Test 5 different OCR prompts
-- Generate visualizations with bounding boxes
-- Save results as `visualization_*.png` files
-
-### 3. Project Structure
-
-```
-deepseek-ocr/
-├── config.yaml              # Truss configuration
-├── model/
-│   ├── __init__.py
-│   └── model.py            # Main model implementation
-├── README.md                # Documentation
-├── test_document_ocr.py     # Working test script
-├── visualizer.py           # Bounding box visualization
-├── Bad-Handwriting.png     # Test image
-└── visualization_*.png     # Generated outputs
-```
-
-## Model Information
-
-- **Model**: DeepSeek OCR v1
-- **Framework**: vLLM + PyTorch + Transformers
-- **GPU**: H100_40GB recommended
-- **Memory**: 16GB RAM
-- **Engine**: AsyncLLMEngine with custom DeepseekOCRForCausalLM
-
-## Usage
-
-### Using the Test Script
-
-The `test_document_ocr.py` script is the main testing interface and provides a complete example of how to use the model:
-
-```python
-# The script tests these prompts:
-prompts = [
-    "<image>\n<|grounding|>Convert the document to markdown.",
-    "<image>\n<|grounding|>OCR this image.",
-    "<image>\nFree OCR.",
-    "<image>\nParse the figure.",
-    "<image>\nDescribe this image in detail.",
-]
-```
-
-**Recommended**: Use `<image>\n<|grounding|>Convert the document to markdown.` for best results with bounding boxes.
-
-For detailed API documentation and examples, see the `test_document_ocr.py` script.
diff --git a/deepseek-ocr/config.yaml b/deepseek-ocr/config.yaml
deleted file mode 100644
index 70f8eddc8..000000000
--- a/deepseek-ocr/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    model: "deepseek-ai/DeepSeek-OCR"
-    messages:
-      - role: user
-        content:
-          - type: image_url
-            image_url:
-              url: "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg=="
-          - type: text
-            text: "<|grounding|>Convert the document to markdown."
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: deepseek-ocr-latest
-base_image:
-  image: lmsysorg/sglang@sha256:bb19265cdc61a65a158b84fb69d84f885f4c5f55e12e4515be88223ec067cf50
-docker_server:
-  start_command: sh -c "python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-OCR --served-model-name deepseek-ai/DeepSeek-OCR --host 0.0.0.0 --port 8000"
-  readiness_endpoint: /health_generate
-  liveness_endpoint: /health_generate
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-runtime:
-  predict_concurrency: 256
diff --git a/deepseek/deepseek-vl2/config.yaml b/deepseek/deepseek-vl2/config.yaml
deleted file mode 100644
index 2aee62d71..000000000
--- a/deepseek/deepseek-vl2/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    model: "deepseek-ai/deepseek-vl2"
-    messages:
-      - role: user
-        content:
-          - type: image_url
-            image_url:
-              url: "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg=="
-          - type: text
-            text: "<|grounding|>Convert the document to markdown."
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: deepseek vl2
-base_image:
-  image: lmsysorg/sglang:v0.5.5
-docker_server:
-  start_command: sh -c "python3 -m sglang.launch_server --model deepseek-ai/deepseek-vl2 --served-model-name deepseek-ai/deepseek-vl2 --host 0.0.0.0 --port 8000"
-  readiness_endpoint: /health_generate
-  liveness_endpoint: /health_generate
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency: 256
diff --git a/deepseek/engine-deepseek-r1-distill-llama-70b/README.md b/deepseek/engine-deepseek-r1-distill-llama-70b/README.md
deleted file mode 100644
index bf3784d87..000000000
--- a/deepseek/engine-deepseek-r1-distill-llama-70b/README.md
+++ /dev/null
@@ -1 +0,0 @@
-# DeepSeek-R1 Distill Llama 70B
diff --git a/deepseek/engine-deepseek-r1-distill-llama-70b/config.yaml b/deepseek/engine-deepseek-r1-distill-llama-70b/config.yaml
deleted file mode 100644
index 943f6f9ed..000000000
--- a/deepseek/engine-deepseek-r1-distill-llama-70b/config.yaml
+++ /dev/null
@@ -1,52 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "user",
-            content: "Which is heavier, a pound of bricks or a pound of feathers?",
-          },
-        ],
-      stream: true,
-      max_tokens: 1024,
-      temperature: 0.6,
-      top_p: 1.0,
-      top_k: 40,
-      frequency_penalty: 1,
-    }
-  repo_id: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-model_name: DeepSeek R1 Distill Llama 70B
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:2
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
-      source: HF
-    num_builder_gpus: 4
-    quantization_type: fp8_kv
-    max_seq_len: 131072
-    tensor_parallel_count: 2
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: true
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 131072
diff --git a/deepseek/engine-deepseek-r1-distill-llama-8b/README.md b/deepseek/engine-deepseek-r1-distill-llama-8b/README.md
deleted file mode 100644
index c55238dc2..000000000
--- a/deepseek/engine-deepseek-r1-distill-llama-8b/README.md
+++ /dev/null
@@ -1 +0,0 @@
-# DeepSeek-R1 Distill Llama 8B
diff --git a/deepseek/engine-deepseek-r1-distill-llama-8b/config.yaml b/deepseek/engine-deepseek-r1-distill-llama-8b/config.yaml
deleted file mode 100644
index db7448e4c..000000000
--- a/deepseek/engine-deepseek-r1-distill-llama-8b/config.yaml
+++ /dev/null
@@ -1,52 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "user",
-            content: "Which is heavier, a pound of bricks or a pound of feathers?",
-          },
-        ],
-      stream: true,
-      max_tokens: 1024,
-      temperature: 0.6,
-      top_p: 1.0,
-      top_k: 40,
-      frequency_penalty: 1,
-    }
-  repo_id: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-model_name: DeepSeek R1 Distill Llama 8B
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    max_seq_len: 131072
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 131072
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-14b/README.md b/deepseek/engine-deepseek-r1-distill-qwen-14b/README.md
deleted file mode 100644
index abd78652a..000000000
--- a/deepseek/engine-deepseek-r1-distill-qwen-14b/README.md
+++ /dev/null
@@ -1 +0,0 @@
-# DeepSeek-R1 Distill Qwen 14B
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml b/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml
deleted file mode 100644
index 8a3b06001..000000000
--- a/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml
+++ /dev/null
@@ -1,48 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "user",
-            content: "Which is heavier, a pound of bricks or a pound of feathers?",
-          },
-        ],
-      stream: true,
-      max_tokens: 1024,
-      temperature: 0.6,
-    }
-  repo_id: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
-model_name: DeepSeek R1 Distill Qwen 14B
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-32b/README.md b/deepseek/engine-deepseek-r1-distill-qwen-32b/README.md
deleted file mode 100644
index e051dfc88..000000000
--- a/deepseek/engine-deepseek-r1-distill-qwen-32b/README.md
+++ /dev/null
@@ -1 +0,0 @@
-# DeepSeek-R1 Distill Qwen 32B
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml b/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml
deleted file mode 100644
index e4eab5457..000000000
--- a/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml
+++ /dev/null
@@ -1,48 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "user",
-            content: "Which is heavier, a pound of bricks or a pound of feathers?",
-          },
-        ],
-      stream: true,
-      max_tokens: 1024,
-      temperature: 0.6,
-    }
-  repo_id: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
-model_name: DeepSeek R1 Distill Qwen 32B
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
-      source: HF
-    num_builder_gpus: 2
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-7b/README.md b/deepseek/engine-deepseek-r1-distill-qwen-7b/README.md
deleted file mode 100644
index 97210c5df..000000000
--- a/deepseek/engine-deepseek-r1-distill-qwen-7b/README.md
+++ /dev/null
@@ -1 +0,0 @@
-# DeepSeek-R1 Distill Qwen 7B
diff --git a/deepseek/engine-deepseek-r1-distill-qwen-7b/config.yaml b/deepseek/engine-deepseek-r1-distill-qwen-7b/config.yaml
deleted file mode 100644
index 5d93c82a1..000000000
--- a/deepseek/engine-deepseek-r1-distill-qwen-7b/config.yaml
+++ /dev/null
@@ -1,48 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "user",
-            content: "Which is heavier, a pound of bricks or a pound of feathers?",
-          },
-        ],
-      stream: true,
-      max_tokens: 1024,
-      temperature: 0.6,
-    }
-  repo_id: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
-model_name: DeepSeek R1 Distill Qwen 7B
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/deepspeed-mii/README.md b/deepspeed-mii/README.md
deleted file mode 100644
index 2322c411b..000000000
--- a/deepspeed-mii/README.md
+++ /dev/null
@@ -1,76 +0,0 @@
-# Llama-2-chat 7B DeepSpeed MII Truss
-
-This is a [Truss](https://truss.baseten.co/) for Llama-2-chat 7B served with [DeepSpeed MII](https://github.com/microsoft/DeepSpeed-MII). Llama 2 is a family of language models released by Meta. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Llama-2-chat 7B.
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp)). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-### Get Llama 2 access
-
-Llama 2 currently requires approval to access. To request access:
-
-1. Go to [https://ai.meta.com/resources/models-and-libraries/llama-downloads/](https://ai.meta.com/resources/models-and-libraries/llama-downloads/) and request access using the email associated with your HuggingFace account.
-2. Go to [https://huggingface.co/meta-llama/Llama-2-7b](https://huggingface.co/meta-llama/Llama-2-7b) and request access.
-
-Once you have Llama access:
-
-1. Create a [HuggingFace access token](https://huggingface.co/settings/tokens)
-2. Set it as a [secret in your Baseten account](https://app.baseten.co/settings/secrets) with the name `hf_access_token`
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd deepspeed-mii
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `deepspeed-mii` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-This seven billion parameter model is running in `float16` so that it fits on an A10G.
-
-## Llama-2-chat 7B API documentation
-
-This section provides an overview of the Llama-2-chat 7B API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __max_tokens__ (optional, default=512): The maximum number of tokens to return, counting input tokens. Maximum of 4096.
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "What is the meaning of life?", "max_tokens": 1024}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "What's the meaning of life?",
-           "max_tokens": 1024
-         }'
-```
diff --git a/deepspeed-mii/config.yaml b/deepspeed-mii/config.yaml
deleted file mode 100644
index 44f423609..000000000
--- a/deepspeed-mii/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-description: Generate text from a prompt with this seven billion parameter language
-  model.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/llama.png
-  example_model_input:
-    max_tokens: 1024
-    prompt: What's the meaning of life?
-  max_length: 4096
-  repo_id: meta-llama/Llama-2-7b-chat-hf
-  tags:
-  - text-generation
-  tensor_parallel: 1
-model_name: Llama-2-chat 7B DeepSpeed MII
-python_version: py311
-requirements:
-- deepspeed-mii==0.1.1
-resources:
-  accelerator: A100
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-runtime:
-  predict_concurrency: 256
-secrets:
-  hf_access_token: null
-system_packages:
-- cuda-toolkit-12-2
diff --git a/deepspeed-mii/model/model.py b/deepspeed-mii/model/model.py
deleted file mode 100644
index 50d4e45eb..000000000
--- a/deepspeed-mii/model/model.py
+++ /dev/null
@@ -1,79 +0,0 @@
-import asyncio
-import logging
-import queue
-import threading
-from typing import Any, Dict
-
-import mii
-from huggingface_hub import login
-
-DEFAULT_RESPONSE_MAX_LENGTH = 512
-
-
-class Model:
-    def __init__(self, **kwargs) -> None:
-        self.hf_access_token = kwargs["secrets"]["hf_access_token"]
-        model_metadata = kwargs["config"]["model_metadata"]
-        self.repo_id = model_metadata["repo_id"]
-        self.max_length = int(model_metadata["max_length"])
-        self.tensor_parallel = int(model_metadata["tensor_parallel"])
-        self.is_live_reload = kwargs["config"]["live_reload"]
-
-    def load(self):
-        login(token=self.hf_access_token)
-        # need to create a new loop because `mii.serve` creates async client at the end,
-        # and `load` function being called from new thread
-        asyncio.set_event_loop(asyncio.new_event_loop())
-        try:
-            mii.serve(
-                self.repo_id,
-                tensor_parallel=self.tensor_parallel,
-                max_length=self.max_length,
-            )
-        except Exception as e:
-            if self.is_live_reload:
-                # in live reload `mii.serve` fails after reload since server is running already
-                logging.info(
-                    "An exception occurred while starting mii server: %s, ignoring the exception due to live reload enabled",
-                    e,
-                )
-            else:
-                raise e
-
-    def predict(self, request: Dict):
-        prompt = request.pop("prompt")
-        generate_args = {
-            "max_new_tokens": request.pop("max_tokens", DEFAULT_RESPONSE_MAX_LENGTH),
-            "ignore_eos": request.pop("ignore_eos", False),
-        }
-
-        if request.pop("stream", False):
-            return self.stream(prompt, generate_args)
-        else:
-            # we need to create new asyncio loop because each request is being server from new thread
-            new_loop = asyncio.new_event_loop()
-            asyncio.set_event_loop(new_loop)
-            client = mii.client(self.repo_id)
-            response = client.generate(prompt, **generate_args)
-            new_loop.close()
-
-            return {"text": "\n".join(response.response)}
-
-    def stream(self, prompt: str, generate_args: Dict[str, Any]):
-        q = queue.Queue()
-
-        def generate():
-            new_loop = asyncio.new_event_loop()
-            asyncio.set_event_loop(new_loop)
-            client = mii.client(self.repo_id)
-            client.generate(prompt, streaming_fn=q.put, **generate_args)
-            q.put(None)
-
-        threading.Thread(target=generate).start()
-
-        while True:
-            item = q.get()  # This will block until an item is available
-            if item is not None:
-                yield "".join(item.response)
-            else:
-                break
diff --git a/dis-segmentation/README.md b/dis-segmentation/README.md
deleted file mode 100644
index f6b09639f..000000000
--- a/dis-segmentation/README.md
+++ /dev/null
@@ -1,104 +0,0 @@
-# DIS Segmentation Truss
-
-This is a [Truss](https://truss.baseten.co/welcome) for [DIS](https://github.com/xuebinqin/DIS/tree/main)
-
-This model can be used to remove backgrounds from an image or create a segmentation mask for an object.
-
-This model takes an input image and creates two output images:
-1. This first output image is the original image with the background removed
-2. The second output image contains the mask of the object in the foreground
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd dis-segmentation
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `dis-segmentation` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### API route: `predict`
-
-The model only has one input:
-
-- `input_image` (required): An image represented as a base74 string.
-
-The output of the model contains two images in the form of a base64 string.
-Example model output:
-```json
-{
-    "img_without_bg": "Base64 string image",
-    "image_mask": "Base64 string image"
-}
-```
-
-## Example usage
-
-You can invoke the model in Python like so:
-
-``` python
-import os
-import json
-import base64
-import requests
-
-# Set essential values
-model_id = ""
-baseten_api_key = ""
-
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/development/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json={
-        "input_image": pil_to_b64(Image.open("path/to/image.png"))
-    }
-)
-# Get output image
-res = res.json()
-background_removed = b64_to_pil(res.get("img_without_bg"))
-mask = b64_to_pil(res.get("image_mask"))
-background_removed.save("image_without_background.png")
-mask.save("image_mask.png")
-```
-
-Here is an example of the outputs given the following input.
-
-Input image:
-
-![input image](https://github.com/basetenlabs/truss-examples/assets/15642666/a0491a6f-795b-4e2a-aa66-8830f4fb86b3)
-
-Image without background:
-
-![removed-bg](https://github.com/basetenlabs/truss-examples/assets/15642666/a94ceba8-9a62-4e38-9fbe-f6f9382c8086)
-
-Image mask:
-
-![mask](https://github.com/basetenlabs/truss-examples/assets/15642666/14bcaea1-2136-4cdb-b896-bbed9e99fa89)
diff --git a/dis-segmentation/config.yaml b/dis-segmentation/config.yaml
deleted file mode 100644
index 86c3e7e79..000000000
--- a/dis-segmentation/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    input_image: <base64 image string>
-model_name: DIS Segmentation
-python_version: py310
-requirements:
-- torch==2.1.0
-- Pillow==9.4.0
-- numpy==1.23.5
-- gdown==4.7.3
-- torchvision==0.16.0
-- torchaudio==2.1.0
-- scikit-image==0.19.3
-resources:
-  accelerator: T4
-  memory: 2Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/embeddings/README.md b/embeddings/README.md
new file mode 100644
index 000000000..727683129
--- /dev/null
+++ b/embeddings/README.md
@@ -0,0 +1,18 @@
+# Embeddings
+
+Truss configurations for text embedding models, rerankers, and classifiers. Includes both TensorRT-optimized and HuggingFace TEI-based deployments covering a broad set of embedding providers.
+
+| Directory | Models | Description |
+|-----------|--------|-------------|
+| [bei](bei/) | 50 | Baseten Embeddings Infrastructure -- TensorRT-optimized embedding, reranking, and classification models from providers including BGE, GTE, Nomic, Jina, Qwen 3, Snowflake, and more |
+| [tei](tei/) | 15 | HuggingFace Text Embeddings Inference server configurations for embedding and reranking models |
+| [clip](clip/) | 1 | OpenAI CLIP model for image and text embeddings |
+| [text-embeddings-inference](text-embeddings-inference/) | 1 | Standalone Text Embeddings Inference server configuration |
+
+## Deploying
+
+Each embedding model can be deployed to Baseten with:
+
+```bash
+truss push
+```
diff --git a/embeddings/bei/alibaba-nlp-gte-modernbert-base-embedding/README.md b/embeddings/bei/alibaba-nlp-gte-modernbert-base-embedding/README.md
new file mode 100644
index 000000000..db6d5f697
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-modernbert-base-embedding/README.md
@@ -0,0 +1,30 @@
+# Alibaba-NLP GTE ModernBERT Base Embedding
+
+Deploy [Alibaba-NLP/gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Alibaba-NLP/gte-modernbert-base"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/alibaba-nlp-gte-modernbert-base-embedding/config.yaml b/embeddings/bei/alibaba-nlp-gte-modernbert-base-embedding/config.yaml
new file mode 100644
index 000000000..b1fa8af5a
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-modernbert-base-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-modernbert-base embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-alibaba-nlp-gte-modernbert-base-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: Alibaba-NLP/gte-modernbert-base
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md b/embeddings/bei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
new file mode 100644
index 000000000..642fb2cda
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
@@ -0,0 +1,30 @@
+# Alibaba-NLP GTE Qwen2 1.5B Instruct Embedding
+
+Deploy [Alibaba-NLP/gte-Qwen2-1.5B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-Qwen2-1.5B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Alibaba-NLP/gte-Qwen2-1.5B-instruct"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml b/embeddings/bei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
new file mode 100644
index 000000000..339d0e512
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-Qwen2-1.5B-instruct embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: Alibaba-NLP/gte-Qwen2-1.5B-instruct
+      revision: main
+      source: HF
+    max_num_tokens: 131072
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md b/embeddings/bei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
new file mode 100644
index 000000000..b729aeb90
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
@@ -0,0 +1,30 @@
+# Alibaba-NLP GTE Qwen2 7B Instruct Embedding
+
+Deploy [Alibaba-NLP/gte-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Alibaba-NLP/gte-Qwen2-7B-instruct"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml b/embeddings/bei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
new file mode 100644
index 000000000..f8d611ee0
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-Qwen2-7B-instruct embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: Alibaba-NLP/gte-Qwen2-7B-instruct
+      revision: main
+      source: HF
+    max_num_tokens: 131072
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/alibaba-nlp-gte-reranker-modernbert-base/README.md b/embeddings/bei/alibaba-nlp-gte-reranker-modernbert-base/README.md
new file mode 100644
index 000000000..4bb975fba
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-reranker-modernbert-base/README.md
@@ -0,0 +1,30 @@
+# Alibaba-NLP GTE Reranker ModernBERT Base
+
+Deploy [Alibaba-NLP/gte-reranker-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-reranker-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/alibaba-nlp-gte-reranker-modernbert-base/config.yaml b/embeddings/bei/alibaba-nlp-gte-reranker-modernbert-base/config.yaml
new file mode 100644
index 000000000..8f935dea3
--- /dev/null
+++ b/embeddings/bei/alibaba-nlp-gte-reranker-modernbert-base/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-reranker-modernbert-base reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    query: What is Baseten?
+    raw_scores: true
+    return_text: true
+    texts:
+    - Deep Learning is ...
+    - Baseten is a fast inference provider
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-Bert-alibaba-nlp-gte-reranker-modernbert-base-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: Alibaba-NLP/gte-reranker-modernbert-base
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /rerank
diff --git a/embeddings/bei/allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md b/embeddings/bei/allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md
new file mode 100644
index 000000000..9d5df943f
--- /dev/null
+++ b/embeddings/bei/allenai-llama-3.1-tulu-3-8b-reward-model-fp8/README.md
@@ -0,0 +1,43 @@
+# AllenAI Llama 3.1 Tulu 3 8B Reward Model
+
+Deploy [allenai/Llama-3.1-Tulu-3-8B-RM](https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B-RM) for classification using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [allenai/Llama-3.1-Tulu-3-8B-RM](https://huggingface.co/allenai/Llama-3.1-Tulu-3-8B-RM) |
+| Task | Classification |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "inputs": [
+    [
+      "Baseten is a fast inference provider"
+    ],
+    [
+      "Classify this separately."
+    ]
+  ],
+  "raw_scores": true,
+  "truncate": true,
+  "truncation_direction": "Right"
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml b/embeddings/bei/allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml
new file mode 100644
index 000000000..ab5866f63
--- /dev/null
+++ b/embeddings/bei/allenai-llama-3.1-tulu-3-8b-reward-model-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "allenai/Llama-3.1-Tulu-3-8B-RM classifier model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-allenai-llama-3.1-tulu-3-8b-reward-model-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: allenai/Llama-3.1-Tulu-3-8B-RM
+      revision: main
+      source: HF
+    max_num_tokens: 131072
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/baai-bge-en-icl-embedding-fp8/README.md b/embeddings/bei/baai-bge-en-icl-embedding-fp8/README.md
new file mode 100644
index 000000000..b3cb98355
--- /dev/null
+++ b/embeddings/bei/baai-bge-en-icl-embedding-fp8/README.md
@@ -0,0 +1,31 @@
+# BAAI BGE EN ICL Embedding
+
+Deploy [BAAI/bge-en-icl](https://huggingface.co/BAAI/bge-en-icl) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-en-icl](https://huggingface.co/BAAI/bge-en-icl) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "BAAI/bge-en-icl"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/baai-bge-en-icl-embedding-fp8/config.yaml b/embeddings/bei/baai-bge-en-icl-embedding-fp8/config.yaml
new file mode 100644
index 000000000..77b9c39b6
--- /dev/null
+++ b/embeddings/bei/baai-bge-en-icl-embedding-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-en-icl embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-baai-bge-en-icl-embedding-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: BAAI/bge-en-icl
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 2
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/baai-bge-large-en-v1.5-embedding/README.md b/embeddings/bei/baai-bge-large-en-v1.5-embedding/README.md
new file mode 100644
index 000000000..727e6c107
--- /dev/null
+++ b/embeddings/bei/baai-bge-large-en-v1.5-embedding/README.md
@@ -0,0 +1,30 @@
+# BAAI BGE Large EN v1.5 Embedding
+
+Deploy [BAAI/bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-large-en-v1.5](https://huggingface.co/BAAI/bge-large-en-v1.5) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "BAAI/bge-large-en-v1.5"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/baai-bge-large-en-v1.5-embedding/config.yaml b/embeddings/bei/baai-bge-large-en-v1.5-embedding/config.yaml
new file mode 100644
index 000000000..4336bcf45
--- /dev/null
+++ b/embeddings/bei/baai-bge-large-en-v1.5-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-large-en-v1.5 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-baai-bge-large-en-v1.5-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: BAAI/bge-large-en-v1.5
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/baai-bge-m3-embedding-dense/README.md b/embeddings/bei/baai-bge-m3-embedding-dense/README.md
new file mode 100644
index 000000000..7b197b215
--- /dev/null
+++ b/embeddings/bei/baai-bge-m3-embedding-dense/README.md
@@ -0,0 +1,30 @@
+# BAAI BGE M3 Embedding Dense
+
+Deploy [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-m3](https://huggingface.co/BAAI/bge-m3) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "BAAI/bge-m3"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/baai-bge-m3-embedding-dense/config.yaml b/embeddings/bei/baai-bge-m3-embedding-dense/config.yaml
new file mode 100644
index 000000000..10edada5d
--- /dev/null
+++ b/embeddings/bei/baai-bge-m3-embedding-dense/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-m3 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-baai-bge-m3-embedding-dense-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: BAAI/bge-m3
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/baai-bge-multilingual-gemma2-multilingual-embedding/README.md b/embeddings/bei/baai-bge-multilingual-gemma2-multilingual-embedding/README.md
new file mode 100644
index 000000000..c7fda032e
--- /dev/null
+++ b/embeddings/bei/baai-bge-multilingual-gemma2-multilingual-embedding/README.md
@@ -0,0 +1,30 @@
+# BAAI BGE Multilingual Gemma2 Embedding
+
+Deploy [BAAI/bge-multilingual-gemma2](https://huggingface.co/BAAI/bge-multilingual-gemma2) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-multilingual-gemma2](https://huggingface.co/BAAI/bge-multilingual-gemma2) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "BAAI/bge-multilingual-gemma2"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml b/embeddings/bei/baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml
new file mode 100644
index 000000000..b13aef4a3
--- /dev/null
+++ b/embeddings/bei/baai-bge-multilingual-gemma2-multilingual-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-multilingual-gemma2 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-baai-bge-multilingual-gemma2-multilingual-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: BAAI/bge-multilingual-gemma2
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/baai-bge-reranker-large/BEI-Bert-baai-bge-reranker-large/README.md b/embeddings/bei/baai-bge-reranker-large/BEI-Bert-baai-bge-reranker-large/README.md
new file mode 100644
index 000000000..fad7cb7b0
--- /dev/null
+++ b/embeddings/bei/baai-bge-reranker-large/BEI-Bert-baai-bge-reranker-large/README.md
@@ -0,0 +1,30 @@
+# BAAI BGE Reranker Large
+
+Deploy [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/baai-bge-reranker-large/BEI-Bert-baai-bge-reranker-large/config.yaml b/embeddings/bei/baai-bge-reranker-large/BEI-Bert-baai-bge-reranker-large/config.yaml
new file mode 100644
index 000000000..8294976a6
--- /dev/null
+++ b/embeddings/bei/baai-bge-reranker-large/BEI-Bert-baai-bge-reranker-large/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-reranker-large reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    query: What is Baseten?
+    raw_scores: true
+    return_text: true
+    texts:
+    - Deep Learning is ...
+    - Baseten is a fast inference provider
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-Bert-baai-bge-reranker-large-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: BAAI/bge-reranker-large
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /rerank
diff --git a/embeddings/bei/baai-bge-reranker-large/README.md b/embeddings/bei/baai-bge-reranker-large/README.md
new file mode 100644
index 000000000..d0db506d5
--- /dev/null
+++ b/embeddings/bei/baai-bge-reranker-large/README.md
@@ -0,0 +1,30 @@
+# BAAI BGE Reranker Large
+
+Deploy [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/baai-bge-reranker-large/config.yaml b/embeddings/bei/baai-bge-reranker-large/config.yaml
new file mode 100644
index 000000000..b02f57fda
--- /dev/null
+++ b/embeddings/bei/baai-bge-reranker-large/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-reranker-large reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    query: What is Baseten?
+    raw_scores: true
+    return_text: true
+    texts:
+    - Deep Learning is ...
+    - Baseten is a fast inference provider
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-baai-bge-reranker-large-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: BAAI/bge-reranker-large
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /rerank
diff --git a/embeddings/bei/baai-bge-reranker-v2-m3-multilingual/README.md b/embeddings/bei/baai-bge-reranker-v2-m3-multilingual/README.md
new file mode 100644
index 000000000..7740d4f55
--- /dev/null
+++ b/embeddings/bei/baai-bge-reranker-v2-m3-multilingual/README.md
@@ -0,0 +1,30 @@
+# BAAI BGE Reranker v2 M3
+
+Deploy [BAAI/bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-reranker-v2-m3](https://huggingface.co/BAAI/bge-reranker-v2-m3) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/baai-bge-reranker-v2-m3-multilingual/config.yaml b/embeddings/bei/baai-bge-reranker-v2-m3-multilingual/config.yaml
new file mode 100644
index 000000000..d8fca2e5e
--- /dev/null
+++ b/embeddings/bei/baai-bge-reranker-v2-m3-multilingual/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-reranker-v2-m3 reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    query: What is Baseten?
+    raw_scores: true
+    return_text: true
+    texts:
+    - Deep Learning is ...
+    - Baseten is a fast inference provider
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-baai-bge-reranker-v2-m3-multilingual-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: BAAI/bge-reranker-v2-m3
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /rerank
diff --git a/embeddings/bei/baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md b/embeddings/bei/baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md
new file mode 100644
index 000000000..e453ca27a
--- /dev/null
+++ b/embeddings/bei/baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/README.md
@@ -0,0 +1,43 @@
+# Baseten Example Meta Llama 3 70B InstructForSequenceClassification
+
+Deploy [baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification](https://huggingface.co/baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification) for classification using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification](https://huggingface.co/baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification) |
+| Task | Classification |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "inputs": [
+    [
+      "Baseten is a fast inference provider"
+    ],
+    [
+      "Classify this separately."
+    ]
+  ],
+  "raw_scores": true,
+  "truncate": true,
+  "truncation_direction": "Right"
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml b/embeddings/bei/baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml
new file mode 100644
index 000000000..97146ab5d
--- /dev/null
+++ b/embeddings/bei/baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification classifier model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-baseten-example-meta-llama-3-70b-instructforsequenceclassification-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: baseten/example-Meta-Llama-3-70B-InstructForSequenceClassification
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+    num_builder_gpus: 2
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/codefuse-ai-f2llm-4b-embedding-fp8/README.md b/embeddings/bei/codefuse-ai-f2llm-4b-embedding-fp8/README.md
new file mode 100644
index 000000000..6d17eb461
--- /dev/null
+++ b/embeddings/bei/codefuse-ai-f2llm-4b-embedding-fp8/README.md
@@ -0,0 +1,31 @@
+# Codefuse-AI F2LLM 4B Embedding
+
+Deploy [codefuse-ai/F2LLM-4B](https://huggingface.co/codefuse-ai/F2LLM-4B) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [codefuse-ai/F2LLM-4B](https://huggingface.co/codefuse-ai/F2LLM-4B) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "codefuse-ai/F2LLM-4B"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/codefuse-ai-f2llm-4b-embedding-fp8/config.yaml b/embeddings/bei/codefuse-ai-f2llm-4b-embedding-fp8/config.yaml
new file mode 100644
index 000000000..a19bf13b9
--- /dev/null
+++ b/embeddings/bei/codefuse-ai-f2llm-4b-embedding-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "codefuse-ai/F2LLM-4B embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-codefuse-ai-f2llm-4b-embedding-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: codefuse-ai/F2LLM-4B
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 2
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/google-embeddinggemma-300m/README.md b/embeddings/bei/google-embeddinggemma-300m/README.md
new file mode 100644
index 000000000..655480877
--- /dev/null
+++ b/embeddings/bei/google-embeddinggemma-300m/README.md
@@ -0,0 +1,30 @@
+# Google EmbeddingGemma 300M Embedding
+
+Deploy [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "google/embeddinggemma-300m"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/google-embeddinggemma-300m/config.yaml b/embeddings/bei/google-embeddinggemma-300m/config.yaml
new file mode 100644
index 000000000..ae39fa87f
--- /dev/null
+++ b/embeddings/bei/google-embeddinggemma-300m/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "google/embeddinggemma-300m embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-google-embeddinggemma-300m-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: google/embeddinggemma-300m
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md b/embeddings/bei/intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md
new file mode 100644
index 000000000..28c68545a
--- /dev/null
+++ b/embeddings/bei/intfloat-e5-mistral-7b-instruct-embedding-fp8/README.md
@@ -0,0 +1,31 @@
+# Intfloat E5 Mistral 7B Instruct Embedding
+
+Deploy [intfloat/e5-mistral-7b-instruct](https://huggingface.co/intfloat/e5-mistral-7b-instruct) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [intfloat/e5-mistral-7b-instruct](https://huggingface.co/intfloat/e5-mistral-7b-instruct) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "intfloat/e5-mistral-7b-instruct"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml b/embeddings/bei/intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml
new file mode 100644
index 000000000..1b7e09d6a
--- /dev/null
+++ b/embeddings/bei/intfloat-e5-mistral-7b-instruct-embedding-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "intfloat/e5-mistral-7b-instruct embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-intfloat-e5-mistral-7b-instruct-embedding-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: intfloat/e5-mistral-7b-instruct
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 2
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/intfloat-multilingual-e5-large-instruct/README.md b/embeddings/bei/intfloat-multilingual-e5-large-instruct/README.md
new file mode 100644
index 000000000..b1fd2d32a
--- /dev/null
+++ b/embeddings/bei/intfloat-multilingual-e5-large-instruct/README.md
@@ -0,0 +1,30 @@
+# Intfloat Multilingual E5 Large Instruct Embedding
+
+Deploy [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "intfloat/multilingual-e5-large-instruct"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/intfloat-multilingual-e5-large-instruct/config.yaml b/embeddings/bei/intfloat-multilingual-e5-large-instruct/config.yaml
new file mode 100644
index 000000000..21e4f4176
--- /dev/null
+++ b/embeddings/bei/intfloat-multilingual-e5-large-instruct/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "intfloat/multilingual-e5-large-instruct embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-intfloat-multilingual-e5-large-instruct-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: intfloat/multilingual-e5-large-instruct
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/jina-ai-jina-embeddings-v2-base-en/README.md b/embeddings/bei/jina-ai-jina-embeddings-v2-base-en/README.md
new file mode 100644
index 000000000..9e2410619
--- /dev/null
+++ b/embeddings/bei/jina-ai-jina-embeddings-v2-base-en/README.md
@@ -0,0 +1,30 @@
+# Jina AI Jina Embeddings v2 Base EN
+
+Deploy [jinaai/jina-embeddings-v2-base-en](https://huggingface.co/jinaai/jina-embeddings-v2-base-en) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [jinaai/jina-embeddings-v2-base-en](https://huggingface.co/jinaai/jina-embeddings-v2-base-en) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "jinaai/jina-embeddings-v2-base-en"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/jina-ai-jina-embeddings-v2-base-en/config.yaml b/embeddings/bei/jina-ai-jina-embeddings-v2-base-en/config.yaml
new file mode 100644
index 000000000..1253f9d1d
--- /dev/null
+++ b/embeddings/bei/jina-ai-jina-embeddings-v2-base-en/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "jinaai/jina-embeddings-v2-base-en embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-jina-ai-jina-embeddings-v2-base-en-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: jinaai/jina-embeddings-v2-base-en
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/jinaai-jina-code-embeddings-0.5b-fp8/README.md b/embeddings/bei/jinaai-jina-code-embeddings-0.5b-fp8/README.md
new file mode 100644
index 000000000..ef3e26f40
--- /dev/null
+++ b/embeddings/bei/jinaai-jina-code-embeddings-0.5b-fp8/README.md
@@ -0,0 +1,31 @@
+# Jina AI Jina Code Embeddings 0.5B
+
+Deploy [jinaai/jina-code-embeddings-0.5b](https://huggingface.co/jinaai/jina-code-embeddings-0.5b) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [jinaai/jina-code-embeddings-0.5b](https://huggingface.co/jinaai/jina-code-embeddings-0.5b) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "jinaai/jina-code-embeddings-0.5b"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/jinaai-jina-code-embeddings-0.5b-fp8/config.yaml b/embeddings/bei/jinaai-jina-code-embeddings-0.5b-fp8/config.yaml
new file mode 100644
index 000000000..534d1f269
--- /dev/null
+++ b/embeddings/bei/jinaai-jina-code-embeddings-0.5b-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "jinaai/jina-code-embeddings-0.5b embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-jinaai-jina-code-embeddings-0.5b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: jinaai/jina-code-embeddings-0.5b
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/jinaai-jina-embeddings-v2-base-code/README.md b/embeddings/bei/jinaai-jina-embeddings-v2-base-code/README.md
new file mode 100644
index 000000000..1a4cdd7dc
--- /dev/null
+++ b/embeddings/bei/jinaai-jina-embeddings-v2-base-code/README.md
@@ -0,0 +1,30 @@
+# Jina AI Jina Embeddings v2 Base Code
+
+Deploy [jinaai/jina-embeddings-v2-base-code](https://huggingface.co/jinaai/jina-embeddings-v2-base-code) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [jinaai/jina-embeddings-v2-base-code](https://huggingface.co/jinaai/jina-embeddings-v2-base-code) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "jinaai/jina-embeddings-v2-base-code"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/jinaai-jina-embeddings-v2-base-code/config.yaml b/embeddings/bei/jinaai-jina-embeddings-v2-base-code/config.yaml
new file mode 100644
index 000000000..54bdd69c8
--- /dev/null
+++ b/embeddings/bei/jinaai-jina-embeddings-v2-base-code/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "jinaai/jina-embeddings-v2-base-code embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-jinaai-jina-embeddings-v2-base-code-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: jinaai/jina-embeddings-v2-base-code
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
new file mode 100644
index 000000000..d57ec58b2
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
@@ -0,0 +1,30 @@
+# Mixedbread AI MxBAI Embed Large v1 Embedding
+
+Deploy [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "mixedbread-ai/mxbai-embed-large-v1"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
new file mode 100644
index 000000000..fc42152d7
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/BEI-mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "mixedbread-ai/mxbai-embed-large-v1 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: mixedbread-ai/mxbai-embed-large-v1
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
new file mode 100644
index 000000000..d57ec58b2
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
@@ -0,0 +1,30 @@
+# Mixedbread AI MxBAI Embed Large v1 Embedding
+
+Deploy [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "mixedbread-ai/mxbai-embed-large-v1"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
new file mode 100644
index 000000000..298732476
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "mixedbread-ai/mxbai-embed-large-v1 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: mixedbread-ai/mxbai-embed-large-v1
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md b/embeddings/bei/mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md
new file mode 100644
index 000000000..b19dda6f1
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/README.md
@@ -0,0 +1,31 @@
+# MxBAI Rerank Base v2 Reranker
+
+Deploy [michaelfeil/mxbai-rerank-base-v2-seq](https://huggingface.co/michaelfeil/mxbai-rerank-base-v2-seq) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/mxbai-rerank-base-v2-seq](https://huggingface.co/michaelfeil/mxbai-rerank-base-v2-seq) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml b/embeddings/bei/mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml
new file mode 100644
index 000000000..cf3b77360
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/mxbai-rerank-base-v2-seq reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-mixedbread-ai-mxbai-rerank-base-v2-reranker-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/mxbai-rerank-base-v2-seq
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 4
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md b/embeddings/bei/mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md
new file mode 100644
index 000000000..3b5ca02ca
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/README.md
@@ -0,0 +1,31 @@
+# MxBAI Rerank Large v2 Reranker
+
+Deploy [michaelfeil/mxbai-rerank-large-v2-seq](https://huggingface.co/michaelfeil/mxbai-rerank-large-v2-seq) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/mxbai-rerank-large-v2-seq](https://huggingface.co/michaelfeil/mxbai-rerank-large-v2-seq) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml b/embeddings/bei/mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml
new file mode 100644
index 000000000..e5739a2dc
--- /dev/null
+++ b/embeddings/bei/mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/mxbai-rerank-large-v2-seq reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-mixedbread-ai-mxbai-rerank-large-v2-reranker-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/mxbai-rerank-large-v2-seq
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 4
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/ncbi-medcpt-cross-encoder-reranker/README.md b/embeddings/bei/ncbi-medcpt-cross-encoder-reranker/README.md
new file mode 100644
index 000000000..ffbf0e99b
--- /dev/null
+++ b/embeddings/bei/ncbi-medcpt-cross-encoder-reranker/README.md
@@ -0,0 +1,30 @@
+# NCBI MedCPT Cross-Encoder Reranker
+
+Deploy [ncbi/MedCPT-Cross-Encoder](https://huggingface.co/ncbi/MedCPT-Cross-Encoder) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [ncbi/MedCPT-Cross-Encoder](https://huggingface.co/ncbi/MedCPT-Cross-Encoder) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/ncbi-medcpt-cross-encoder-reranker/config.yaml b/embeddings/bei/ncbi-medcpt-cross-encoder-reranker/config.yaml
new file mode 100644
index 000000000..e29873927
--- /dev/null
+++ b/embeddings/bei/ncbi-medcpt-cross-encoder-reranker/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "ncbi/MedCPT-Cross-Encoder reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    query: What is Baseten?
+    raw_scores: true
+    return_text: true
+    texts:
+    - Deep Learning is ...
+    - Baseten is a fast inference provider
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-ncbi-medcpt-cross-encoder-reranker-truss-example
+python_version: py39
+resources:
+  accelerator: A10G
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: ncbi/MedCPT-Cross-Encoder
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /rerank
diff --git a/embeddings/bei/ner-bert-base-ner-uncased/README.md b/embeddings/bei/ner-bert-base-ner-uncased/README.md
new file mode 100644
index 000000000..daa8077a0
--- /dev/null
+++ b/embeddings/bei/ner-bert-base-ner-uncased/README.md
@@ -0,0 +1,30 @@
+# BERT Base NER Uncased Classification
+
+Deploy [baseten-admin/bert-base-ner-uncased](https://huggingface.co/baseten-admin/bert-base-ner-uncased) for classification using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten-admin/bert-base-ner-uncased](https://huggingface.co/baseten-admin/bert-base-ner-uncased) |
+| Task | Classification |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/ner-bert-base-ner-uncased/config.yaml b/embeddings/bei/ner-bert-base-ner-uncased/config.yaml
new file mode 100644
index 000000000..c4daafbb1
--- /dev/null
+++ b/embeddings/bei/ner-bert-base-ner-uncased/config.yaml
@@ -0,0 +1,32 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BERT Base NER Uncased named entity recognition model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+      - - Apple is looking at buying U.K. startup for $1 billion
+      - - John works at Google in Mountain View, California
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-Bert-ner-bert-base-ner-uncased-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: baseten-admin/bert-base-ner-uncased
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /rerank
+  version_overrides:
+    engine_builder_version: null
+    bei_bert_version: 1.8.6.ner
diff --git a/embeddings/bei/nomic-ai-nomic-embed-code-fp8/README.md b/embeddings/bei/nomic-ai-nomic-embed-code-fp8/README.md
new file mode 100644
index 000000000..a153cc7bf
--- /dev/null
+++ b/embeddings/bei/nomic-ai-nomic-embed-code-fp8/README.md
@@ -0,0 +1,31 @@
+# Nomic AI Nomic Embed Code
+
+Deploy [nomic-ai/nomic-embed-code](https://huggingface.co/nomic-ai/nomic-embed-code) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nomic-ai/nomic-embed-code](https://huggingface.co/nomic-ai/nomic-embed-code) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "nomic-ai/nomic-embed-code"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/nomic-ai-nomic-embed-code-fp8/config.yaml b/embeddings/bei/nomic-ai-nomic-embed-code-fp8/config.yaml
new file mode 100644
index 000000000..7f967e657
--- /dev/null
+++ b/embeddings/bei/nomic-ai-nomic-embed-code-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "nomic-ai/nomic-embed-code embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-nomic-ai-nomic-embed-code-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: nomic-ai/nomic-embed-code
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/nomic-ai-nomic-embed-text-v1.5/README.md b/embeddings/bei/nomic-ai-nomic-embed-text-v1.5/README.md
new file mode 100644
index 000000000..4aced1333
--- /dev/null
+++ b/embeddings/bei/nomic-ai-nomic-embed-text-v1.5/README.md
@@ -0,0 +1,30 @@
+# Nomic AI Nomic Embed Text v1.5
+
+Deploy [nomic-ai/nomic-embed-text-v1.5](https://huggingface.co/nomic-ai/nomic-embed-text-v1.5) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nomic-ai/nomic-embed-text-v1.5](https://huggingface.co/nomic-ai/nomic-embed-text-v1.5) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "nomic-ai/nomic-embed-text-v1.5"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/nomic-ai-nomic-embed-text-v1.5/config.yaml b/embeddings/bei/nomic-ai-nomic-embed-text-v1.5/config.yaml
new file mode 100644
index 000000000..a1616f62d
--- /dev/null
+++ b/embeddings/bei/nomic-ai-nomic-embed-text-v1.5/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "nomic-ai/nomic-embed-text-v1.5 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-nomic-ai-nomic-embed-text-v1.5-truss-example
+python_version: py39
+resources:
+  accelerator: A10G
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: nomic-ai/nomic-embed-text-v1.5
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/nomic-ai-nomic-embed-text-v2-moe/README.md b/embeddings/bei/nomic-ai-nomic-embed-text-v2-moe/README.md
new file mode 100644
index 000000000..b836ddc4d
--- /dev/null
+++ b/embeddings/bei/nomic-ai-nomic-embed-text-v2-moe/README.md
@@ -0,0 +1,30 @@
+# Nomic AI Nomic Embed Text v2 MoE
+
+Deploy [nomic-ai/nomic-embed-text-v2-moe](https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nomic-ai/nomic-embed-text-v2-moe](https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "nomic-ai/nomic-embed-text-v2-moe"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/nomic-ai-nomic-embed-text-v2-moe/config.yaml b/embeddings/bei/nomic-ai-nomic-embed-text-v2-moe/config.yaml
new file mode 100644
index 000000000..5c263ec3f
--- /dev/null
+++ b/embeddings/bei/nomic-ai-nomic-embed-text-v2-moe/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "nomic-ai/nomic-embed-text-v2-moe embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-nomic-ai-nomic-embed-text-v2-moe-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: nomic-ai/nomic-embed-text-v2-moe
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/nvidia-llama-embed-nemotron-8b/README.md b/embeddings/bei/nvidia-llama-embed-nemotron-8b/README.md
new file mode 100644
index 000000000..dc93b9d02
--- /dev/null
+++ b/embeddings/bei/nvidia-llama-embed-nemotron-8b/README.md
@@ -0,0 +1,30 @@
+# NVIDIA Llama Embed Nemotron 8B
+
+Deploy [nvidia/llama-embed-nemotron-8b](https://huggingface.co/nvidia/llama-embed-nemotron-8b) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/llama-embed-nemotron-8b](https://huggingface.co/nvidia/llama-embed-nemotron-8b) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "nvidia/llama-embed-nemotron-8b"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/nvidia-llama-embed-nemotron-8b/config.yaml b/embeddings/bei/nvidia-llama-embed-nemotron-8b/config.yaml
new file mode 100644
index 000000000..bcf7bb234
--- /dev/null
+++ b/embeddings/bei/nvidia-llama-embed-nemotron-8b/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "nvidia/llama-embed-nemotron-8b embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-nvidia-llama-embed-nemotron-8b-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: nvidia/llama-embed-nemotron-8b
+      revision: main
+      source: HF
+    max_num_tokens: 131072
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/nvidia-llama-nemotron-embed-1b-v2/README.md b/embeddings/bei/nvidia-llama-nemotron-embed-1b-v2/README.md
new file mode 100644
index 000000000..88abdaa96
--- /dev/null
+++ b/embeddings/bei/nvidia-llama-nemotron-embed-1b-v2/README.md
@@ -0,0 +1,30 @@
+# NVIDIA Llama Nemotron Embed 1B v2
+
+Deploy [nvidia/llama-nemotron-embed-1b-v2](https://huggingface.co/nvidia/llama-nemotron-embed-1b-v2) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/llama-nemotron-embed-1b-v2](https://huggingface.co/nvidia/llama-nemotron-embed-1b-v2) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "nvidia/llama-nemotron-embed-1b-v2"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/nvidia-llama-nemotron-embed-1b-v2/config.yaml b/embeddings/bei/nvidia-llama-nemotron-embed-1b-v2/config.yaml
new file mode 100644
index 000000000..af49d4d68
--- /dev/null
+++ b/embeddings/bei/nvidia-llama-nemotron-embed-1b-v2/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "nvidia/llama-nemotron-embed-1b-v2 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-nvidia-llama-nemotron-embed-1b-v2-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: nvidia/llama-nemotron-embed-1b-v2
+      revision: main
+      source: HF
+    max_num_tokens: 131072
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/papluca-xlm-roberta-base-language-detection-classification/README.md b/embeddings/bei/papluca-xlm-roberta-base-language-detection-classification/README.md
new file mode 100644
index 000000000..574205f9d
--- /dev/null
+++ b/embeddings/bei/papluca-xlm-roberta-base-language-detection-classification/README.md
@@ -0,0 +1,42 @@
+# Papluca XLM-RoBERTa Base Language Detection Classification
+
+Deploy [papluca/xlm-roberta-base-language-detection](https://huggingface.co/papluca/xlm-roberta-base-language-detection) for classification using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [papluca/xlm-roberta-base-language-detection](https://huggingface.co/papluca/xlm-roberta-base-language-detection) |
+| Task | Classification |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "inputs": [
+    [
+      "Baseten is a fast inference provider"
+    ],
+    [
+      "Classify this separately."
+    ]
+  ],
+  "raw_scores": true,
+  "truncate": true,
+  "truncation_direction": "Right"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/papluca-xlm-roberta-base-language-detection-classification/config.yaml b/embeddings/bei/papluca-xlm-roberta-base-language-detection-classification/config.yaml
new file mode 100644
index 000000000..5c454a3f0
--- /dev/null
+++ b/embeddings/bei/papluca-xlm-roberta-base-language-detection-classification/config.yaml
@@ -0,0 +1,29 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "papluca/xlm-roberta-base-language-detection classifier model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-papluca-xlm-roberta-base-language-detection-classification-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: papluca/xlm-roberta-base-language-detection
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/qwen-qwen3-embedding-0.6b-fp8/README.md b/embeddings/bei/qwen-qwen3-embedding-0.6b-fp8/README.md
new file mode 100644
index 000000000..e447f59e5
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-0.6b-fp8/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Embedding 0.6B
+
+Deploy [michaelfeil/Qwen3-Embedding-0.6B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-0.6B-auto) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Embedding-0.6B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-0.6B-auto) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "michaelfeil/Qwen3-Embedding-0.6B-auto"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/qwen-qwen3-embedding-0.6b-fp8/config.yaml b/embeddings/bei/qwen-qwen3-embedding-0.6b-fp8/config.yaml
new file mode 100644
index 000000000..a3db4d55d
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-0.6b-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Embedding-0.6B-auto embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 4
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/qwen-qwen3-embedding-4b-fp4/README.md b/embeddings/bei/qwen-qwen3-embedding-4b-fp4/README.md
new file mode 100644
index 000000000..24da50130
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-4b-fp4/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Embedding 4B FP4
+
+Deploy [michaelfeil/Qwen3-Embedding-4B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-4B-auto) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Embedding-4B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-4B-auto) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | B200 |
+| Quantization | FP4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "michaelfeil/Qwen3-Embedding-4B-auto"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4**
diff --git a/embeddings/bei/qwen-qwen3-embedding-4b-fp4/config.yaml b/embeddings/bei/qwen-qwen3-embedding-4b-fp4/config.yaml
new file mode 100644
index 000000000..ee7823199
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-4b-fp4/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Embedding-4B-auto embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-qwen-qwen3-embedding-4b-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Embedding-4B-auto
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 1
+    quantization_type: fp4
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/qwen-qwen3-embedding-4b-fp8/README.md b/embeddings/bei/qwen-qwen3-embedding-4b-fp8/README.md
new file mode 100644
index 000000000..d810134b8
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-4b-fp8/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Embedding 4B FP8
+
+Deploy [michaelfeil/Qwen3-Embedding-4B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-4B-auto) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Embedding-4B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-4B-auto) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "michaelfeil/Qwen3-Embedding-4B-auto"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/qwen-qwen3-embedding-4b-fp8/config.yaml b/embeddings/bei/qwen-qwen3-embedding-4b-fp8/config.yaml
new file mode 100644
index 000000000..b0b0aa5e0
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-4b-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Embedding-4B-auto embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-qwen-qwen3-embedding-4b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Embedding-4B-auto
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/qwen-qwen3-embedding-8b-fp8/README.md b/embeddings/bei/qwen-qwen3-embedding-8b-fp8/README.md
new file mode 100644
index 000000000..8c58f9763
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-8b-fp8/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Embedding 8B
+
+Deploy [michaelfeil/Qwen3-Embedding-8B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-8B-auto) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Embedding-8B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-8B-auto) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "michaelfeil/Qwen3-Embedding-8B-auto"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/qwen-qwen3-embedding-8b-fp8/config.yaml b/embeddings/bei/qwen-qwen3-embedding-8b-fp8/config.yaml
new file mode 100644
index 000000000..8abbbd778
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-embedding-8b-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Embedding-8B-auto embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-qwen-qwen3-embedding-8b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Embedding-8B-auto
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/qwen-qwen3-reranker-0.6b-fp8/README.md b/embeddings/bei/qwen-qwen3-reranker-0.6b-fp8/README.md
new file mode 100644
index 000000000..33f1e95ea
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-0.6b-fp8/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Reranker 0.6B
+
+Deploy [michaelfeil/Qwen3-Reranker-0.6B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-0.6B-seq) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Reranker-0.6B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-0.6B-seq) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/qwen-qwen3-reranker-0.6b-fp8/config.yaml b/embeddings/bei/qwen-qwen3-reranker-0.6b-fp8/config.yaml
new file mode 100644
index 000000000..42d35697d
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-0.6b-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Reranker-0.6B-seq reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-qwen-qwen3-reranker-0.6b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Reranker-0.6B-seq
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 4
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/qwen-qwen3-reranker-4b-fp8/README.md b/embeddings/bei/qwen-qwen3-reranker-4b-fp8/README.md
new file mode 100644
index 000000000..848395237
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-4b-fp8/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Reranker 4B
+
+Deploy [michaelfeil/Qwen3-Reranker-4B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-4B-seq) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Reranker-4B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-4B-seq) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/qwen-qwen3-reranker-4b-fp8/config.yaml b/embeddings/bei/qwen-qwen3-reranker-4b-fp8/config.yaml
new file mode 100644
index 000000000..e06fcb410
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-4b-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Reranker-4B-seq reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-qwen-qwen3-reranker-4b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Reranker-4B-seq
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/qwen-qwen3-reranker-8b-fp4/README.md b/embeddings/bei/qwen-qwen3-reranker-8b-fp4/README.md
new file mode 100644
index 000000000..32cbc9dc0
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-8b-fp4/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Reranker 8B FP4
+
+Deploy [michaelfeil/Qwen3-Reranker-8B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-8B-seq) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Reranker-8B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-8B-seq) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | B200 |
+| Quantization | FP4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4**
diff --git a/embeddings/bei/qwen-qwen3-reranker-8b-fp4/config.yaml b/embeddings/bei/qwen-qwen3-reranker-8b-fp4/config.yaml
new file mode 100644
index 000000000..f955c1593
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-8b-fp4/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Reranker-8B-seq reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-qwen-qwen3-reranker-8b-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Reranker-8B-seq
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 1
+    quantization_type: fp4
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/qwen-qwen3-reranker-8b-fp8/README.md b/embeddings/bei/qwen-qwen3-reranker-8b-fp8/README.md
new file mode 100644
index 000000000..60d6aebb5
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-8b-fp8/README.md
@@ -0,0 +1,31 @@
+# Qwen3 Reranker 8B FP8
+
+Deploy [michaelfeil/Qwen3-Reranker-8B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-8B-seq) as a reranker using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/Qwen3-Reranker-8B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-8B-seq) |
+| Task | Reranking |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/qwen-qwen3-reranker-8b-fp8/config.yaml b/embeddings/bei/qwen-qwen3-reranker-8b-fp8/config.yaml
new file mode 100644
index 000000000..c807cc5fb
--- /dev/null
+++ b/embeddings/bei/qwen-qwen3-reranker-8b-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "michaelfeil/Qwen3-Reranker-8B-seq reranker model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-qwen-qwen3-reranker-8b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: michaelfeil/Qwen3-Reranker-8B-seq
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/redis-langcache-embed-v2/README.md b/embeddings/bei/redis-langcache-embed-v2/README.md
new file mode 100644
index 000000000..2bed91e4a
--- /dev/null
+++ b/embeddings/bei/redis-langcache-embed-v2/README.md
@@ -0,0 +1,30 @@
+# Redis LangCache Embed v2
+
+Deploy [redis/langcache-embed-v2](https://huggingface.co/redis/langcache-embed-v2) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [redis/langcache-embed-v2](https://huggingface.co/redis/langcache-embed-v2) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "redis/langcache-embed-v2"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/redis-langcache-embed-v2/config.yaml b/embeddings/bei/redis-langcache-embed-v2/config.yaml
new file mode 100644
index 000000000..dcefdf264
--- /dev/null
+++ b/embeddings/bei/redis-langcache-embed-v2/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "redis/langcache-embed-v2 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-redis-langcache-embed-v2-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: redis/langcache-embed-v2
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/salesforce-sfr-embedding-mistral-fp8/README.md b/embeddings/bei/salesforce-sfr-embedding-mistral-fp8/README.md
new file mode 100644
index 000000000..5c211bdd9
--- /dev/null
+++ b/embeddings/bei/salesforce-sfr-embedding-mistral-fp8/README.md
@@ -0,0 +1,31 @@
+# Salesforce SFR Embedding Mistral
+
+Deploy [Salesforce/SFR-Embedding-Mistral](https://huggingface.co/Salesforce/SFR-Embedding-Mistral) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Salesforce/SFR-Embedding-Mistral](https://huggingface.co/Salesforce/SFR-Embedding-Mistral) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Salesforce/SFR-Embedding-Mistral"}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/salesforce-sfr-embedding-mistral-fp8/config.yaml b/embeddings/bei/salesforce-sfr-embedding-mistral-fp8/config.yaml
new file mode 100644
index 000000000..74c92696f
--- /dev/null
+++ b/embeddings/bei/salesforce-sfr-embedding-mistral-fp8/config.yaml
@@ -0,0 +1,28 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Salesforce/SFR-Embedding-Mistral embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-salesforce-sfr-embedding-mistral-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: Salesforce/SFR-Embedding-Mistral
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/samlowe-roberta-base-go_emotions-classification/README.md b/embeddings/bei/samlowe-roberta-base-go_emotions-classification/README.md
new file mode 100644
index 000000000..8002a837c
--- /dev/null
+++ b/embeddings/bei/samlowe-roberta-base-go_emotions-classification/README.md
@@ -0,0 +1,42 @@
+# SamLowe RoBERTa Base GoEmotions Classification
+
+Deploy [SamLowe/roberta-base-go_emotions](https://huggingface.co/SamLowe/roberta-base-go_emotions) for classification using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [SamLowe/roberta-base-go_emotions](https://huggingface.co/SamLowe/roberta-base-go_emotions) |
+| Task | Classification |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "inputs": [
+    [
+      "Baseten is a fast inference provider"
+    ],
+    [
+      "Classify this separately."
+    ]
+  ],
+  "raw_scores": true,
+  "truncate": true,
+  "truncation_direction": "Right"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/samlowe-roberta-base-go_emotions-classification/config.yaml b/embeddings/bei/samlowe-roberta-base-go_emotions-classification/config.yaml
new file mode 100644
index 000000000..c0e5048af
--- /dev/null
+++ b/embeddings/bei/samlowe-roberta-base-go_emotions-classification/config.yaml
@@ -0,0 +1,29 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "SamLowe/roberta-base-go_emotions classifier model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-samlowe-roberta-base-go_emotions-classification-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: SamLowe/roberta-base-go_emotions
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/sentence-transformers-all-minilm-l6-v2-embedding/README.md b/embeddings/bei/sentence-transformers-all-minilm-l6-v2-embedding/README.md
new file mode 100644
index 000000000..ed45b7dcd
--- /dev/null
+++ b/embeddings/bei/sentence-transformers-all-minilm-l6-v2-embedding/README.md
@@ -0,0 +1,30 @@
+# Sentence Transformers All-MiniLM-L6-v2 Embedding
+
+Deploy [sentence-transformers/all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [sentence-transformers/all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "sentence-transformers/all-MiniLM-L6-v2"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/sentence-transformers-all-minilm-l6-v2-embedding/config.yaml b/embeddings/bei/sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
new file mode 100644
index 000000000..2e60c4570
--- /dev/null
+++ b/embeddings/bei/sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "sentence-transformers/all-MiniLM-L6-v2 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-sentence-transformers-all-minilm-l6-v2-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: sentence-transformers/all-MiniLM-L6-v2
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md b/embeddings/bei/skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md
new file mode 100644
index 000000000..d13af8b3b
--- /dev/null
+++ b/embeddings/bei/skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/README.md
@@ -0,0 +1,43 @@
+# Skywork Reward Llama 3.1 8B v0.2
+
+Deploy [Skywork/Skywork-Reward-Llama-3.1-8B-v0.2](https://huggingface.co/Skywork/Skywork-Reward-Llama-3.1-8B-v0.2) for classification using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Skywork/Skywork-Reward-Llama-3.1-8B-v0.2](https://huggingface.co/Skywork/Skywork-Reward-Llama-3.1-8B-v0.2) |
+| Task | Classification |
+| Engine | BEI (TensorRT) |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "inputs": [
+    [
+      "Baseten is a fast inference provider"
+    ],
+    [
+      "Classify this separately."
+    ]
+  ],
+  "raw_scores": true,
+  "truncate": true,
+  "truncation_direction": "Right"
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
diff --git a/embeddings/bei/skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml b/embeddings/bei/skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml
new file mode 100644
index 000000000..22829dae1
--- /dev/null
+++ b/embeddings/bei/skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Skywork/Skywork-Reward-Llama-3.1-8B-v0.2 classifier model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+    - - Baseten is a fast inference provider
+    - - Classify this separately.
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-skywork-skywork-reward-llama-3.1-8b-v0.2-reward-model-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: Skywork/Skywork-Reward-Llama-3.1-8B-v0.2
+      revision: main
+      source: HF
+    max_num_tokens: 131072
+    num_builder_gpus: 1
+    quantization_type: fp8
+  runtime:
+    webserver_default_route: /predict
diff --git a/embeddings/bei/snowflake-snowflake-arctic-embed-l-v2.0/README.md b/embeddings/bei/snowflake-snowflake-arctic-embed-l-v2.0/README.md
new file mode 100644
index 000000000..ecd0abe55
--- /dev/null
+++ b/embeddings/bei/snowflake-snowflake-arctic-embed-l-v2.0/README.md
@@ -0,0 +1,30 @@
+# Snowflake Arctic Embed L v2.0
+
+Deploy [Snowflake/snowflake-arctic-embed-l-v2.0](https://huggingface.co/Snowflake/snowflake-arctic-embed-l-v2.0) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Snowflake/snowflake-arctic-embed-l-v2.0](https://huggingface.co/Snowflake/snowflake-arctic-embed-l-v2.0) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Snowflake/snowflake-arctic-embed-l-v2.0"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/snowflake-snowflake-arctic-embed-l-v2.0/config.yaml b/embeddings/bei/snowflake-snowflake-arctic-embed-l-v2.0/config.yaml
new file mode 100644
index 000000000..394d3c0b0
--- /dev/null
+++ b/embeddings/bei/snowflake-snowflake-arctic-embed-l-v2.0/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Snowflake/snowflake-arctic-embed-l-v2.0 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-snowflake-snowflake-arctic-embed-l-v2.0-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: Snowflake/snowflake-arctic-embed-l-v2.0
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/tanaos-tanaos-ner-v1/README.md b/embeddings/bei/tanaos-tanaos-ner-v1/README.md
new file mode 100644
index 000000000..8219220db
--- /dev/null
+++ b/embeddings/bei/tanaos-tanaos-ner-v1/README.md
@@ -0,0 +1,30 @@
+# Tanaos NER v1
+
+Deploy [tanaos/tanaos-NER-v1](https://huggingface.co/tanaos/tanaos-NER-v1) for classification using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [tanaos/tanaos-NER-v1](https://huggingface.co/tanaos/tanaos-NER-v1) |
+| Task | Classification |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/tanaos-tanaos-ner-v1/config.yaml b/embeddings/bei/tanaos-tanaos-ner-v1/config.yaml
new file mode 100644
index 000000000..025cbdecf
--- /dev/null
+++ b/embeddings/bei/tanaos-tanaos-ner-v1/config.yaml
@@ -0,0 +1,32 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Tanaos NER v1 named entity recognition model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    inputs:
+      - - Apple is looking at buying U.K. startup for $1 billion
+      - - John works at Google in Mountain View, California
+    raw_scores: true
+    truncate: true
+    truncation_direction: Right
+model_name: BEI-Bert-tanaos-tanaos-ner-v1-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: tanaos/tanaos-NER-v1
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /rerank
+  version_overrides:
+    engine_builder_version: null
+    bei_bert_version: 1.8.6.ner
diff --git a/embeddings/bei/taylorai-bge-micro-v2/README.md b/embeddings/bei/taylorai-bge-micro-v2/README.md
new file mode 100644
index 000000000..07c3c3eca
--- /dev/null
+++ b/embeddings/bei/taylorai-bge-micro-v2/README.md
@@ -0,0 +1,30 @@
+# TaylorAI BGE Micro v2
+
+Deploy [TaylorAI/bge-micro-v2](https://huggingface.co/TaylorAI/bge-micro-v2) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [TaylorAI/bge-micro-v2](https://huggingface.co/TaylorAI/bge-micro-v2) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "TaylorAI/bge-micro-v2"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/taylorai-bge-micro-v2/config.yaml b/embeddings/bei/taylorai-bge-micro-v2/config.yaml
new file mode 100644
index 000000000..6f0f4848d
--- /dev/null
+++ b/embeddings/bei/taylorai-bge-micro-v2/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "TaylorAI/bge-micro-v2 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-taylorai-bge-micro-v2-truss-example
+python_version: py39
+resources:
+  accelerator: A10G
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: TaylorAI/bge-micro-v2
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/voyageai-voyage-4-nano/README.md b/embeddings/bei/voyageai-voyage-4-nano/README.md
new file mode 100644
index 000000000..9f95a2482
--- /dev/null
+++ b/embeddings/bei/voyageai-voyage-4-nano/README.md
@@ -0,0 +1,30 @@
+# VoyageAI Voyage 4 Nano
+
+Deploy [voyageai/voyage-4-nano](https://huggingface.co/voyageai/voyage-4-nano) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [voyageai/voyage-4-nano](https://huggingface.co/voyageai/voyage-4-nano) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "voyageai/voyage-4-nano"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/voyageai-voyage-4-nano/config.yaml b/embeddings/bei/voyageai-voyage-4-nano/config.yaml
new file mode 100644
index 000000000..944d2befc
--- /dev/null
+++ b/embeddings/bei/voyageai-voyage-4-nano/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "voyageai/voyage-4-nano embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-Bert-voyageai-voyage-4-nano-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_bert
+    checkpoint_repository:
+      repo: voyageai/voyage-4-nano
+      revision: main
+      source: HF
+    max_num_tokens: 40960
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/bei/whereisai-uae-large-v1-embedding/README.md b/embeddings/bei/whereisai-uae-large-v1-embedding/README.md
new file mode 100644
index 000000000..790e4cb91
--- /dev/null
+++ b/embeddings/bei/whereisai-uae-large-v1-embedding/README.md
@@ -0,0 +1,30 @@
+# WhereIsAI UAE Large V1 Embedding
+
+Deploy [WhereIsAI/UAE-Large-V1](https://huggingface.co/WhereIsAI/UAE-Large-V1) for generating text embeddings using a BEI (TensorRT) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [WhereIsAI/UAE-Large-V1](https://huggingface.co/WhereIsAI/UAE-Large-V1) |
+| Task | Embeddings |
+| Engine | BEI (TensorRT) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "WhereIsAI/UAE-Large-V1"}'
+```
+
+## Configuration highlights
+
+- Engine: **BEI (TensorRT)**
diff --git a/embeddings/bei/whereisai-uae-large-v1-embedding/config.yaml b/embeddings/bei/whereisai-uae-large-v1-embedding/config.yaml
new file mode 100644
index 000000000..2ae321c7f
--- /dev/null
+++ b/embeddings/bei/whereisai-uae-large-v1-embedding/config.yaml
@@ -0,0 +1,26 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "WhereIsAI/UAE-Large-V1 embedding model"
+model_metadata:
+  tags:
+    - force-legacy-api-non-openai-compatible
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-whereisai-uae-large-v1-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder
+    checkpoint_repository:
+      repo: WhereIsAI/UAE-Large-V1
+      revision: main
+      source: HF
+    max_num_tokens: 16384
+  runtime:
+    webserver_default_route: /v1/embeddings
diff --git a/embeddings/clip/README.md b/embeddings/clip/README.md
new file mode 100644
index 000000000..ce5715828
--- /dev/null
+++ b/embeddings/clip/README.md
@@ -0,0 +1,31 @@
+# clip-example
+
+Deploy clip-example for generating text embeddings using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Embeddings |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://images.pexels.com/photos/1170986/pexels-photo-1170986.jpeg?auto=compress&cs=tinysrgb&w=1600"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/embeddings/clip/config.yaml b/embeddings/clip/config.yaml
new file mode 100644
index 000000000..450b28a8f
--- /dev/null
+++ b/embeddings/clip/config.yaml
@@ -0,0 +1,20 @@
+description: "CLIP for image-text embedding"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "openai/clip-vit-large-patch14"
+  example_model_input:
+    url: https://images.pexels.com/photos/1170986/pexels-photo-1170986.jpeg?auto=compress&cs=tinysrgb&w=1600
+model_name: clip-example
+python_version: py311
+requirements:
+- transformers==4.47.1
+- pillow==10.1.0
+- torch==2.1.0
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/comfyui-truss/model/__init__.py b/embeddings/clip/model/__init__.py
similarity index 100%
rename from comfyui-truss/model/__init__.py
rename to embeddings/clip/model/__init__.py
diff --git a/clip/model/model.py b/embeddings/clip/model/model.py
similarity index 100%
rename from clip/model/model.py
rename to embeddings/clip/model/model.py
diff --git a/embeddings/tei/alibaba-nlp-gte-modernbert-base-embedding/README.md b/embeddings/tei/alibaba-nlp-gte-modernbert-base-embedding/README.md
new file mode 100644
index 000000000..821df1ca9
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-modernbert-base-embedding/README.md
@@ -0,0 +1,32 @@
+# Alibaba-NLP GTE ModernBERT Base Embedding
+
+Deploy [Alibaba-NLP/gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-modernbert-base) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Alibaba-NLP/gte-modernbert-base"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/alibaba-nlp-gte-modernbert-base-embedding/config.yaml b/embeddings/tei/alibaba-nlp-gte-modernbert-base-embedding/config.yaml
new file mode 100644
index 000000000..5833f5282
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-modernbert-base-embedding/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-modernbert-base embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: Alibaba-NLP/gte-modernbert-base
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-alibaba-nlp-gte-modernbert-base-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md b/embeddings/tei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
new file mode 100644
index 000000000..a3d32f5e9
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/README.md
@@ -0,0 +1,32 @@
+# Alibaba-NLP GTE Qwen2 1.5B Instruct Embedding
+
+Deploy [Alibaba-NLP/gte-Qwen2-1.5B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-Qwen2-1.5B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-1.5B-instruct) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Alibaba-NLP/gte-Qwen2-1.5B-instruct"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml b/embeddings/tei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
new file mode 100644
index 000000000..b7afa13fa
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-qwen2-1.5b-instruct-embedding/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-Qwen2-1.5B-instruct embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: Alibaba-NLP/gte-Qwen2-1.5B-instruct
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-alibaba-nlp-gte-qwen2-1.5b-instruct-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md b/embeddings/tei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
new file mode 100644
index 000000000..5eb543a6e
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/README.md
@@ -0,0 +1,32 @@
+# Alibaba-NLP GTE Qwen2 7B Instruct Embedding
+
+Deploy [Alibaba-NLP/gte-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-Qwen2-7B-instruct](https://huggingface.co/Alibaba-NLP/gte-Qwen2-7B-instruct) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "Alibaba-NLP/gte-Qwen2-7B-instruct"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:hopper-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml b/embeddings/tei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
new file mode 100644
index 000000000..bcd3dbfe6
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-qwen2-7b-instruct-embedding/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-Qwen2-7B-instruct embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:hopper-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: Alibaba-NLP/gte-Qwen2-7B-instruct
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-alibaba-nlp-gte-qwen2-7b-instruct-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/alibaba-nlp-gte-reranker-modernbert-base/README.md b/embeddings/tei/alibaba-nlp-gte-reranker-modernbert-base/README.md
new file mode 100644
index 000000000..46552d612
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-reranker-modernbert-base/README.md
@@ -0,0 +1,32 @@
+# Alibaba-NLP GTE Reranker ModernBERT Base
+
+Deploy [Alibaba-NLP/gte-reranker-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base) as a reranker using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Alibaba-NLP/gte-reranker-modernbert-base](https://huggingface.co/Alibaba-NLP/gte-reranker-modernbert-base) |
+| Task | Reranking |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/alibaba-nlp-gte-reranker-modernbert-base/config.yaml b/embeddings/tei/alibaba-nlp-gte-reranker-modernbert-base/config.yaml
new file mode 100644
index 000000000..6e5d26047
--- /dev/null
+++ b/embeddings/tei/alibaba-nlp-gte-reranker-modernbert-base/config.yaml
@@ -0,0 +1,43 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Alibaba-NLP/gte-reranker-modernbert-base reranker model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /rerank
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: Alibaba-NLP/gte-reranker-modernbert-base
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    query: What is Baseten?
+    raw_scores: true
+    return_text: true
+    texts:
+    - Deep Learning is ...
+    - Baseten is a fast inference provider
+    truncate: true
+    truncation_direction: Right
+model_name: TEI-alibaba-nlp-gte-reranker-modernbert-base-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/baai-bge-reranker-large/README.md b/embeddings/tei/baai-bge-reranker-large/README.md
new file mode 100644
index 000000000..87e674306
--- /dev/null
+++ b/embeddings/tei/baai-bge-reranker-large/README.md
@@ -0,0 +1,32 @@
+# BAAI BGE Reranker Large
+
+Deploy [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large) as a reranker using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-reranker-large](https://huggingface.co/BAAI/bge-reranker-large) |
+| Task | Reranking |
+| Engine | TEI (HuggingFace) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/rerank \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"query": "What is deep learning?", "texts": ["Deep learning is a subset of machine learning.", "The weather is nice today."], "raw_scores": true}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:hopper-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/baai-bge-reranker-large/config.yaml b/embeddings/tei/baai-bge-reranker-large/config.yaml
new file mode 100644
index 000000000..cabfa4eaf
--- /dev/null
+++ b/embeddings/tei/baai-bge-reranker-large/config.yaml
@@ -0,0 +1,43 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "BAAI/bge-reranker-large reranker model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:hopper-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /rerank
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: BAAI/bge-reranker-large
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    query: What is Baseten?
+    raw_scores: true
+    return_text: true
+    texts:
+    - Deep Learning is ...
+    - Baseten is a fast inference provider
+    truncate: true
+    truncation_direction: Right
+model_name: TEI-baai-bge-reranker-large-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/google-embeddinggemma-300m/README.md b/embeddings/tei/google-embeddinggemma-300m/README.md
new file mode 100644
index 000000000..a919671af
--- /dev/null
+++ b/embeddings/tei/google-embeddinggemma-300m/README.md
@@ -0,0 +1,32 @@
+# Google EmbeddingGemma 300M Embedding
+
+Deploy [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [google/embeddinggemma-300m](https://huggingface.co/google/embeddinggemma-300m) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "google/embeddinggemma-300m"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/google-embeddinggemma-300m/config.yaml b/embeddings/tei/google-embeddinggemma-300m/config.yaml
new file mode 100644
index 000000000..14fd09943
--- /dev/null
+++ b/embeddings/tei/google-embeddinggemma-300m/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "google/embeddinggemma-300m embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: google/embeddinggemma-300m
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-google-embeddinggemma-300m-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/intfloat-multilingual-e5-large-instruct/README.md b/embeddings/tei/intfloat-multilingual-e5-large-instruct/README.md
new file mode 100644
index 000000000..9611a669f
--- /dev/null
+++ b/embeddings/tei/intfloat-multilingual-e5-large-instruct/README.md
@@ -0,0 +1,32 @@
+# Intfloat Multilingual E5 Large Instruct Embedding
+
+Deploy [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [intfloat/multilingual-e5-large-instruct](https://huggingface.co/intfloat/multilingual-e5-large-instruct) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "intfloat/multilingual-e5-large-instruct"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/intfloat-multilingual-e5-large-instruct/config.yaml b/embeddings/tei/intfloat-multilingual-e5-large-instruct/config.yaml
new file mode 100644
index 000000000..7787cd1cd
--- /dev/null
+++ b/embeddings/tei/intfloat-multilingual-e5-large-instruct/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "intfloat/multilingual-e5-large-instruct embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: intfloat/multilingual-e5-large-instruct
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-intfloat-multilingual-e5-large-instruct-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/jina-ai-jina-embeddings-v2-base-en/README.md b/embeddings/tei/jina-ai-jina-embeddings-v2-base-en/README.md
new file mode 100644
index 000000000..3e7078bd1
--- /dev/null
+++ b/embeddings/tei/jina-ai-jina-embeddings-v2-base-en/README.md
@@ -0,0 +1,32 @@
+# Jina AI Jina Embeddings v2 Base EN
+
+Deploy [jinaai/jina-embeddings-v2-base-en](https://huggingface.co/jinaai/jina-embeddings-v2-base-en) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [jinaai/jina-embeddings-v2-base-en](https://huggingface.co/jinaai/jina-embeddings-v2-base-en) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "jinaai/jina-embeddings-v2-base-en"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/jina-ai-jina-embeddings-v2-base-en/config.yaml b/embeddings/tei/jina-ai-jina-embeddings-v2-base-en/config.yaml
new file mode 100644
index 000000000..f69abfc71
--- /dev/null
+++ b/embeddings/tei/jina-ai-jina-embeddings-v2-base-en/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "jinaai/jina-embeddings-v2-base-en embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: jinaai/jina-embeddings-v2-base-en
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-jina-ai-jina-embeddings-v2-base-en-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/jinaai-jina-embeddings-v2-base-code/README.md b/embeddings/tei/jinaai-jina-embeddings-v2-base-code/README.md
new file mode 100644
index 000000000..8a85a0556
--- /dev/null
+++ b/embeddings/tei/jinaai-jina-embeddings-v2-base-code/README.md
@@ -0,0 +1,32 @@
+# Jina AI Jina Embeddings v2 Base Code
+
+Deploy [jinaai/jina-embeddings-v2-base-code](https://huggingface.co/jinaai/jina-embeddings-v2-base-code) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [jinaai/jina-embeddings-v2-base-code](https://huggingface.co/jinaai/jina-embeddings-v2-base-code) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "jinaai/jina-embeddings-v2-base-code"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/jinaai-jina-embeddings-v2-base-code/config.yaml b/embeddings/tei/jinaai-jina-embeddings-v2-base-code/config.yaml
new file mode 100644
index 000000000..acd5f4826
--- /dev/null
+++ b/embeddings/tei/jinaai-jina-embeddings-v2-base-code/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "jinaai/jina-embeddings-v2-base-code embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: jinaai/jina-embeddings-v2-base-code
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-jinaai-jina-embeddings-v2-base-code-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/mixedbread-ai-mxbai-embed-large-v1-embedding/README.md b/embeddings/tei/mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
new file mode 100644
index 000000000..b2573e0f0
--- /dev/null
+++ b/embeddings/tei/mixedbread-ai-mxbai-embed-large-v1-embedding/README.md
@@ -0,0 +1,32 @@
+# Mixedbread AI MxBAI Embed Large v1 Embedding
+
+Deploy [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mixedbread-ai/mxbai-embed-large-v1](https://huggingface.co/mixedbread-ai/mxbai-embed-large-v1) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "mixedbread-ai/mxbai-embed-large-v1"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml b/embeddings/tei/mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
new file mode 100644
index 000000000..7c1cf267f
--- /dev/null
+++ b/embeddings/tei/mixedbread-ai-mxbai-embed-large-v1-embedding/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "mixedbread-ai/mxbai-embed-large-v1 embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: mixedbread-ai/mxbai-embed-large-v1
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-mixedbread-ai-mxbai-embed-large-v1-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/nomic-ai-nomic-embed-text-v1.5/README.md b/embeddings/tei/nomic-ai-nomic-embed-text-v1.5/README.md
new file mode 100644
index 000000000..cd90b155e
--- /dev/null
+++ b/embeddings/tei/nomic-ai-nomic-embed-text-v1.5/README.md
@@ -0,0 +1,32 @@
+# Nomic AI Nomic Embed Text v1.5
+
+Deploy [nomic-ai/nomic-embed-text-v1.5](https://huggingface.co/nomic-ai/nomic-embed-text-v1.5) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nomic-ai/nomic-embed-text-v1.5](https://huggingface.co/nomic-ai/nomic-embed-text-v1.5) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "nomic-ai/nomic-embed-text-v1.5"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:86-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/nomic-ai-nomic-embed-text-v1.5/config.yaml b/embeddings/tei/nomic-ai-nomic-embed-text-v1.5/config.yaml
new file mode 100644
index 000000000..4041a2b8e
--- /dev/null
+++ b/embeddings/tei/nomic-ai-nomic-embed-text-v1.5/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "nomic-ai/nomic-embed-text-v1.5 embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:86-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: nomic-ai/nomic-embed-text-v1.5
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-nomic-ai-nomic-embed-text-v1.5-truss-example
+python_version: py39
+resources:
+  accelerator: A10G
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/nomic-ai-nomic-embed-text-v2-moe/README.md b/embeddings/tei/nomic-ai-nomic-embed-text-v2-moe/README.md
new file mode 100644
index 000000000..2567e1724
--- /dev/null
+++ b/embeddings/tei/nomic-ai-nomic-embed-text-v2-moe/README.md
@@ -0,0 +1,32 @@
+# Nomic AI Nomic Embed Text v2 MoE
+
+Deploy [nomic-ai/nomic-embed-text-v2-moe](https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nomic-ai/nomic-embed-text-v2-moe](https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "nomic-ai/nomic-embed-text-v2-moe"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/nomic-ai-nomic-embed-text-v2-moe/config.yaml b/embeddings/tei/nomic-ai-nomic-embed-text-v2-moe/config.yaml
new file mode 100644
index 000000000..715bc3f8a
--- /dev/null
+++ b/embeddings/tei/nomic-ai-nomic-embed-text-v2-moe/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "nomic-ai/nomic-embed-text-v2-moe embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: nomic-ai/nomic-embed-text-v2-moe
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-nomic-ai-nomic-embed-text-v2-moe-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/redis-langcache-embed-v2/README.md b/embeddings/tei/redis-langcache-embed-v2/README.md
new file mode 100644
index 000000000..550801a53
--- /dev/null
+++ b/embeddings/tei/redis-langcache-embed-v2/README.md
@@ -0,0 +1,32 @@
+# Redis LangCache Embed v2
+
+Deploy [redis/langcache-embed-v2](https://huggingface.co/redis/langcache-embed-v2) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [redis/langcache-embed-v2](https://huggingface.co/redis/langcache-embed-v2) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "redis/langcache-embed-v2"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/redis-langcache-embed-v2/config.yaml b/embeddings/tei/redis-langcache-embed-v2/config.yaml
new file mode 100644
index 000000000..7ead7013c
--- /dev/null
+++ b/embeddings/tei/redis-langcache-embed-v2/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "redis/langcache-embed-v2 embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:89-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: redis/langcache-embed-v2
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-redis-langcache-embed-v2-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/sentence-transformers-all-minilm-l6-v2-embedding/README.md b/embeddings/tei/sentence-transformers-all-minilm-l6-v2-embedding/README.md
new file mode 100644
index 000000000..06d53d936
--- /dev/null
+++ b/embeddings/tei/sentence-transformers-all-minilm-l6-v2-embedding/README.md
@@ -0,0 +1,32 @@
+# Sentence Transformers All-MiniLM-L6-v2 Embedding
+
+Deploy [sentence-transformers/all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [sentence-transformers/all-MiniLM-L6-v2](https://huggingface.co/sentence-transformers/all-MiniLM-L6-v2) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | T4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "sentence-transformers/all-MiniLM-L6-v2"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:turing-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/sentence-transformers-all-minilm-l6-v2-embedding/config.yaml b/embeddings/tei/sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
new file mode 100644
index 000000000..1533f40a3
--- /dev/null
+++ b/embeddings/tei/sentence-transformers-all-minilm-l6-v2-embedding/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "sentence-transformers/all-MiniLM-L6-v2 embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:turing-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: sentence-transformers/all-MiniLM-L6-v2
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-sentence-transformers-all-minilm-l6-v2-embedding-truss-example
+python_version: py39
+resources:
+  accelerator: T4
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/tei/taylorai-bge-micro-v2/README.md b/embeddings/tei/taylorai-bge-micro-v2/README.md
new file mode 100644
index 000000000..fd2071a80
--- /dev/null
+++ b/embeddings/tei/taylorai-bge-micro-v2/README.md
@@ -0,0 +1,32 @@
+# TaylorAI BGE Micro v2
+
+Deploy [TaylorAI/bge-micro-v2](https://huggingface.co/TaylorAI/bge-micro-v2) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [TaylorAI/bge-micro-v2](https://huggingface.co/TaylorAI/bge-micro-v2) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "TaylorAI/bge-micro-v2"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:86-1.8.3`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/embeddings/tei/taylorai-bge-micro-v2/config.yaml b/embeddings/tei/taylorai-bge-micro-v2/config.yaml
new file mode 100644
index 000000000..267755b81
--- /dev/null
+++ b/embeddings/tei/taylorai-bge-micro-v2/config.yaml
@@ -0,0 +1,38 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "TaylorAI/bge-micro-v2 embedding model"
+base_image:
+  image: baseten/text-embeddings-inference-mirror:86-1.8.3
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/embeddings
+  readiness_endpoint: /health
+  server_port: 7997
+  start_command: bash -c "truss-transfer-cli && text-embeddings-router --port 7997
+    --model-id /app/model_cache/cached_model --max-client-batch-size 128 --max-concurrent-requests
+    1024 --max-batch-tokens 16384 --auto-truncate --tokenization-workers 3"
+model_cache:
+- ignore_patterns:
+  - '*.pt'
+  - '*.ckpt'
+  - '*.onnx'
+  repo_id: TaylorAI/bge-micro-v2
+  revision: main
+  use_volume: true
+  volume_folder: cached_model
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: TEI-taylorai-bge-micro-v2-truss-example
+python_version: py39
+resources:
+  accelerator: A10G
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/embeddings/text-embeddings-inference/README.md b/embeddings/text-embeddings-inference/README.md
new file mode 100644
index 000000000..8d39c0145
--- /dev/null
+++ b/embeddings/text-embeddings-inference/README.md
@@ -0,0 +1,33 @@
+# text-embeddings-inference trussless
+
+Deploy [BAAI/bge-base-en-v1.5](https://huggingface.co/BAAI/bge-base-en-v1.5) for generating text embeddings using a TEI (HuggingFace) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-base-en-v1.5](https://huggingface.co/BAAI/bge-base-en-v1.5) |
+| Task | Embeddings |
+| Engine | TEI (HuggingFace) |
+| GPU | L4 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"input": "What is deep learning?", "model": "BAAI/bge-base-en-v1.5"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/text-embeddings-inference-mirror:89-1.6`
+- Predict concurrency: **40**
+- Server port: **7997**
diff --git a/embeddings/text-embeddings-inference/config.yaml b/embeddings/text-embeddings-inference/config.yaml
new file mode 100644
index 000000000..048a9f3cc
--- /dev/null
+++ b/embeddings/text-embeddings-inference/config.yaml
@@ -0,0 +1,31 @@
+description: "BGE Base EN v1.5 via Text Embeddings Inference server"
+base_image:
+  # select an image: L4
+  # CPU	baseten/text-embeddings-inference-mirror:cpu-1.6
+  # Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.6
+  # Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.6
+  # Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.6
+  # Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.6
+  # Hopper (H100/H100 40GB)	baseten/text-embeddings-inference-mirror:hopper-1.6
+  image: baseten/text-embeddings-inference-mirror:89-1.6
+model_metadata:
+  repo_id: BAAI/bge-base-en-v1.5
+  example_model_input: {"input": "What is deep learning?", "model": "BAAI/bge-base-en-v1.5"}
+docker_server:
+  start_command: sh -c "text-embeddings-router --port 7997 --model-id /data/local-model --max-client-batch-size 32 --max-concurrent-requests 40 --max-batch-tokens 32768"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  # change to /rerank or /predict if you want to use the rerank or predict endpoint
+  # https://huggingface.github.io/text-embeddings-inference/
+  predict_endpoint: /v1/embeddings
+  server_port: 7997
+resources:
+  accelerator: L4
+  use_gpu: true
+model_name: text-embeddings-inference trussless
+build_commands: # optional step to download the weights of the model into the image
+- git clone https://huggingface.co/BAAI/bge-base-en-v1.5 /data/local-model
+runtime:
+  predict_concurrency : 40
+environment_variables:
+  hf_access_token: null
diff --git a/falcon/falcon3-10B-trt-llm-spec-dec/README.md b/falcon/falcon3-10B-trt-llm-spec-dec/README.md
deleted file mode 100644
index 15b89047e..000000000
--- a/falcon/falcon3-10B-trt-llm-spec-dec/README.md
+++ /dev/null
@@ -1,49 +0,0 @@
-# Falcon 3 10B Instruct using TensorRT-LLM with Speculative Decoding
-
-This directory is [Truss](https://truss.baseten.co/) template for deploying model Llama 3.1 8B Instruct using our TensorRT-LLM (TRT-LLM) [engine builder](https://docs.baseten.co/performance/engine-builder-overview). This configuration is optimized for high-throughput scenarios.
-
-## Use case
-
-This deployment is tailored for applications that require processing large volumes of data with moderate to low latency, such as:
-* Content moderation for social media platforms
-* Bulk article summarization
-* Large-scale Retrieval-Augmented Generation (RAG) systems
-
-## Configuration
-
-The template uses the following key configuration parameters:
-
-| Property            | Value    | Description                                                                    |
-| ------------------- | -------- | ------------------------------------------------------------------------------ |
-| GPU                 | 1xH100   | Single NVIDIA H100 GPU                                                         |
-| `max_batch_size`    | 32       | Allows processing up to 32 requests simultaneously                             |
-| `quantization_type` | `fp8_kv` | FP8 quantization, balancing performance and accuracy. See this [blog](https://www.baseten.co/blog/33-faster-llm-inference-with-fp8-quantization/). |
-| `max_input_len`     | 4096     | Maximum number of input tokens that the model will accept                      |
-| `max_output_len`    | 1024     | Maximum number of output tokens the model can generate                         |
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd falcon/falcon3-10B-trt-llm-spec-dec
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `falcon/falcon3-10B-trt-llm-spec-dec` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted --publish
-```
-
-Paste your Baseten API key if prompted. Also ensure the `hf_access_token` secret is properly setup in your Baseten Account to access this model.
-
-**Note**: TensorRT-LLM with engine builder will only work under a Baseten production deployment
-
-For more information, refer to the [Truss documentation](https://docs.baseten.co/performance/engine-builder-overview).
diff --git a/falcon/falcon3-3B-trt-llm-engine-high-throughput/README.md b/falcon/falcon3-3B-trt-llm-engine-high-throughput/README.md
deleted file mode 100644
index 9a832a539..000000000
--- a/falcon/falcon3-3B-trt-llm-engine-high-throughput/README.md
+++ /dev/null
@@ -1,49 +0,0 @@
-# Falcon 3.1 3B Instruct using TensorRT-LLM with High Throughput
-
-This directory is [Truss](https://truss.baseten.co/) template for deploying model Falcon 3.1 3B Instruct using our TensorRT-LLM (TRT-LLM) [engine builder](https://docs.baseten.co/performance/engine-builder-overview). This configuration is optimized for high-throughput scenarios.
-
-## Use case
-
-This deployment is tailored for applications that require processing large volumes of data with moderate to low latency, such as:
-* Content moderation for social media platforms
-* Bulk article summarization
-* Large-scale Retrieval-Augmented Generation (RAG) systems
-
-## Configuration
-
-The template uses the following key configuration parameters:
-
-| Property            | Value    | Description                                                                    |
-| ------------------- | -------- | ------------------------------------------------------------------------------ |
-| GPU                 | 1xH100   | Single NVIDIA H100 GPU                                                         |
-| `max_batch_size`    | 32       | Allows processing up to 32 requests simultaneously                             |
-| `quantization_type` | `fp8_kv` | FP8 quantization, balancing performance and accuracy. See this [blog](https://www.baseten.co/blog/33-faster-llm-inference-with-fp8-quantization/). |
-| `max_input_len`     | 4096     | Maximum number of input tokens that the model will accept                      |
-| `max_output_len`    | 1024     | Maximum number of output tokens the model can generate                         |
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd falcon/falcon3-3B-trt-llm-engine-high-throughput
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `falcon/falcon3-3B-trt-llm-engine-high-throughput` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted --publish
-```
-
-Paste your Baseten API key if prompted. Also ensure the `hf_access_token` secret is properly setup in your Baseten Account to access this model.
-
-**Note**: TensorRT-LLM with engine builder will only work under a Baseten production deployment
-
-For more information, refer to the [Truss documentation](https://docs.baseten.co/performance/engine-builder-overview).
diff --git a/falcon/falcon3-3B-trt-llm-engine-high-throughput/config.yaml b/falcon/falcon3-3B-trt-llm-engine-high-throughput/config.yaml
deleted file mode 100644
index 862407cf1..000000000
--- a/falcon/falcon3-3B-trt-llm-engine-high-throughput/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    frequency_penalty: 1
-    max_tokens: 512
-    messages:
-      - content: You are a knowledgable, engaging, biology teacher.
-        role: system
-      - content: What makes falcons effective hunters?
-        role: user
-    stream: true
-    temperature: 0.6
-  repo_id: tiiuae/Falcon3-3B-Instruct
-model_name: Falcon 3 3B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: A10G
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: tiiuae/Falcon3-3B-Instruct
-      source: HF
-    max_seq_len: 8192
-    num_builder_gpus: 1
-    plugin_configuration:
-      paged_kv_cache: true
-      use_paged_context_fmha: true
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    kv_cache_free_gpu_mem_fraction: 0.85
-    request_default_max_tokens: 8192
diff --git a/flux.1-dev-trt-b200/README.md b/flux.1-dev-trt-b200/README.md
deleted file mode 100644
index 80ff0cfbe..000000000
--- a/flux.1-dev-trt-b200/README.md
+++ /dev/null
@@ -1,280 +0,0 @@
-# Flux 1.0 Dev - TensorRT
-
-This model provides high-quality text-to-image generation using [Flux.1-dev model](https://huggingface.co/black-forest-labs/FLUX.1-dev) optimized with TensorRT for the B200 GPU.
-
-## Model Information
-
-- **Model**: Flux 1.0 Dev (black-forest-labs/FLUX.1-dev)
-- **Optimization**: TensorRT 8.6.1
-- **Hardware**: NVIDIA B200 GPU
-- **Framework**: PyTorch with TensorRT acceleration
-
-## Features
-
-- High-quality image generation from text prompts
-- TensorRT optimization for fast inference
-- Support for custom image dimensions (must be multiples of 8)
-- Configurable denoising steps and guidance scale
-- CUDA graph optimization support
-
-## Usage
-
-### Basic Usage
-
-```bash
-curl -X POST https://app.baseten.co/models/{MODEL_ID}/predict \
-  -H "Authorization: Api-Key API_KEY" \
-  -H "Content-Type: application/json" \
-  -d '{
-    "prompt": "A beautiful landscape with mountains and a lake, photorealistic, high quality",
-    "negative_prompt": "blurry, low quality, distorted",
-    "height": 1024,
-    "width": 1024,
-    "num_inference_steps": 30,
-    "guidance_scale": 3.5,
-    "seed": 42,
-    "batch_size": 1,
-    "batch_count": 1
-  }' | python show.py
-```
-
-### Batch Processing
-
-The model supports efficient batch processing for generating multiple images in a single request. You can generate up to 4 images simultaneously with either the same prompt or different prompts for each image.
-
-#### Same Prompt for All Images
-
-```bash
-curl -X POST https://app.baseten.co/models/{MODEL_ID}/predict \
-  -H "Authorization: Api-Key API_KEY" \
-  -H "Content-Type: application/json" \
-  -d '{
-    "prompt": "A beautiful landscape with mountains and a lake, photorealistic, high quality",
-    "negative_prompt": "blurry, low quality, distorted",
-    "height": 1024,
-    "width": 1024,
-    "num_inference_steps": 30,
-    "guidance_scale": 3.5,
-    "seed": 42,
-    "batch_size": 4,
-    "batch_count": 1
-  }' | python show_batch.py
-```
-
-#### Different Prompts for Each Image
-
-```bash
-curl -X POST https://app.baseten.co/models/{MODEL_ID}/predict \
-  -H "Authorization: Api-Key API_KEY" \
-  -H "Content-Type: application/json" \
-  -d '{
-    "prompt": [
-      "A beautiful landscape with mountains and a lake, photorealistic, high quality",
-      "A futuristic city skyline at sunset, neon lights, cyberpunk style, high quality",
-      "A cute cat sitting in a garden, soft lighting, detailed, high quality",
-      "Abstract geometric patterns in vibrant colors, modern art style, high quality"
-    ],
-    "negative_prompt": [
-      "blurry, low quality, distorted",
-      "blurry, low quality, distorted",
-      "blurry, low quality, distorted",
-      "blurry, low quality, distorted"
-    ],
-    "height": 1024,
-    "width": 1024,
-    "num_inference_steps": 30,
-    "guidance_scale": 3.5,
-    "seed": 42,
-    "batch_size": 4,
-    "batch_count": 1
-  }' | python show_batch.py
-```
-
-#### Batch Processing Benefits
-
-- **Parallel Processing**: All images in a batch are generated simultaneously, not sequentially
-- **Better GPU Utilization**: More efficient use of GPU resources compared to separate requests
-- **Faster Total Time**: Generating 4 images in a batch is significantly faster than 4 separate API calls
-- **Consistent Parameters**: All images in a batch use the same dimensions, inference steps, and guidance scale
-
-#### Batch Processing Limitations
-
-- **Maximum Batch Size**: Limited to 4 images per batch (MAX_BATCH_SIZE = 4)
-- **Prompt Array Length**: When using different prompts, the prompt array length must match the batch_size
-- **Memory Requirements**: Larger batches require more GPU memory
-
-#### Displaying Batch Results
-
-Use the included `show_batch.py` script to handle multiple images in the response:
-
-```bash
-# The script automatically detects single vs multiple images
-curl ... | python show_batch.py
-```
-
-The script will:
-- Save each image with a unique filename
-- Automatically open all generated images
-- Print status information about the batch
-
-### Load Test
-Before running the load test, update `load_test.py` with your actual endpoint URL and API key. Replace the placeholder values for `api_url` and `api_key` with your deployment's information (lines 43 and 44).
-
-```bash
-python load_test.py --save-all-images --use-varied-prompts --concurrent --num-requests 30
-
-🚀 Starting Flux Truss API test...
-==================================================
-🎨 Using 30 varied prompts for load testing
-📝 First 3 prompts to be tested:
-   1. a beautiful photograph of Mt. Fuji during cherry blossom, photorealistic, high quality
-   2. a majestic dragon soaring through a mystical forest, digital art, detailed
-   3. a cozy coffee shop interior with warm lighting, people working on laptops, photorealistic
-   ... and 27 more
-🚀 Starting concurrent load test with 30 requests, max 5 workers
-============================================================
-📤 Sending request 1/30: 'a beautiful photograph of Mt. Fuji during cherry b...'
-Testing Truss API endpoint with prompt: 'a beautiful photograph of Mt. Fuji during cherry blossom, photorealistic, high quality'
-
-...
-
-============================================================
-📊 LOAD TEST SUMMARY
-============================================================
-Total requests: 30
-Successful: 30
-Failed: 0
-Success rate: 100.0%
-Total time: 74.72 seconds
-Average request time: 11.63 seconds
-Min request time: 2.91 seconds
-Max request time: 12.86 seconds
-Throughput: 0.40 requests/second
-
-==================================================
-
-💾 Saving 30 successful images...
-✅ Successfully saved 30/30 images to './output'
-📁 Opened output directory
-
-==================================================
-```
-
-### Advanced Usage
-
-```python
-{
-    "prompt": "A beautiful landscape with mountains and a lake, photorealistic, high quality",
-    "prompt2": "Additional prompt for T5 tokenizer",  # Optional, uses prompt if not provided
-    "negative_prompt": "blurry, low quality, distorted",  # Optional
-    "height": 1024,  # Must be multiple of 8
-    "width": 1024,   # Must be multiple of 8
-    "denoising_steps": 50,  # Number of denoising steps
-    "guidance_scale": 3.5,   # Classifier-free guidance scale (must be > 1)
-    "seed": 42,  # Random seed for reproducible results
-    "batch_size": 1,  # Number of images to generate
-    "batch_count": 1,  # Number of batches
-    "num_warmup_runs": 0,  # Number of warmup runs
-    "max_sequence_length": 512,  # Max sequence length (up to 512 for flux.1-dev)
-    "t5_ws_percentage": None,  # T5 weight streaming percentage
-    "transformer_ws_percentage": None  # Transformer weight streaming percentage
-}
-```
-
-## Parameters
-
-| Parameter | Type | Default | Description |
-|-----------|------|---------|-------------|
-| `prompt` | string/array | required | Text prompt(s) for image generation. Can be a single string or array of strings for different prompts per batch item |
-| `prompt2` | string/array | same as prompt | Additional prompt(s) for T5 tokenizer. Can be a single string or array of strings |
-| `negative_prompt` | string/array | "" | Negative prompt(s) to avoid certain elements. Can be a single string or array of strings |
-| `height` | int | 1024 | Image height (must be multiple of 8) |
-| `width` | int | 1024 | Image width (must be multiple of 8) |
-| `denoising_steps` | int | 50 | Number of denoising steps |
-| `guidance_scale` | float | 3.5 | Classifier-free guidance scale (> 1) |
-| `seed` | int | None | Random seed for reproducibility |
-| `batch_size` | int | 1 | Number of images per batch |
-| `batch_count` | int | 1 | Number of batches |
-| `num_warmup_runs` | int | 0 | Number of warmup runs |
-| `max_sequence_length` | int | 512 | Maximum sequence length (≤ 512) |
-| `t5_ws_percentage` | int | None | T5 weight streaming percentage |
-| `transformer_ws_percentage` | int | None | Transformer weight streaming percentage |
-
-## Response Format
-
-The model returns a JSON response with the following structure:
-
-### Single Image Response
-
-```json
-{
-    "status": "success",
-    "data": "base64_encoded_image",
-    "time": 2.34,
-    "prompt": "A beautiful landscape with mountains and a lake, photorealistic, high quality",
-    "negative_prompt": "blurry, low quality, distorted",
-    "height": 1024,
-    "width": 1024,
-    "num_inference_steps": 30,
-    "guidance_scale": 3.5,
-    "seed": 42
-}
-```
-
-### Batch Response (Multiple Images)
-
-```json
-{
-    "status": "success",
-    "data": [
-        "base64_encoded_image_1",
-        "base64_encoded_image_2",
-        "base64_encoded_image_3",
-        "base64_encoded_image_4"
-    ],
-    "time": 7.23,
-    "prompt": [
-        "A beautiful landscape with mountains and a lake, photorealistic, high quality",
-        "A futuristic city skyline at sunset, neon lights, cyberpunk style, high quality",
-        "A cute cat sitting in a garden, soft lighting, detailed, high quality",
-        "Abstract geometric patterns in vibrant colors, modern art style, high quality"
-    ],
-    "negative_prompt": [
-        "blurry, low quality, distorted",
-        "blurry, low quality, distorted",
-        "blurry, low quality, distorted",
-        "blurry, low quality, distorted"
-    ],
-    "height": 1024,
-    "width": 1024,
-    "num_inference_steps": 30,
-    "guidance_scale": 3.5,
-    "seed": 42
-}
-```
-
-## Performance Notes
-
-The model is optimized with a pre-compiled TensorRT engine for the NVIDIA B200 GPU.
-Performance characteristic is described below for the [basic usage](#basic-usage)
-
-
-```text
-|------------------|--------------|
-|     Module       |   Latency    |
-|------------------|--------------|
-|      CLIP        |      2.02 ms |
-|       T5         |      6.43 ms |
-| Transformer x 50 |   2361.44 ms |
-|     VAE-Dec      |     11.67 ms |
-|------------------|--------------|
-|    Pipeline      |   2382.45 ms |
-|------------------|--------------|
-```
-
-## Model Variants
-
-This implementation supports the `flux.1-dev` variant with:
-- Maximum sequence length: 512 tokens
-- Default image dimensions: 1024x1024
-- Optimized for high-quality image generation
diff --git a/flux.1-dev-trt-b200/config.yaml b/flux.1-dev-trt-b200/config.yaml
deleted file mode 100644
index d2b6b2e2f..000000000
--- a/flux.1-dev-trt-b200/config.yaml
+++ /dev/null
@@ -1,17 +0,0 @@
-base_image:
-  image: nvcr.io/nvidia/pytorch:25.06-py3
-description: Generate high-quality images from text prompts using Black Forest Labs's Flux model with TensorRT optimization.
-external_package_dirs: []
-model_name: Flux 1.0 Dev - TensorRT
-requirements_file: ./requirements.txt
-resources:
-  accelerator: B200
-  use_gpu: true
-runtime:
-  predict_concurrency: 1
-secrets:
-  hf_access_token: null
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/flux/README.md b/flux/README.md
deleted file mode 100644
index 3f7fa0911..000000000
--- a/flux/README.md
+++ /dev/null
@@ -1,82 +0,0 @@
-# FLUX.1
-
-These are trusses for the brand new FLUX.1 schnell and dev model.
-
-## Deploy FLUX.1
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd flux
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Make sure you have access to the [FLUX.1 model](https://huggingface.co/black-forest-labs/FLUX.1-schnell) and add a huggingface access token with read access as a Baseten secret named `hf_access_token`.
-
-With `flux/dev` or `flux/schnell` as your working directory, you can deploy the model with:
-
-```
-truss push --trusted --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using FLUX.1 through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-## Invoking FLUX.1
-
-The output will be a dictionary with a key `data` mapping to a base64 encoded image. It's processed with this script:
-
-```python
-import httpx
-import os
-import base64
-from PIL import Image
-from io import BytesIO
-
-# Replace the empty string with your model id below
-model_id = ""
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-
-# Function used to convert a base64 string to a PIL image
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str)))
-
-data = {
-  "prompt": "a little boy looking through a large magical portal, the boy sees a futuristic human civilization in that portal, extremely detailed, trending on artstation, 8k"
-}
-
-# Call model endpoint
-res = httpx.post(
-    f"https://model-{model_id}.api.baseten.co/production/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=data
-)
-
-# Get output image
-res = res.json()
-output = res.get("data")
-
-# Convert the base64 model output to an image
-img = b64_to_pil(output)
-img.save("output_image.jpg")
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://model-{model_id}.api.baseten.co/production/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "A tree in a field under the night sky"
-         }'
-```
-
-Again, the model will return a dictionary containing the base64-encoded image, which will need to be decoded and saved.
diff --git a/flux/dev/config.yaml b/flux/dev/config.yaml
deleted file mode 100644
index 61e1a7d40..000000000
--- a/flux/dev/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-external_package_dirs: []
-model_metadata:
-  example_model_input: {"prompt": 'black forest gateau cake spelling out the words "FLUX DEV", tasty, food photography, dynamic shot'}
-  repo_id: black-forest-labs/FLUX.1-dev
-model_name: Flux.1-dev
-python_version: py311
-requirements:
-  - git+https://github.com/huggingface/diffusers.git@fc6a91e3834c35e57b398ad1c0d99f6f83557e04
-  - transformers
-  - accelerate
-  - sentencepiece
-  - protobuf
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages:
-  - ffmpeg
-  - libsm6
-  - libxext6
diff --git a/flux/schnell/README.md b/flux/schnell/README.md
deleted file mode 100644
index c4f768de4..000000000
--- a/flux/schnell/README.md
+++ /dev/null
@@ -1,81 +0,0 @@
-# b10‑transfer + torch.compile example (Flux image generation)
-
-Speed up cold starts for a diffusers Flux pipeline on Baseten by caching PyTorch compilation artifacts with b10‑tcache. This example shows how to:
-
-- Compile the heavy parts of the pipeline with torch.compile
-
-- Warm up the model across supported resolutions
-
-- Load a previously saved compile cache on startup and save it after first warmup
-
-In this case, we brought the cold start time from 900s to 70s.
-
-See the full docs on [b10_transfer](https://docs.baseten.co/development/model/b10-transfer).
-
-## Prerequisites
-
-Add the cache helper to your requirements in `config.yaml`:
-
-```yaml
-requirements:
-  - b10-transfer
-```
-
-## How it works
-
-Load (and later save) compile cache via b10_transfer:
-
-```python
-from b10_transfer import load_compile_cache, save_compile_cache, OperationStatus
-
-cache_loaded = load_compile_cache()
-
-if cache_loaded == OperationStatus.ERROR:
-    logging.info("Run in eager mode, skipping torch compile")
-else:
-    self.compile()
-
-if cache_loaded == OperationStatus.DOES_NOT_EXIST:
-    save_compile_cache()
-```
-
-Compile the hotspots of the pipeline with Torch Dynamo/TorchInductor:
-
-```python
-self.pipe.transformer = torch.compile(
-    self.pipe.transformer, mode="max-autotune-no-cudagraphs", dynamic=False
-)
-self.pipe.vae.decode = torch.compile(
-    self.pipe.vae.decode, mode="max-autotune-no-cudagraphs", dynamic=False
-)
-```
-
-Warm up with dummy prompts across every resolution you intend to serve (so later requests hit already‑compiled kernels):
-
-```python
-for width, height in [(1024, 1024), (1216, 832), (896, 1152)]:
-    self.pipe(
-        prompt="dummy prompt",
-        prompt_2=None,
-        guidance_scale=0.0,
-        max_sequence_length=256,
-        num_inference_steps=4,
-        width=width,
-        height=height,
-        output_type="pil",
-        generator=generator,
-    )
-    self.pipe(
-        prompt="extra dummy prompt",
-        prompt_2=None,
-        guidance_scale=0.0,
-        max_sequence_length=256,
-        num_inference_steps=4,
-        width=width,
-        height=height,
-        output_type="pil",
-        generator=generator,
-    )
-```
-
-On subsequent cold starts, load_compile_cache() restores previously compiled artifacts and dramatically reduces compile latency.
diff --git a/flux/schnell/config.yaml b/flux/schnell/config.yaml
deleted file mode 100644
index 0e19a616f..000000000
--- a/flux/schnell/config.yaml
+++ /dev/null
@@ -1,22 +0,0 @@
-external_package_dirs: []
-model_metadata:
-  example_model_input: {"prompt": 'black forest gateau cake spelling out the words "FLUX SCHNELL", tasty, food photography, dynamic shot'}
-  repo_id: black-forest-labs/FLUX.1-schnell
-model_name: Flux.1-schnell
-python_version: py311
-requirements:
-  - git+https://github.com/huggingface/diffusers.git@fc6a91e3834c35e57b398ad1c0d99f6f83557e04
-  - transformers
-  - accelerate
-  - sentencepiece
-  - protobuf
-  - b10-transfer
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages:
-  - ffmpeg
-  - libsm6
-  - libxext6
diff --git a/fotographer/zenctrl/README.md b/fotographer/zenctrl/README.md
deleted file mode 100644
index 27bac2687..000000000
--- a/fotographer/zenctrl/README.md
+++ /dev/null
@@ -1,17 +0,0 @@
-![Header Image](images/banner_1.png)
-
-# Fotographer AI ZenCtrl
-
-Image-to-image model for generating in-context product photography.
-
-Deploy with `truss push --promote`
-
-Call with `python call.py` after providing `model_id` from deployed model.
-
-### Example input image
-
-![Speaker Input](images/speaker-input.png)
-
-### Example output image
-
-![Speaker Input](images/speaker-output.png)
diff --git a/fotographer/zenctrl/config.yaml b/fotographer/zenctrl/config.yaml
deleted file mode 100644
index f199b2185..000000000
--- a/fotographer/zenctrl/config.yaml
+++ /dev/null
@@ -1,20 +0,0 @@
-base_image:
-  image: fotographerai/zenctrlstage:latest
-model_metadata: {}
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python -m uvicorn app:app --host 0.0.0.0 --port 8000 --log-level debug"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /generate
-  server_port: 8000
-resources:
-  accelerator: H100
-  use_gpu: true
-model_name: ZenCtrl
-environment_variables:
-  PORT: 8000
-  HF_TOKEN: null
-runtime:
-  predict_concurrency: 8
-secrets:
-  hf_access_token: null
diff --git a/gemma/gemma-2-27b-it-vllm/README.md b/gemma/gemma-2-27b-it-vllm/README.md
deleted file mode 100644
index 093531597..000000000
--- a/gemma/gemma-2-27b-it-vllm/README.md
+++ /dev/null
@@ -1,55 +0,0 @@
-# Gemma 2 27B
-
-This is a [Truss](https://truss.baseten.co/) for Gemma 2 27B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Gemma 2 27B Instruct.
-
-## Gemma 2 27B Instruct Implementation
-
-This implementation of Gemma 2 uses [vLLM](https://github.com/vllm-project/vllm).
-
-Since Gemma 2 is a gated model, you will also need to provide your Huggingface access token after making sure you have access to [the model](https://huggingface.co/google/gemma-2-27b-it). Please use the [following guide](https://docs.baseten.co/deploy/guides/secrets) to add your Huggingface access token as a secret.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd gemma/gemma-2-27b-it-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `gemma/gemma-2-27b-it-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Gemma 2 27B Instruct API documentation
-
-This section provides an overview of the Gemma 2 27B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __max_tokens__: The maximum number of output tokens.
-
-## Example usage
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{"prompt": "what came before, the chicken or the egg?", "max_tokens": 64}'
-```
diff --git a/gemma/gemma-2-27b-it-vllm/config.yaml b/gemma/gemma-2-27b-it-vllm/config.yaml
deleted file mode 100644
index 26593477f..000000000
--- a/gemma/gemma-2-27b-it-vllm/config.yaml
+++ /dev/null
@@ -1,17 +0,0 @@
-model_name: "Gemma 2 27B Instruct VLLM"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: google/gemma-2-27b-it
-  tensor_parallel: 1
-  max_num_seqs: 16
-requirements:
-  - vllm==0.5.1
-  - https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp311-cp311-linux_x86_64.whl
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
diff --git a/gemma/gemma-2-9b-it-vllm/README.md b/gemma/gemma-2-9b-it-vllm/README.md
deleted file mode 100644
index b660f4896..000000000
--- a/gemma/gemma-2-9b-it-vllm/README.md
+++ /dev/null
@@ -1,55 +0,0 @@
-# Gemma 2 9B
-
-This is a [Truss](https://truss.baseten.co/) for Gemma 2 9B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Gemma 2 9B Instruct.
-
-## Gemma 2 9B Instruct Implementation
-
-This implementation of Gemma 2 uses [vLLM](https://github.com/vllm-project/vllm).
-
-Since Gemma 2 is a gated model, you will also need to provide your Huggingface access token after making sure you have access to [the model](https://huggingface.co/google/gemma-2-9b-it). Please use the [following guide](https://docs.baseten.co/deploy/guides/secrets) to add your Huggingface access token as a secret.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd gemma/gemma-2-9b-it-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `gemma/gemma-2-9b-it-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Gemma 2 9B Instruct API documentation
-
-This section provides an overview of the Gemma 2 9B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __max_tokens__: The maximum number of output tokens.
-
-## Example usage
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{"prompt": "what came before, the chicken or the egg?", "max_tokens": 64}'
-```
diff --git a/gemma/gemma-2-9b-it-vllm/config.yaml b/gemma/gemma-2-9b-it-vllm/config.yaml
deleted file mode 100644
index d6b1306a5..000000000
--- a/gemma/gemma-2-9b-it-vllm/config.yaml
+++ /dev/null
@@ -1,16 +0,0 @@
-model_name: "Gemma 2 9B Instruct VLLM"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: google/gemma-2-9b-it
-  tensor_parallel: 1
-requirements:
-  - vllm==0.5.1
-  - https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp311-cp311-linux_x86_64.whl
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
diff --git a/gemma/gemma-3-27b-it/config.yaml b/gemma/gemma-3-27b-it/config.yaml
deleted file mode 100644
index 9c28b9aee..000000000
--- a/gemma/gemma-3-27b-it/config.yaml
+++ /dev/null
@@ -1,61 +0,0 @@
-base_image:
-  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:8a4a2efc6fc32cdc30e4e35ba3f8c64dcd0aa1d0
-build_commands:
-  - pip install git+https://github.com/huggingface/transformers@071a161d3e38f56dbda2743b979f0afeed2cd4f1
-model_metadata:
-  repo_id: google/gemma-3-27b-it
-  example_model_input: {
-    "model": "gemma",
-    "messages": [
-      {
-      "role": "user",
-      "content": [
-        {
-        "type": "text",
-        "text": "Describe this image in one sentence."
-        },
-        {
-        "type": "image_url",
-        "image_url": {
-          "url": "https://picsum.photos/id/237/200/300"
-        }
-        }
-      ]
-      }
-    ],
-    "stream": true,
-    "max_tokens": 512,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: "sh -c \"truss-transfer-cli && VLLM_USE_V1=1 HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve /app/model_cache/gemma --served-model-name gemma --max-num-seqs 8 --max-model-len 16384 --limit_mm_per_prompt 'image=1' --hf-overrides '{\\\"do_pan_and_scan\\\": true}' --gpu-memory-utilization 0.95\""
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
-model_cache:
-  - repo_id: google/gemma-3-27b-it
-    revision: 005ad3404e59d6023443cb575daa05336842228a
-    use_volume: true
-    volume_folder: gemma
-requirements:
-- huggingface_hub
-- hf_transfer
-- datasets
-resources:
-  accelerator: H100
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  health_checks:
-    restart_check_delay_seconds: 300      # Waits 5 minutes after deployment before starting health checks
-    restart_threshold_seconds: 300       # Triggers a restart if health checks fail for 5 minutes
-    stop_traffic_threshold_seconds: 120  # Stops traffic if health checks fail for 2 minutes
-  predict_concurrency : 8
-  truss_server_version_override: "0.11.4"
-model_name: Gemma 27B Instruct
diff --git a/gfp-gan/README.md b/gfp-gan/README.md
deleted file mode 100644
index 611703192..000000000
--- a/gfp-gan/README.md
+++ /dev/null
@@ -1,62 +0,0 @@
-# GFP-GAN Truss
-
-This is a [Truss](https://truss.baseten.co/) for serving an implementation of TencentARC
-[GFPGAN](https://github.com/TencentARC/GFPGAN). GFPGAN is an algorithm for real-world face restoration.
-It can be used on old photos of faces to remove blur, and increase clarity and resolution.
-
-It leverages rich and diverse priors encapsulated in a pretrained face GAN (e.g., StyleGAN2) for
-"blind face" restoration.
-
-## Deploying GFP-GAN
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd gfp-gan-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `gfp-gan-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## GFP-GAN API documentation
-
-### Input
-
-The input should be a dictionary with the following key:
-
-- `image` - the image to be restored, encoded as base64.
-
-### Output
-
-The model returns a dictionary containing the base64-encoded restored image:
-
-- `status` - either `success` or `failed`
-- `data` - the restored image, encoded as base64
-- `message` - will contain details in the case of errors
-
-## Example usage
-
-```sh
-truss predict -d '{"image": "{BASE_64_INPUT}"}'
-```
-
-You can also invoke this model on Baseten with the following cURL command (just fill in the model version ID and API Key):
-
-```
-$ curl -X POST https://app.baseten.co/models/{MODEL_VERSION_ID}/predict \
-    -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-    -d '{"image": "{BASE_64_INPUT}"}'
-```
diff --git a/gfp-gan/config.yaml b/gfp-gan/config.yaml
deleted file mode 100644
index e047715ee..000000000
--- a/gfp-gan/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-description: Restore photos with this image-to-image model.
-environment_variables: {}
-external_data:
-- local_data_path: RealESRGAN_x2plus.pth
-  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/gfp-gan/RealESRGAN_x2plus.pth
-- local_data_path: GFPGANv1.3.pth
-  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/gfp-gan/GFPGANv1.3.pth
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/tencent.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/gfp-gan.png
-  example_model_input_file: input.json
-  tags:
-  - image-restoration
-model_name: GFP-GAN
-python_version: py39
-requirements:
-- gfpgan==1.3.8
-- realesrgan==0.3.0
-- basicsr==1.4.2
-- torchvision==0.16.2
-- numpy==1.26.4
-resources:
-  cpu: '3'
-  memory: 8Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/grpc/README.md b/grpc/README.md
deleted file mode 100644
index 90df3f483..000000000
--- a/grpc/README.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# gRPC on Baseten
-
-This example demonstrates how to deploy a gRPC model on Baseten using Truss.
-
-# Prerequisites
-
-1. **Install Truss:**
-   ```bash
-   pip install --upgrade truss
-   ```
-
-2. **Install Protocol Buffer compiler:**
-   ```bash
-   # On macOS
-   brew install protobuf
-
-   # On Ubuntu/Debian
-   sudo apt-get install protobuf-compiler
-
-   # On other systems, see: https://protobuf.dev/getting-started/
-   ```
-
-3. **Install gRPC tools:**
-   ```bash
-   pip install grpcio-tools
-   ```
-
-# Steps to Deploy
-
-### Step 1: Generate Protocol Buffer Code
-
-Generate the Python code from your `.proto` file:
-
-```bash
-python -m grpc_tools.protoc --python_out=. --grpc_python_out=. --proto_path . example.proto
-```
-
-### Step 2: Build and Push Docker Image
-
-Build and push your Docker image to a container registry:
-
-```bash
-docker build -t your-registry/truss-grpc-demo:latest . --platform linux/amd64
-docker push your-registry/truss-grpc-demo:latest
-```
-
-### Step 3: Configure your Truss
-
-Update the `config.yaml` file with your model name and Docker image:
-
-```yaml
-model_name: "gRPC Model Example"
-base_image:
-    image: your-registry/truss-grpc-demo:latest
-```
-
-### Step 4: Deploy with Truss
-
-Deploy your model using the Truss CLI:
-
-```bash
-truss push --promote
-```
-
-### Step 5: Invoke the Model
-
-Update the code in `client.py` to connect to your deployed model. Replace `{MODEL_ID}` with your actual model ID,
-and the API_KEY with your Baseten API key:
-
-Run your client to test the deployed model:
-
-```bash
-python client.py
-```
diff --git a/grpc/config.yaml b/grpc/config.yaml
deleted file mode 100644
index b8802073a..000000000
--- a/grpc/config.yaml
+++ /dev/null
@@ -1,15 +0,0 @@
-model_name: "gRPC Model Example"
-base_image:
-  image: your/repository:tag
-docker_server:
-  start_command: python model.py
-  server_port: 8080
-  predict_endpoint: /
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-resources:
-  accelerator: A10G  # or your preferred GPU
-  use_gpu: true
-runtime:
-  transport:
-    kind: "grpc"
diff --git a/image-segmentation/config.yaml b/image-segmentation/config.yaml
deleted file mode 100644
index 6602b1035..000000000
--- a/image-segmentation/config.yaml
+++ /dev/null
@@ -1,13 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: Image segmentation
-python_version: py39
-requirements:
-- torchvision==0.9.1
-resources:
-  cpu: 3000m
-  memory: 8Gi
-  use_gpu: false
-secrets: {}
-spec_version: 2.0
-system_packages: []
diff --git a/image/README.md b/image/README.md
new file mode 100644
index 000000000..d43d01bf6
--- /dev/null
+++ b/image/README.md
@@ -0,0 +1,29 @@
+# Image Models
+
+Truss configurations for image generation, editing, and segmentation models. Covers Stable Diffusion, Flux, and a variety of specialized image processing pipelines.
+
+| Directory                                           | Variants | Description                                                                                                                   |
+| --------------------------------------------------- | -------- | ----------------------------------------------------------------------------------------------------------------------------- |
+| [stable-diffusion](stable-diffusion/)               | 18       | Stable Diffusion family including SD 1.x, SDXL, SD 3, turbo, LCM, LoRA, ControlNet, inpainting, TensorRT, and video diffusion |
+| [flux](flux/)                                       | 2        | Black Forest Labs Flux models (dev, schnell)                                                                                  |
+| [flux-dev-trt-b200](flux-dev-trt-b200/)             | 1        | Flux Dev optimized with TensorRT for B200 GPUs                                                                                |
+| [sana](sana/)                                       | 2        | Sana image generation models (600M, 1600M)                                                                                    |
+| [comfyui](comfyui/)                                 | 1        | ComfyUI workflow server for node-based image generation                                                                       |
+| [control-net-qrcode](control-net-qrcode/)           | 1        | ControlNet QR code art generation                                                                                             |
+| [deepfloyd-xl](deepfloyd-xl/)                       | 1        | DeepFloyd IF XL text-to-image model                                                                                           |
+| [fotographer](fotographer/)                         | 1        | Fotographer AI portrait generation model                                                                                      |
+| [gfp-gan](gfp-gan/)                                 | 1        | GFPGAN face restoration and enhancement                                                                                       |
+| [ip-adapter](ip-adapter/)                           | 1        | IP-Adapter for image-prompted generation                                                                                      |
+| [magic-animate](magic-animate/)                     | 1        | MagicAnimate human image animation                                                                                            |
+| [playground-v2-aesthetic](playground-v2-aesthetic/) | 1        | Playground v2 aesthetic image generation                                                                                      |
+| [segment-anything](segment-anything/)               | 1        | Meta SAM universal image segmentation                                                                                         |
+| [dis-segmentation](dis-segmentation/)               | 1        | Dichotomous image segmentation for high-accuracy cutouts                                                                      |
+| [image-segmentation](image-segmentation/)           | 1        | General-purpose image segmentation                                                                                            |
+
+## Deploying
+
+Each image model can be deployed to Baseten with:
+
+```bash
+truss push
+```
diff --git a/image/comfyui/README.md b/image/comfyui/README.md
new file mode 100644
index 000000000..d957e93e2
--- /dev/null
+++ b/image/comfyui/README.md
@@ -0,0 +1,36 @@
+# ComfyUI Workflow
+
+Deploy a ComfyUI workflow as a Truss
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "workflow_values": {
+    "controlnet_image": "https://storage.googleapis.com/logos-bucket-01/baseten_logo.png",
+    "negative_prompt": "blurry, text, low quality",
+    "positive_prompt": "An igloo on a snowy day, 4k, hd"
+  }
+}'
+```
+
+## Configuration highlights
+
+- Base image: `bolabaseten/comfyui-truss-base:6a7bc35`
+- System packages: `ffmpeg, libgl1-mesa-glx`
diff --git a/image/comfyui/config.yaml b/image/comfyui/config.yaml
new file mode 100644
index 000000000..f24e7d599
--- /dev/null
+++ b/image/comfyui/config.yaml
@@ -0,0 +1,28 @@
+base_image:
+  image: bolabaseten/comfyui-truss-base:6a7bc35
+  python_executable_path: /usr/bin/python3
+description: Deploy a ComfyUI workflow as a Truss
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-diffusion-xl-base-1.0"
+  example_model_input:
+    workflow_values:
+      controlnet_image: https://storage.googleapis.com/logos-bucket-01/baseten_logo.png
+      negative_prompt: blurry, text, low quality
+      positive_prompt: An igloo on a snowy day, 4k, hd
+model_name: ComfyUI Workflow
+python_version: py39
+requirements:
+- websocket-client==1.6.4
+- accelerate==0.23.0
+- opencv-python==4.8.1.78
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
+- libgl1-mesa-glx
diff --git a/comfyui-truss/data/comfy_ui_workflow.json b/image/comfyui/data/comfy_ui_workflow.json
similarity index 100%
rename from comfyui-truss/data/comfy_ui_workflow.json
rename to image/comfyui/data/comfy_ui_workflow.json
diff --git a/comfyui-truss/data/model.json b/image/comfyui/data/model.json
similarity index 100%
rename from comfyui-truss/data/model.json
rename to image/comfyui/data/model.json
diff --git a/comfyui-truss/examples/animate-diff/model.json b/image/comfyui/examples/animate-diff/model.json
similarity index 100%
rename from comfyui-truss/examples/animate-diff/model.json
rename to image/comfyui/examples/animate-diff/model.json
diff --git a/comfyui-truss/examples/animate-diff/workflow.json b/image/comfyui/examples/animate-diff/workflow.json
similarity index 100%
rename from comfyui-truss/examples/animate-diff/workflow.json
rename to image/comfyui/examples/animate-diff/workflow.json
diff --git a/image/comfyui/examples/anime-style-transfer/README.md b/image/comfyui/examples/anime-style-transfer/README.md
new file mode 100644
index 000000000..8493cc762
--- /dev/null
+++ b/image/comfyui/examples/anime-style-transfer/README.md
@@ -0,0 +1,31 @@
+# ComfyUI Anime Pet Style Transfer
+
+Deploy ComfyUI Anime Pet Style Transfer for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- System packages: `wget, ffmpeg, libgl1-mesa-glx`
diff --git a/image/comfyui/examples/anime-style-transfer/config.yaml b/image/comfyui/examples/anime-style-transfer/config.yaml
new file mode 100644
index 000000000..abfc3c94e
--- /dev/null
+++ b/image/comfyui/examples/anime-style-transfer/config.yaml
@@ -0,0 +1,35 @@
+description: "ComfyUI Anime Pet Style Transfer for image generation"
+build_commands:
+- git clone https://github.com/comfyanonymous/ComfyUI.git
+- cd ComfyUI && git checkout b1fd26fe9e55163f780bf9e5f56bf9bf5f035c93 && pip install -r requirements.txt
+- cd ComfyUI/custom_nodes && git clone https://github.com/LykosAI/ComfyUI-Inference-Core-Nodes --recursive && cd ComfyUI-Inference-Core-Nodes && pip install -e .[cuda12]
+- cd ComfyUI/custom_nodes && git clone https://github.com/ZHO-ZHO-ZHO/ComfyUI-Gemini --recursive && cd ComfyUI-Gemini && pip install -r requirements.txt
+- cd ComfyUI/custom_nodes && git clone https://github.com/kijai/ComfyUI-Marigold --recursive && cd ComfyUI-Marigold && pip install -r requirements.txt
+- cd ComfyUI/custom_nodes && git clone https://github.com/omar92/ComfyUI-QualityOfLifeSuit_Omar92 --recursive
+- cd ComfyUI/custom_nodes && git clone https://github.com/Fannovel16/comfyui_controlnet_aux --recursive && cd comfyui_controlnet_aux && pip install -r requirements.txt
+- cd ComfyUI/models/controlnet && wget -O control-lora-canny-rank256.safetensors https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-canny-rank256.safetensors
+- cd ComfyUI/models/controlnet && wget -O control-lora-depth-rank256.safetensors https://huggingface.co/stabilityai/control-lora/resolve/main/control-LoRAs-rank256/control-lora-depth-rank256.safetensors
+- cd ComfyUI/models/checkpoints && wget -O dreamshaperXL_v21TurboDPMSDE.safetensors https://civitai.com/api/download/models/351306
+- cd ComfyUI/models/loras && wget -O StudioGhibli.Redmond-StdGBRRedmAF-StudioGhibli.safetensors https://huggingface.co/artificialguybr/StudioGhibli.Redmond-V2/resolve/main/StudioGhibli.Redmond-StdGBRRedmAF-StudioGhibli.safetensors
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-diffusion-xl-base-1.0"
+  example_model_input:
+    workflow_values:
+      prompt: A cute cat sitting on a windowsill
+      input_image: https://example.com/image.jpg
+model_name: ComfyUI Anime Pet Style Transfer
+python_version: py310
+requirements:
+  - websocket-client==1.6.4
+  - accelerate==0.23.0
+  - opencv-python==4.8.1.78
+resources:
+  accelerator: A100
+  use_gpu: true
+secrets: {}
+system_packages:
+  - wget
+  - ffmpeg
+  - libgl1-mesa-glx
diff --git a/comfyui-truss/examples/anime-style-transfer/workflow.json b/image/comfyui/examples/anime-style-transfer/workflow.json
similarity index 100%
rename from comfyui-truss/examples/anime-style-transfer/workflow.json
rename to image/comfyui/examples/anime-style-transfer/workflow.json
diff --git a/comfyui-truss/examples/sdxl-controlnet/model.json b/image/comfyui/examples/sdxl-controlnet/model.json
similarity index 100%
rename from comfyui-truss/examples/sdxl-controlnet/model.json
rename to image/comfyui/examples/sdxl-controlnet/model.json
diff --git a/comfyui-truss/examples/sdxl-controlnet/workflow.json b/image/comfyui/examples/sdxl-controlnet/workflow.json
similarity index 100%
rename from comfyui-truss/examples/sdxl-controlnet/workflow.json
rename to image/comfyui/examples/sdxl-controlnet/workflow.json
diff --git a/comfyui-truss/examples/sdxl-with-refiner/model.json b/image/comfyui/examples/sdxl-with-refiner/model.json
similarity index 100%
rename from comfyui-truss/examples/sdxl-with-refiner/model.json
rename to image/comfyui/examples/sdxl-with-refiner/model.json
diff --git a/comfyui-truss/examples/sdxl-with-refiner/workflow.json b/image/comfyui/examples/sdxl-with-refiner/workflow.json
similarity index 100%
rename from comfyui-truss/examples/sdxl-with-refiner/workflow.json
rename to image/comfyui/examples/sdxl-with-refiner/workflow.json
diff --git a/control-net-qrcode/model/__init__.py b/image/comfyui/model/__init__.py
similarity index 100%
rename from control-net-qrcode/model/__init__.py
rename to image/comfyui/model/__init__.py
diff --git a/comfyui-truss/model/helpers.py b/image/comfyui/model/helpers.py
similarity index 100%
rename from comfyui-truss/model/helpers.py
rename to image/comfyui/model/helpers.py
diff --git a/comfyui-truss/model/model.py b/image/comfyui/model/model.py
similarity index 100%
rename from comfyui-truss/model/model.py
rename to image/comfyui/model/model.py
diff --git a/image/control-net-qrcode/README.md b/image/control-net-qrcode/README.md
new file mode 100644
index 000000000..b60bc1d15
--- /dev/null
+++ b/image/control-net-qrcode/README.md
@@ -0,0 +1,32 @@
+# control-net-qrcode
+
+Deploy control-net-qrcode for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "A cubism painting of the Garden of Eaden with animals walking around, Andreas Rocha, matte painting concept art, a detailed matte painting",
+  "qr_code_content": "https://www.baseten.co"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/control-net-qrcode/config.yaml b/image/control-net-qrcode/config.yaml
new file mode 100644
index 000000000..5e8a55731
--- /dev/null
+++ b/image/control-net-qrcode/config.yaml
@@ -0,0 +1,27 @@
+description: "ControlNet QR Code for artistic QR code image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "monster-labs/control_v1p_sd15_qrcode_monster"
+  example_model_input:
+    prompt: A cubism painting of the Garden of Eaden with animals walking around,
+      Andreas Rocha, matte painting concept art, a detailed matte painting
+    qr_code_content: https://www.baseten.co
+model_name: control-net-qrcode
+python_version: py310
+requirements:
+- diffusers==0.21.1
+- torch==2.0.1
+- ftfy==6.1.1
+- scipy==1.9.3
+- transformers==4.25.1
+- accelerate==0.20.3
+- qrcode==7.4.2
+- xformers==0.0.21
+resources:
+  accelerator: T4
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/control-net-qrcode/controlnet_qr_code_results.gif b/image/control-net-qrcode/controlnet_qr_code_results.gif
similarity index 100%
rename from control-net-qrcode/controlnet_qr_code_results.gif
rename to image/control-net-qrcode/controlnet_qr_code_results.gif
diff --git a/deepfloyd-xl/model/__init__.py b/image/control-net-qrcode/model/__init__.py
similarity index 100%
rename from deepfloyd-xl/model/__init__.py
rename to image/control-net-qrcode/model/__init__.py
diff --git a/control-net-qrcode/model/model.py b/image/control-net-qrcode/model/model.py
similarity index 100%
rename from control-net-qrcode/model/model.py
rename to image/control-net-qrcode/model/model.py
diff --git a/control-net-qrcode/twitter_mask.jpeg b/image/control-net-qrcode/twitter_mask.jpeg
similarity index 100%
rename from control-net-qrcode/twitter_mask.jpeg
rename to image/control-net-qrcode/twitter_mask.jpeg
diff --git a/control-net-qrcode/twitter_output.jpg b/image/control-net-qrcode/twitter_output.jpg
similarity index 100%
rename from control-net-qrcode/twitter_output.jpg
rename to image/control-net-qrcode/twitter_output.jpg
diff --git a/image/deepfloyd-xl/README.md b/image/deepfloyd-xl/README.md
new file mode 100644
index 000000000..7ed98c40a
--- /dev/null
+++ b/image/deepfloyd-xl/README.md
@@ -0,0 +1,33 @@
+# Deepfloyd XL
+
+Generate original images from text prompts.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/deepfloyd-xl/config.yaml b/image/deepfloyd-xl/config.yaml
new file mode 100644
index 000000000..d9c410018
--- /dev/null
+++ b/image/deepfloyd-xl/config.yaml
@@ -0,0 +1,32 @@
+description: Generate original images from text prompts.
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "DeepFloyd/IF-I-XL-v1.0"
+  avatar_url: https://cdn.baseten.co/production/static/explore/deep-floyd.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/deepfloyd-cover.png
+  example_model_input:
+    prompt: A photo of an astronaut riding a horse
+  tags:
+  - image-generation
+model_name: Deepfloyd XL
+python_version: py39
+requirements:
+- diffusers==0.24.0
+- transformers==4.36.0
+- torch==2.1.0
+- scipy==1.11.4
+- accelerate==0.25.0
+- pillow==10.1.0
+- bitsandbytes==0.41.3
+- sentencepiece==0.1.99
+- huggingface_hub==0.19.4
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets:
+  hf_access_token: ENTER HF API KEY HERE
+spec_version: "2.0"
+system_packages: []
diff --git a/deepspeed-mii/model/__init__.py b/image/deepfloyd-xl/model/__init__.py
similarity index 100%
rename from deepspeed-mii/model/__init__.py
rename to image/deepfloyd-xl/model/__init__.py
diff --git a/deepfloyd-xl/model/model.py b/image/deepfloyd-xl/model/model.py
similarity index 100%
rename from deepfloyd-xl/model/model.py
rename to image/deepfloyd-xl/model/model.py
diff --git a/image/dis-segmentation/README.md b/image/dis-segmentation/README.md
new file mode 100644
index 000000000..362b2c20c
--- /dev/null
+++ b/image/dis-segmentation/README.md
@@ -0,0 +1,31 @@
+# DIS Segmentation
+
+Deploy DIS Segmentation for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "input_image": "<base64 image string>"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/dis-segmentation/config.yaml b/image/dis-segmentation/config.yaml
new file mode 100644
index 000000000..549b0a66d
--- /dev/null
+++ b/image/dis-segmentation/config.yaml
@@ -0,0 +1,23 @@
+description: "DIS for high-accuracy image segmentation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "ECCV2022/dis-background-removal"
+  example_model_input:
+    input_image: <base64 image string>
+model_name: DIS Segmentation
+python_version: py310
+requirements:
+- torch==2.1.0
+- Pillow==9.4.0
+- numpy==1.23.5
+- gdown==4.7.3
+- torchvision==0.16.0
+- torchaudio==2.1.0
+- scikit-image==0.19.3
+resources:
+  accelerator: T4
+  memory: 2Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/dis-segmentation/model/__init__.py b/image/dis-segmentation/model/__init__.py
similarity index 100%
rename from dis-segmentation/model/__init__.py
rename to image/dis-segmentation/model/__init__.py
diff --git a/dis-segmentation/model/clone_repo_helper.py b/image/dis-segmentation/model/clone_repo_helper.py
similarity index 100%
rename from dis-segmentation/model/clone_repo_helper.py
rename to image/dis-segmentation/model/clone_repo_helper.py
diff --git a/dis-segmentation/model/helpers.py b/image/dis-segmentation/model/helpers.py
similarity index 100%
rename from dis-segmentation/model/helpers.py
rename to image/dis-segmentation/model/helpers.py
diff --git a/dis-segmentation/model/model.py b/image/dis-segmentation/model/model.py
similarity index 100%
rename from dis-segmentation/model/model.py
rename to image/dis-segmentation/model/model.py
diff --git a/image/flux-dev-trt-b200/README.md b/image/flux-dev-trt-b200/README.md
new file mode 100644
index 000000000..98759196b
--- /dev/null
+++ b/image/flux-dev-trt-b200/README.md
@@ -0,0 +1,34 @@
+# Flux 1.0 Dev - TensorRT
+
+Generate high-quality images from text prompts using Black Forest Labs's Flux model with TensorRT optimization.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | B200 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Base image: `nvcr.io/nvidia/pytorch:25.06-py3`
+- Predict concurrency: **1**
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/image/flux-dev-trt-b200/config.yaml b/image/flux-dev-trt-b200/config.yaml
new file mode 100644
index 000000000..c72a45919
--- /dev/null
+++ b/image/flux-dev-trt-b200/config.yaml
@@ -0,0 +1,21 @@
+base_image:
+  image: nvcr.io/nvidia/pytorch:25.06-py3
+description: Generate high-quality images from text prompts using Black Forest Labs's Flux model with TensorRT optimization.
+external_package_dirs: []
+model_metadata:
+  repo_id: "black-forest-labs/FLUX.1-dev"
+  example_model_input:
+    prompt: A photo of an astronaut riding a horse
+model_name: Flux 1.0 Dev - TensorRT
+requirements_file: ./requirements.txt
+resources:
+  accelerator: B200
+  use_gpu: true
+runtime:
+  predict_concurrency: 1
+secrets:
+  hf_access_token: null
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/flux.1-dev-trt-b200/load_test.py b/image/flux-dev-trt-b200/load_test.py
similarity index 100%
rename from flux.1-dev-trt-b200/load_test.py
rename to image/flux-dev-trt-b200/load_test.py
diff --git a/flux.1-dev-trt-b200/model/__init__.py b/image/flux-dev-trt-b200/model/__init__.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/__init__.py
rename to image/flux-dev-trt-b200/model/__init__.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/__init__.py b/image/flux-dev-trt-b200/model/demo_diffusion/__init__.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/__init__.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/__init__.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/dd_argparse.py b/image/flux-dev-trt-b200/model/demo_diffusion/dd_argparse.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/dd_argparse.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/dd_argparse.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/dynamic_import.py b/image/flux-dev-trt-b200/model/demo_diffusion/dynamic_import.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/dynamic_import.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/dynamic_import.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/engine.py b/image/flux-dev-trt-b200/model/demo_diffusion/engine.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/engine.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/engine.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/image/__init__.py b/image/flux-dev-trt-b200/model/demo_diffusion/image/__init__.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/image/__init__.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/image/__init__.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/image/load.py b/image/flux-dev-trt-b200/model/demo_diffusion/image/load.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/image/load.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/image/load.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/image/resize.py b/image/flux-dev-trt-b200/model/demo_diffusion/image/resize.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/image/resize.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/image/resize.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/image/video.py b/image/flux-dev-trt-b200/model/demo_diffusion/image/video.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/image/video.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/image/video.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/__init__.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/__init__.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/__init__.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/__init__.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/base_model.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/base_model.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/base_model.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/base_model.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/clip.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/clip.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/clip.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/clip.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/diffusion_transformer.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/diffusion_transformer.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/diffusion_transformer.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/diffusion_transformer.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/gan.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/gan.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/gan.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/gan.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/load.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/load.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/load.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/load.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/lora.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/lora.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/lora.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/lora.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/optimizer.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/optimizer.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/optimizer.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/optimizer.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/scheduler.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/scheduler.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/scheduler.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/scheduler.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/t5.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/t5.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/t5.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/t5.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/tokenizer.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/tokenizer.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/tokenizer.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/tokenizer.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/unet.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/unet.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/unet.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/unet.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/model/vae.py b/image/flux-dev-trt-b200/model/demo_diffusion/model/vae.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/model/vae.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/model/vae.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/path/__init__.py b/image/flux-dev-trt-b200/model/demo_diffusion/path/__init__.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/path/__init__.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/path/__init__.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/path/dd_path.py b/image/flux-dev-trt-b200/model/demo_diffusion/path/dd_path.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/path/dd_path.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/path/dd_path.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/path/resolve_path.py b/image/flux-dev-trt-b200/model/demo_diffusion/path/resolve_path.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/path/resolve_path.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/path/resolve_path.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/__init__.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/__init__.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/__init__.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/__init__.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/calibrate.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/calibrate.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/calibrate.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/calibrate.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/diffusion_pipeline.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/diffusion_pipeline.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/diffusion_pipeline.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/diffusion_pipeline.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/flux_pipeline.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/flux_pipeline.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/flux_pipeline.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/flux_pipeline.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/model_memory_manager.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/model_memory_manager.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/model_memory_manager.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/model_memory_manager.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_cascade_pipeline.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_cascade_pipeline.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_cascade_pipeline.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_cascade_pipeline.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_35_pipeline.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_35_pipeline.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_35_pipeline.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_35_pipeline.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_3_pipeline.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_3_pipeline.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_3_pipeline.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_3_pipeline.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_pipeline.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_pipeline.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_pipeline.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_diffusion_pipeline.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_video_diffusion_pipeline.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_video_diffusion_pipeline.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/stable_video_diffusion_pipeline.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/stable_video_diffusion_pipeline.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/pipeline/type.py b/image/flux-dev-trt-b200/model/demo_diffusion/pipeline/type.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/pipeline/type.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/pipeline/type.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/utils_modelopt.py b/image/flux-dev-trt-b200/model/demo_diffusion/utils_modelopt.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/utils_modelopt.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/utils_modelopt.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/__init__.py b/image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/__init__.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/__init__.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/__init__.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/mmdit.py b/image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/mmdit.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/mmdit.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/mmdit.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/other_impls.py b/image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/other_impls.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/other_impls.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/other_impls.py
diff --git a/flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/sd3_impls.py b/image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/sd3_impls.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/demo_diffusion/utils_sd3/sd3_impls.py
rename to image/flux-dev-trt-b200/model/demo_diffusion/utils_sd3/sd3_impls.py
diff --git a/flux.1-dev-trt-b200/model/model.py b/image/flux-dev-trt-b200/model/model.py
similarity index 100%
rename from flux.1-dev-trt-b200/model/model.py
rename to image/flux-dev-trt-b200/model/model.py
diff --git a/flux.1-dev-trt-b200/requirements.txt b/image/flux-dev-trt-b200/requirements.txt
similarity index 100%
rename from flux.1-dev-trt-b200/requirements.txt
rename to image/flux-dev-trt-b200/requirements.txt
diff --git a/flux.1-dev-trt-b200/show.py b/image/flux-dev-trt-b200/show.py
similarity index 100%
rename from flux.1-dev-trt-b200/show.py
rename to image/flux-dev-trt-b200/show.py
diff --git a/flux.1-dev-trt-b200/show_batch.py b/image/flux-dev-trt-b200/show_batch.py
similarity index 100%
rename from flux.1-dev-trt-b200/show_batch.py
rename to image/flux-dev-trt-b200/show_batch.py
diff --git a/image/flux/README.md b/image/flux/README.md
new file mode 100644
index 000000000..bcfec3909
--- /dev/null
+++ b/image/flux/README.md
@@ -0,0 +1,46 @@
+# FLUX.1
+
+Deploy [FLUX.1](https://huggingface.co/black-forest-labs/FLUX.1-schnell) image generation models on Baseten. This directory contains two variants:
+
+| Variant | Path | GPU | HuggingFace |
+|---------|------|-----|-------------|
+| FLUX.1 Dev | [`dev/`](dev/) | H100 40GB | [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) |
+| FLUX.1 Schnell | [`schnell/`](schnell/) | H100 40GB | [black-forest-labs/FLUX.1-schnell](https://huggingface.co/black-forest-labs/FLUX.1-schnell) |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push image/flux/dev
+# or
+truss push image/flux/schnell
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "black forest gateau cake spelling out the words FLUX DEV, tasty, food photography, dynamic shot"}'
+```
+
+The model returns a dictionary with a `data` key containing a base64-encoded image:
+
+```python
+import requests
+import base64
+from PIL import Image
+from io import BytesIO
+
+res = requests.post(
+    "https://model-<model_id>.api.baseten.co/predict",
+    headers={"Authorization": "Api-Key YOUR_BASETEN_API_KEY"},
+    json={"prompt": "A tree in a field under the night sky"},
+)
+
+output = res.json()["data"]
+img = Image.open(BytesIO(base64.b64decode(output)))
+img.save("output_image.jpg")
+```
diff --git a/image/flux/dev/README.md b/image/flux/dev/README.md
new file mode 100644
index 000000000..8f0eeeadd
--- /dev/null
+++ b/image/flux/dev/README.md
@@ -0,0 +1,34 @@
+# Flux.1-dev
+
+Deploy [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [black-forest-labs/FLUX.1-dev](https://huggingface.co/black-forest-labs/FLUX.1-dev) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | H100_40GB |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "black forest gateau cake spelling out the words \"FLUX DEV\", tasty, food photography, dynamic shot"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/image/flux/dev/config.yaml b/image/flux/dev/config.yaml
new file mode 100644
index 000000000..c7bb29409
--- /dev/null
+++ b/image/flux/dev/config.yaml
@@ -0,0 +1,22 @@
+description: "black-forest-labs/FLUX.1-dev for image generation"
+external_package_dirs: []
+model_metadata:
+  example_model_input: {"prompt": 'black forest gateau cake spelling out the words "FLUX DEV", tasty, food photography, dynamic shot'}
+  repo_id: black-forest-labs/FLUX.1-dev
+model_name: Flux.1-dev
+python_version: py311
+requirements:
+  - git+https://github.com/huggingface/diffusers.git@fc6a91e3834c35e57b398ad1c0d99f6f83557e04
+  - transformers==4.36.0
+  - accelerate==0.25.0
+  - sentencepiece==0.1.99
+  - protobuf==4.25.1
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+secrets:
+  hf_access_token: null
+system_packages:
+  - ffmpeg
+  - libsm6
+  - libxext6
diff --git a/flux/dev/model/__init__.py b/image/flux/dev/model/__init__.py
similarity index 100%
rename from flux/dev/model/__init__.py
rename to image/flux/dev/model/__init__.py
diff --git a/flux/dev/model/model.py b/image/flux/dev/model/model.py
similarity index 100%
rename from flux/dev/model/model.py
rename to image/flux/dev/model/model.py
diff --git a/image/flux/schnell/README.md b/image/flux/schnell/README.md
new file mode 100644
index 000000000..a7cf3d25f
--- /dev/null
+++ b/image/flux/schnell/README.md
@@ -0,0 +1,34 @@
+# Flux.1-schnell
+
+Deploy [black-forest-labs/FLUX.1-schnell](https://huggingface.co/black-forest-labs/FLUX.1-schnell) for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [black-forest-labs/FLUX.1-schnell](https://huggingface.co/black-forest-labs/FLUX.1-schnell) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | H100_40GB |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "black forest gateau cake spelling out the words \"FLUX SCHNELL\", tasty, food photography, dynamic shot"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/image/flux/schnell/config.yaml b/image/flux/schnell/config.yaml
new file mode 100644
index 000000000..7a01e68c9
--- /dev/null
+++ b/image/flux/schnell/config.yaml
@@ -0,0 +1,23 @@
+description: "black-forest-labs/FLUX.1-schnell for image generation"
+external_package_dirs: []
+model_metadata:
+  example_model_input: {"prompt": 'black forest gateau cake spelling out the words "FLUX SCHNELL", tasty, food photography, dynamic shot'}
+  repo_id: black-forest-labs/FLUX.1-schnell
+model_name: Flux.1-schnell
+python_version: py311
+requirements:
+  - git+https://github.com/huggingface/diffusers.git@fc6a91e3834c35e57b398ad1c0d99f6f83557e04
+  - transformers==4.36.0
+  - accelerate==0.25.0
+  - sentencepiece==0.1.99
+  - protobuf==4.25.1
+  - b10-transfer==0.0.5
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+secrets:
+  hf_access_token: null
+system_packages:
+  - ffmpeg
+  - libsm6
+  - libxext6
diff --git a/flux/schnell/model/__init__.py b/image/flux/schnell/model/__init__.py
similarity index 100%
rename from flux/schnell/model/__init__.py
rename to image/flux/schnell/model/__init__.py
diff --git a/flux/schnell/model/model.py b/image/flux/schnell/model/model.py
similarity index 100%
rename from flux/schnell/model/model.py
rename to image/flux/schnell/model/model.py
diff --git a/image/fotographer/zenctrl/README.md b/image/fotographer/zenctrl/README.md
new file mode 100644
index 000000000..5ed287a46
--- /dev/null
+++ b/image/fotographer/zenctrl/README.md
@@ -0,0 +1,34 @@
+# ZenCtrl
+
+Deploy ZenCtrl for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Docker Server |
+| GPU | H100 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/generate \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Base image: `fotographerai/zenctrlstage:latest`
+- Predict concurrency: **8**
+- Environment variables: `PORT`
diff --git a/fotographer/zenctrl/call.py b/image/fotographer/zenctrl/call.py
similarity index 100%
rename from fotographer/zenctrl/call.py
rename to image/fotographer/zenctrl/call.py
diff --git a/image/fotographer/zenctrl/config.yaml b/image/fotographer/zenctrl/config.yaml
new file mode 100644
index 000000000..ec0a19e82
--- /dev/null
+++ b/image/fotographer/zenctrl/config.yaml
@@ -0,0 +1,26 @@
+description: "ZenCtrl for image generation"
+base_image:
+  image: fotographerai/zenctrlstage:latest
+model_metadata:
+  repo_id: "Kijai/ZenCtrl-Flux-comfy"
+  example_model_input:
+    image: "<base64-encoded image>"
+    prompt: A man holding a camera facing the objective
+    steps: 10
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python -m uvicorn app:app --host 0.0.0.0 --port 8000 --log-level debug"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /generate
+  server_port: 8000
+resources:
+  accelerator: H100
+  use_gpu: true
+model_name: ZenCtrl
+environment_variables:
+  PORT: 8000
+  HF_TOKEN: null
+runtime:
+  predict_concurrency: 8
+secrets:
+  hf_access_token: null
diff --git a/fotographer/zenctrl/images/banner_1.png b/image/fotographer/zenctrl/images/banner_1.png
similarity index 100%
rename from fotographer/zenctrl/images/banner_1.png
rename to image/fotographer/zenctrl/images/banner_1.png
diff --git a/fotographer/zenctrl/images/camera.png b/image/fotographer/zenctrl/images/camera.png
similarity index 100%
rename from fotographer/zenctrl/images/camera.png
rename to image/fotographer/zenctrl/images/camera.png
diff --git a/fotographer/zenctrl/images/speaker-input.png b/image/fotographer/zenctrl/images/speaker-input.png
similarity index 100%
rename from fotographer/zenctrl/images/speaker-input.png
rename to image/fotographer/zenctrl/images/speaker-input.png
diff --git a/fotographer/zenctrl/images/speaker-output.png b/image/fotographer/zenctrl/images/speaker-output.png
similarity index 100%
rename from fotographer/zenctrl/images/speaker-output.png
rename to image/fotographer/zenctrl/images/speaker-output.png
diff --git a/fotographer/zenctrl/requirements.txt b/image/fotographer/zenctrl/requirements.txt
similarity index 100%
rename from fotographer/zenctrl/requirements.txt
rename to image/fotographer/zenctrl/requirements.txt
diff --git a/whisper/whisper-truss/LICENSE b/image/gfp-gan/LICENSE
similarity index 100%
rename from whisper/whisper-truss/LICENSE
rename to image/gfp-gan/LICENSE
diff --git a/image/gfp-gan/README.md b/image/gfp-gan/README.md
new file mode 100644
index 000000000..8a5ef5398
--- /dev/null
+++ b/image/gfp-gan/README.md
@@ -0,0 +1,31 @@
+# GFP-GAN
+
+Restore photos with this image-to-image model.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | GPU (unspecified) |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/image/gfp-gan/config.yaml b/image/gfp-gan/config.yaml
new file mode 100644
index 000000000..86d93b6ad
--- /dev/null
+++ b/image/gfp-gan/config.yaml
@@ -0,0 +1,34 @@
+description: Restore photos with this image-to-image model.
+environment_variables: {}
+external_data:
+- local_data_path: RealESRGAN_x2plus.pth
+  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/gfp-gan/RealESRGAN_x2plus.pth
+- local_data_path: GFPGANv1.3.pth
+  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/gfp-gan/GFPGANv1.3.pth
+external_package_dirs: []
+model_metadata:
+  repo_id: "TencentARC/gfpgan"
+  avatar_url: https://cdn.baseten.co/production/static/explore/tencent.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/gfp-gan.png
+  example_model_input:
+    image: "<base64-encoded image>"
+  example_model_input_file: input.json
+  tags:
+  - image-restoration
+model_name: GFP-GAN
+python_version: py39
+requirements:
+- gfpgan==1.3.8
+- realesrgan==0.3.0
+- basicsr==1.4.2
+- torchvision==0.16.2
+- numpy==1.26.4
+resources:
+  cpu: '3'
+  memory: 8Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/__init__ copy.py b/image/gfp-gan/data/.gitkeep
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/__init__ copy.py
rename to image/gfp-gan/data/.gitkeep
diff --git a/gfp-gan/input.json b/image/gfp-gan/input.json
similarity index 100%
rename from gfp-gan/input.json
rename to image/gfp-gan/input.json
diff --git a/gemma/gemma-2-27b-it-vllm/model/__init__.py b/image/gfp-gan/model/__init__.py
similarity index 100%
rename from gemma/gemma-2-27b-it-vllm/model/__init__.py
rename to image/gfp-gan/model/__init__.py
diff --git a/gfp-gan/model/model.py b/image/gfp-gan/model/model.py
similarity index 100%
rename from gfp-gan/model/model.py
rename to image/gfp-gan/model/model.py
diff --git a/image/image-segmentation/README.md b/image/image-segmentation/README.md
new file mode 100644
index 000000000..976ed76aa
--- /dev/null
+++ b/image/image-segmentation/README.md
@@ -0,0 +1,31 @@
+# Image segmentation
+
+Deploy Image segmentation for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | CPU |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/image-segmentation/config.yaml b/image/image-segmentation/config.yaml
new file mode 100644
index 000000000..97e572e75
--- /dev/null
+++ b/image/image-segmentation/config.yaml
@@ -0,0 +1,19 @@
+description: "Image segmentation for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "nvidia/segformer-b0-finetuned-ade-512-512"
+  example_model_input:
+    instances:
+    - image_url: https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg
+model_name: Image segmentation
+python_version: py39
+requirements:
+- torchvision==0.9.1
+resources:
+  cpu: 3000m
+  memory: 8Gi
+  use_gpu: false
+secrets: {}
+spec_version: "2.0"
+system_packages: []
diff --git a/gemma/gemma-2-9b-it-vllm/model/__init__.py b/image/image-segmentation/model/__init__.py
similarity index 100%
rename from gemma/gemma-2-9b-it-vllm/model/__init__.py
rename to image/image-segmentation/model/__init__.py
diff --git a/image-segmentation/model/model.py b/image/image-segmentation/model/model.py
similarity index 100%
rename from image-segmentation/model/model.py
rename to image/image-segmentation/model/model.py
diff --git a/image/ip-adapter/README.md b/image/ip-adapter/README.md
new file mode 100644
index 000000000..0c6c1066f
--- /dev/null
+++ b/image/ip-adapter/README.md
@@ -0,0 +1,31 @@
+# IP Adapter
+
+Deploy IP Adapter for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/ip-adapter/config.yaml b/image/ip-adapter/config.yaml
new file mode 100644
index 000000000..8d39291ef
--- /dev/null
+++ b/image/ip-adapter/config.yaml
@@ -0,0 +1,20 @@
+description: "IP Adapter for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "h94/IP-Adapter"
+  example_model_input:
+    image: "<base64-encoded image>"
+model_name: IP Adapter
+python_version: py311
+requirements:
+- torch==2.1.1
+- diffusers==0.24.0
+- transformers==4.35.2
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 15Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/gfp-gan/model/__init__.py b/image/ip-adapter/model/__init__.py
similarity index 100%
rename from gfp-gan/model/__init__.py
rename to image/ip-adapter/model/__init__.py
diff --git a/ip-adapter/model/model.py b/image/ip-adapter/model/model.py
similarity index 100%
rename from ip-adapter/model/model.py
rename to image/ip-adapter/model/model.py
diff --git a/image/magic-animate/README.md b/image/magic-animate/README.md
new file mode 100644
index 000000000..a1723d1aa
--- /dev/null
+++ b/image/magic-animate/README.md
@@ -0,0 +1,35 @@
+# Magic Animate
+
+Deploy Magic Animate for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "guidance_scale": 7.5,
+  "motion_sequence": "<BASE64 MP4 FILE>",
+  "reference_image": "<BASE64 IMAGE>",
+  "seed": 1,
+  "steps": 10
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/image/magic-animate/config.yaml b/image/magic-animate/config.yaml
new file mode 100644
index 000000000..e3b649a04
--- /dev/null
+++ b/image/magic-animate/config.yaml
@@ -0,0 +1,38 @@
+description: "Magic Animate for video animation from images"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "zcxu-eric/MagicAnimate"
+  example_model_input:
+    guidance_scale: 7.5
+    motion_sequence: <BASE64 MP4 FILE>
+    reference_image: <BASE64 IMAGE>
+    seed: 1
+    steps: 10
+model_name: Magic Animate
+python_version: py310
+requirements:
+- torch==2.0.1
+- torchvision==0.15.2
+- xformers==0.0.22
+- diffusers==0.21.4
+- pillow==9.5.0
+- numpy==1.24.4
+- omegaconf==2.3.0
+- transformers==4.32.0
+- einops==0.6.1
+- imageio==2.9.0
+- imageio-ffmpeg==0.4.3
+- tqdm==4.66.1
+- websockets==11.0.3
+- accelerate==0.22.0
+- huggingface-hub==0.16.4
+- av==11.0.0
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 15Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
diff --git a/image-segmentation/model/__init__.py b/image/magic-animate/model/__init__.py
similarity index 100%
rename from image-segmentation/model/__init__.py
rename to image/magic-animate/model/__init__.py
diff --git a/magic-animate/model/configs/inference/inference.yaml b/image/magic-animate/model/configs/inference/inference.yaml
similarity index 100%
rename from magic-animate/model/configs/inference/inference.yaml
rename to image/magic-animate/model/configs/inference/inference.yaml
diff --git a/magic-animate/model/configs/prompts/animation.yaml b/image/magic-animate/model/configs/prompts/animation.yaml
similarity index 100%
rename from magic-animate/model/configs/prompts/animation.yaml
rename to image/magic-animate/model/configs/prompts/animation.yaml
diff --git a/magic-animate/model/demo/animate.py b/image/magic-animate/model/demo/animate.py
similarity index 100%
rename from magic-animate/model/demo/animate.py
rename to image/magic-animate/model/demo/animate.py
diff --git a/magic-animate/model/magicanimate/models/appearance_encoder.py b/image/magic-animate/model/magicanimate/models/appearance_encoder.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/appearance_encoder.py
rename to image/magic-animate/model/magicanimate/models/appearance_encoder.py
diff --git a/magic-animate/model/magicanimate/models/attention.py b/image/magic-animate/model/magicanimate/models/attention.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/attention.py
rename to image/magic-animate/model/magicanimate/models/attention.py
diff --git a/magic-animate/model/magicanimate/models/controlnet.py b/image/magic-animate/model/magicanimate/models/controlnet.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/controlnet.py
rename to image/magic-animate/model/magicanimate/models/controlnet.py
diff --git a/magic-animate/model/magicanimate/models/embeddings.py b/image/magic-animate/model/magicanimate/models/embeddings.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/embeddings.py
rename to image/magic-animate/model/magicanimate/models/embeddings.py
diff --git a/magic-animate/model/magicanimate/models/motion_module.py b/image/magic-animate/model/magicanimate/models/motion_module.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/motion_module.py
rename to image/magic-animate/model/magicanimate/models/motion_module.py
diff --git a/magic-animate/model/magicanimate/models/mutual_self_attention.py b/image/magic-animate/model/magicanimate/models/mutual_self_attention.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/mutual_self_attention.py
rename to image/magic-animate/model/magicanimate/models/mutual_self_attention.py
diff --git a/magic-animate/model/magicanimate/models/orig_attention.py b/image/magic-animate/model/magicanimate/models/orig_attention.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/orig_attention.py
rename to image/magic-animate/model/magicanimate/models/orig_attention.py
diff --git a/magic-animate/model/magicanimate/models/resnet.py b/image/magic-animate/model/magicanimate/models/resnet.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/resnet.py
rename to image/magic-animate/model/magicanimate/models/resnet.py
diff --git a/magic-animate/model/magicanimate/models/stable_diffusion_controlnet_reference.py b/image/magic-animate/model/magicanimate/models/stable_diffusion_controlnet_reference.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/stable_diffusion_controlnet_reference.py
rename to image/magic-animate/model/magicanimate/models/stable_diffusion_controlnet_reference.py
diff --git a/magic-animate/model/magicanimate/models/unet.py b/image/magic-animate/model/magicanimate/models/unet.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/unet.py
rename to image/magic-animate/model/magicanimate/models/unet.py
diff --git a/magic-animate/model/magicanimate/models/unet_3d_blocks.py b/image/magic-animate/model/magicanimate/models/unet_3d_blocks.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/unet_3d_blocks.py
rename to image/magic-animate/model/magicanimate/models/unet_3d_blocks.py
diff --git a/magic-animate/model/magicanimate/models/unet_controlnet.py b/image/magic-animate/model/magicanimate/models/unet_controlnet.py
similarity index 100%
rename from magic-animate/model/magicanimate/models/unet_controlnet.py
rename to image/magic-animate/model/magicanimate/models/unet_controlnet.py
diff --git a/magic-animate/model/magicanimate/pipelines/animation.py b/image/magic-animate/model/magicanimate/pipelines/animation.py
similarity index 100%
rename from magic-animate/model/magicanimate/pipelines/animation.py
rename to image/magic-animate/model/magicanimate/pipelines/animation.py
diff --git a/magic-animate/model/magicanimate/pipelines/context.py b/image/magic-animate/model/magicanimate/pipelines/context.py
similarity index 100%
rename from magic-animate/model/magicanimate/pipelines/context.py
rename to image/magic-animate/model/magicanimate/pipelines/context.py
diff --git a/magic-animate/model/magicanimate/pipelines/pipeline_animation.py b/image/magic-animate/model/magicanimate/pipelines/pipeline_animation.py
similarity index 100%
rename from magic-animate/model/magicanimate/pipelines/pipeline_animation.py
rename to image/magic-animate/model/magicanimate/pipelines/pipeline_animation.py
diff --git a/magic-animate/model/magicanimate/utils/dist_tools.py b/image/magic-animate/model/magicanimate/utils/dist_tools.py
similarity index 100%
rename from magic-animate/model/magicanimate/utils/dist_tools.py
rename to image/magic-animate/model/magicanimate/utils/dist_tools.py
diff --git a/magic-animate/model/magicanimate/utils/util.py b/image/magic-animate/model/magicanimate/utils/util.py
similarity index 100%
rename from magic-animate/model/magicanimate/utils/util.py
rename to image/magic-animate/model/magicanimate/utils/util.py
diff --git a/magic-animate/model/magicanimate/utils/videoreader.py b/image/magic-animate/model/magicanimate/utils/videoreader.py
similarity index 100%
rename from magic-animate/model/magicanimate/utils/videoreader.py
rename to image/magic-animate/model/magicanimate/utils/videoreader.py
diff --git a/magic-animate/model/model.py b/image/magic-animate/model/model.py
similarity index 100%
rename from magic-animate/model/model.py
rename to image/magic-animate/model/model.py
diff --git a/image/playground-v2-aesthetic/README.md b/image/playground-v2-aesthetic/README.md
new file mode 100644
index 000000000..979cf58a8
--- /dev/null
+++ b/image/playground-v2-aesthetic/README.md
@@ -0,0 +1,33 @@
+# Playground V2 Aesthetic
+
+Generate original images from text prompts.
+
+| Property | Value |
+|----------|-------|
+| Model | [playgroundai/playground-v2-1024px-aesthetic](https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "num_inference_steps": 50,
+  "prompt": "A scenic mountain landscape"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/playground-v2-aesthetic/config.yaml b/image/playground-v2-aesthetic/config.yaml
similarity index 100%
rename from playground-v2-aesthetic/config.yaml
rename to image/playground-v2-aesthetic/config.yaml
diff --git a/ip-adapter/model/__init__.py b/image/playground-v2-aesthetic/model/__init__.py
similarity index 100%
rename from ip-adapter/model/__init__.py
rename to image/playground-v2-aesthetic/model/__init__.py
diff --git a/playground-v2-aesthetic/model/model.py b/image/playground-v2-aesthetic/model/model.py
similarity index 100%
rename from playground-v2-aesthetic/model/model.py
rename to image/playground-v2-aesthetic/model/model.py
diff --git a/playground-v2-aesthetic/show.py b/image/playground-v2-aesthetic/show.py
similarity index 100%
rename from playground-v2-aesthetic/show.py
rename to image/playground-v2-aesthetic/show.py
diff --git a/image/sana/sana-1600m/README.md b/image/sana/sana-1600m/README.md
new file mode 100644
index 000000000..539ead590
--- /dev/null
+++ b/image/sana/sana-1600m/README.md
@@ -0,0 +1,40 @@
+# Sana 1600M
+
+Deploy Sana 1600M for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | H100_40GB |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "a photo of an astronaut riding a horse on mars",
+  "height": 1024,
+  "width": 1024,
+  "guidance_scale": 5.0,
+  "pag_guidance_scale": 2.0,
+  "num_inference_steps": 18,
+  "seed": 4096
+}'
+```
+
+## Configuration highlights
+
+- Base image: `alphatozeta/cuda-python:12.1.1-cudnn8-devel-ubuntu22.04`
+- System packages: `ffmpeg, libsm6, libxext6, python3.10-venv`
diff --git a/image/sana/sana-1600m/config.yaml b/image/sana/sana-1600m/config.yaml
new file mode 100644
index 000000000..1e1bf8e9c
--- /dev/null
+++ b/image/sana/sana-1600m/config.yaml
@@ -0,0 +1,33 @@
+description: "Sana 1600M for image generation"
+build_commands: []
+base_image:
+  image: alphatozeta/cuda-python:12.1.1-cudnn8-devel-ubuntu22.04
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "Efficient-Large-Model/Sana_1600M_1024px_MultiLing_diffusers"
+  example_model_input: {
+    "prompt": "a photo of an astronaut riding a horse on mars",
+    "height": 1024,
+    "width": 1024,
+    "guidance_scale": 5.0,
+    "pag_guidance_scale": 2.0,
+    "num_inference_steps": 18,
+    "seed": 4096,
+  }
+model_name: Sana 1600M
+python_version: py311
+requirements:
+- git+https://github.com/NVlabs/Sana.git@d7945026d8d85008aca1d1e6db5717a1069f5c84
+- huggingface-hub==0.26.3
+- hf-transfer==0.1.8
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+secrets:
+  hf_access_token: "null"
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
+- python3.10-venv
diff --git a/jsonformatter/model/__init__.py b/image/sana/sana-1600m/model/__init__.py
similarity index 100%
rename from jsonformatter/model/__init__.py
rename to image/sana/sana-1600m/model/__init__.py
diff --git a/sana/sana_1600M/model/model.py b/image/sana/sana-1600m/model/model.py
similarity index 100%
rename from sana/sana_1600M/model/model.py
rename to image/sana/sana-1600m/model/model.py
diff --git a/sana/sana_1600M/packages/Sana/CITATION.bib b/image/sana/sana-1600m/packages/Sana/CITATION.bib
similarity index 100%
rename from sana/sana_1600M/packages/Sana/CITATION.bib
rename to image/sana/sana-1600m/packages/Sana/CITATION.bib
diff --git a/sana/sana_1600M/packages/Sana/CIs/add_license_all.sh b/image/sana/sana-1600m/packages/Sana/CIs/add_license_all.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/CIs/add_license_all.sh
rename to image/sana/sana-1600m/packages/Sana/CIs/add_license_all.sh
diff --git a/sana/sana_1600M/packages/Sana/Dockerfile b/image/sana/sana-1600m/packages/Sana/Dockerfile
similarity index 100%
rename from sana/sana_1600M/packages/Sana/Dockerfile
rename to image/sana/sana-1600m/packages/Sana/Dockerfile
diff --git a/sana/sana_1600M/packages/Sana/LICENSE b/image/sana/sana-1600m/packages/Sana/LICENSE
similarity index 100%
rename from sana/sana_1600M/packages/Sana/LICENSE
rename to image/sana/sana-1600m/packages/Sana/LICENSE
diff --git a/sana/sana_1600M/packages/Sana/README.md b/image/sana/sana-1600m/packages/Sana/README.md
similarity index 100%
rename from sana/sana_1600M/packages/Sana/README.md
rename to image/sana/sana-1600m/packages/Sana/README.md
diff --git a/sana/sana_1600M/packages/Sana/app/app_sana.py b/image/sana/sana-1600m/packages/Sana/app/app_sana.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/app/app_sana.py
rename to image/sana/sana-1600m/packages/Sana/app/app_sana.py
diff --git a/sana/sana_1600M/packages/Sana/app/app_sana_multithread.py b/image/sana/sana-1600m/packages/Sana/app/app_sana_multithread.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/app/app_sana_multithread.py
rename to image/sana/sana-1600m/packages/Sana/app/app_sana_multithread.py
diff --git a/sana/sana_1600M/packages/Sana/app/safety_check.py b/image/sana/sana-1600m/packages/Sana/app/safety_check.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/app/safety_check.py
rename to image/sana/sana-1600m/packages/Sana/app/safety_check.py
diff --git a/sana/sana_1600M/packages/Sana/app/sana_pipeline.py b/image/sana/sana-1600m/packages/Sana/app/sana_pipeline.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/app/sana_pipeline.py
rename to image/sana/sana-1600m/packages/Sana/app/sana_pipeline.py
diff --git a/sana/sana_1600M/packages/Sana/asset/Sana.jpg b/image/sana/sana-1600m/packages/Sana/asset/Sana.jpg
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/Sana.jpg
rename to image/sana/sana-1600m/packages/Sana/asset/Sana.jpg
diff --git a/sana/sana_1600M/packages/Sana/asset/docs/metrics_toolkit.md b/image/sana/sana-1600m/packages/Sana/asset/docs/metrics_toolkit.md
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/docs/metrics_toolkit.md
rename to image/sana/sana-1600m/packages/Sana/asset/docs/metrics_toolkit.md
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/00000000.png b/image/sana/sana-1600m/packages/Sana/asset/example_data/00000000.png
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/00000000.png
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/00000000.png
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/00000000.txt b/image/sana/sana-1600m/packages/Sana/asset/example_data/00000000.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/00000000.txt
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/00000000.txt
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/00000000_InternVL2-26B.json b/image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_InternVL2-26B.json
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/00000000_InternVL2-26B.json
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_InternVL2-26B.json
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json b/image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json b/image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json b/image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/00000000_prompt_clip_score.json b/image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_prompt_clip_score.json
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/00000000_prompt_clip_score.json
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/00000000_prompt_clip_score.json
diff --git a/sana/sana_1600M/packages/Sana/asset/example_data/meta_data.json b/image/sana/sana-1600m/packages/Sana/asset/example_data/meta_data.json
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/example_data/meta_data.json
rename to image/sana/sana-1600m/packages/Sana/asset/example_data/meta_data.json
diff --git a/sana/sana_1600M/packages/Sana/asset/examples.py b/image/sana/sana-1600m/packages/Sana/asset/examples.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/examples.py
rename to image/sana/sana-1600m/packages/Sana/asset/examples.py
diff --git a/sana/sana_1600M/packages/Sana/asset/logo.png b/image/sana/sana-1600m/packages/Sana/asset/logo.png
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/logo.png
rename to image/sana/sana-1600m/packages/Sana/asset/logo.png
diff --git a/sana/sana_1600M/packages/Sana/asset/model-incremental.jpg b/image/sana/sana-1600m/packages/Sana/asset/model-incremental.jpg
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/model-incremental.jpg
rename to image/sana/sana-1600m/packages/Sana/asset/model-incremental.jpg
diff --git a/sana/sana_1600M/packages/Sana/asset/model_paths.txt b/image/sana/sana-1600m/packages/Sana/asset/model_paths.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/model_paths.txt
rename to image/sana/sana-1600m/packages/Sana/asset/model_paths.txt
diff --git a/sana/sana_1600M/packages/Sana/asset/samples.txt b/image/sana/sana-1600m/packages/Sana/asset/samples.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/samples.txt
rename to image/sana/sana-1600m/packages/Sana/asset/samples.txt
diff --git a/sana/sana_1600M/packages/Sana/asset/samples_mini.txt b/image/sana/sana-1600m/packages/Sana/asset/samples_mini.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/asset/samples_mini.txt
rename to image/sana/sana-1600m/packages/Sana/asset/samples_mini.txt
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_base.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_base.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_base.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_base.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml
diff --git a/sana/sana_1600M/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml b/image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml
rename to image/sana/sana-1600m/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml
diff --git a/sana/sana_1600M/packages/Sana/diffusion/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/builder.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/builder.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/builder.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/builder.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/datasets/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/datasets/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/datasets/sana_data.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/sana_data.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/datasets/sana_data.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/sana_data.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/datasets/utils.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/utils.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/datasets/utils.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/datasets/utils.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/transforms.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/transforms.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/transforms.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/transforms.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/wids/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/wids/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/wids/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/wids/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/wids/wids.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/wids/wids.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_dl.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_dl.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_dl.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_dl.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_lru.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_lru.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_lru.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_lru.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_mmtar.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_mmtar.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_mmtar.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_mmtar.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_specs.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_specs.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_specs.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_specs.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_tar.py b/image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_tar.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/data/wids/wids_tar.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/data/wids/wids_tar.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/dpm_solver.py b/image/sana/sana-1600m/packages/Sana/diffusion/dpm_solver.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/dpm_solver.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/dpm_solver.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/flow_euler_sampler.py b/image/sana/sana-1600m/packages/Sana/diffusion/flow_euler_sampler.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/flow_euler_sampler.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/flow_euler_sampler.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/iddpm.py b/image/sana/sana-1600m/packages/Sana/diffusion/iddpm.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/iddpm.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/iddpm.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/lcm_scheduler.py b/image/sana/sana-1600m/packages/Sana/diffusion/lcm_scheduler.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/lcm_scheduler.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/lcm_scheduler.py
diff --git a/kokoro/model/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/__init__.py
old mode 100644
new mode 100755
similarity index 100%
rename from kokoro/model/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/act.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/act.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/act.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/act.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/builder.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/builder.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/builder.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/builder.py
diff --git a/layoutlm-document-qa/model/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py
similarity index 100%
rename from layoutlm-document-qa/model/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py
diff --git a/llama/llama-2-13b-chat/model/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py
similarity index 100%
rename from llama/llama-2-13b-chat/model/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py
diff --git a/llama/llama-2-13b/model/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py
similarity index 100%
rename from llama/llama-2-13b/model/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/diffusion_utils.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/diffusion_utils.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/diffusion_utils.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/diffusion_utils.py
diff --git a/image/sana/sana-1600m/packages/Sana/diffusion/model/dpm_solver.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/dpm_solver.py
new file mode 100755
index 000000000..a791dcd38
--- /dev/null
+++ b/image/sana/sana-1600m/packages/Sana/diffusion/model/dpm_solver.py
@@ -0,0 +1,1908 @@
+# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# This file is modified from https://github.com/PixArt-alpha/PixArt-sigma
+import os
+
+import torch
+from tqdm import tqdm
+
+from .nets.sana_blocks import (
+    PAGCFGIdentitySelfAttnProcessorLiteLA,
+    PAGIdentitySelfAttnProcessorLiteLA,
+    SelfAttnProcessorLiteLA,
+)
+
+
+class NoiseScheduleVP:
+    def __init__(
+        self,
+        schedule="discrete",
+        betas=None,
+        alphas_cumprod=None,
+        continuous_beta_0=0.1,
+        continuous_beta_1=20.0,
+        dtype=torch.float32,
+    ):
+        r"""Create a wrapper class for the forward SDE (VP type).
+
+        ***
+        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
+                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
+        ***
+
+        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
+        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
+        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
+
+            log_alpha_t = self.marginal_log_mean_coeff(t)
+            sigma_t = self.marginal_std(t)
+            lambda_t = self.marginal_lambda(t)
+
+        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
+
+            t = self.inverse_lambda(lambda_t)
+
+        ===============================================================
+
+        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
+
+        1. For discrete-time DPMs:
+
+            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
+                t_i = (i + 1) / N
+            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
+            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
+
+            Args:
+                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
+                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
+
+            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
+
+            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
+                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
+                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
+                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
+                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
+                and
+                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
+
+
+        2. For continuous-time DPMs:
+
+            We support the linear VPSDE for the continuous time setting. The hyperparameters for the noise
+            schedule are the default settings in Yang Song's ScoreSDE:
+
+            Args:
+                beta_min: A `float` number. The smallest beta for the linear schedule.
+                beta_max: A `float` number. The largest beta for the linear schedule.
+                T: A `float` number. The ending time of the forward process.
+
+        ===============================================================
+
+        Args:
+            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
+                    'linear' for continuous-time DPMs.
+        Returns:
+            A wrapper object of the forward SDE (VP type).
+
+        ===============================================================
+
+        Example:
+
+        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', betas=betas)
+
+        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
+
+        # For continuous-time DPMs (VPSDE), linear schedule:
+        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
+
+        """
+
+        if schedule not in ["discrete", "linear"]:
+            raise ValueError(
+                f"Unsupported noise schedule {schedule}. The schedule needs to be 'discrete' or 'linear'"
+            )
+
+        self.schedule = schedule
+        if schedule == "discrete":
+            if betas is not None:
+                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
+            else:
+                assert alphas_cumprod is not None
+                log_alphas = 0.5 * torch.log(alphas_cumprod)
+            self.T = 1.0
+            self.log_alpha_array = (
+                self.numerical_clip_alpha(log_alphas)
+                .reshape(
+                    (
+                        1,
+                        -1,
+                    )
+                )
+                .to(dtype=dtype)
+            )
+            self.total_N = self.log_alpha_array.shape[1]
+            self.t_array = (
+                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
+                .reshape((1, -1))
+                .to(dtype=dtype)
+            )
+        else:
+            self.T = 1.0
+            self.total_N = 1000
+            self.beta_0 = continuous_beta_0
+            self.beta_1 = continuous_beta_1
+
+    def numerical_clip_alpha(self, log_alphas, clipped_lambda=-5.1):
+        """
+        For some beta schedules such as cosine schedule, the log-SNR has numerical isssues.
+        We clip the log-SNR near t=T within -5.1 to ensure the stability.
+        Such a trick is very useful for diffusion models with the cosine schedule, such as i-DDPM, guided-diffusion and GLIDE.
+        """
+        log_sigmas = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_alphas))
+        lambs = log_alphas - log_sigmas
+        idx = torch.searchsorted(torch.flip(lambs, [0]), clipped_lambda)
+        if idx > 0:
+            log_alphas = log_alphas[:-idx]
+        return log_alphas
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        if self.schedule == "discrete":
+            return interpolate_fn(
+                t.reshape((-1, 1)),
+                self.t_array.to(t.device),
+                self.log_alpha_array.to(t.device),
+            ).reshape(-1)
+        elif self.schedule == "linear":
+            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return torch.exp(self.marginal_log_mean_coeff(t))
+
+    def marginal_std(self, t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
+        return log_mean_coeff - log_std
+
+    def inverse_lambda(self, lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        if self.schedule == "linear":
+            tmp = (
+                2.0
+                * (self.beta_1 - self.beta_0)
+                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
+            )
+            Delta = self.beta_0**2 + tmp
+            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
+        elif self.schedule == "discrete":
+            log_alpha = -0.5 * torch.logaddexp(
+                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
+            )
+            t = interpolate_fn(
+                log_alpha.reshape((-1, 1)),
+                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
+                torch.flip(self.t_array.to(lamb.device), [1]),
+            )
+            return t.reshape((-1,))
+
+
+class NoiseScheduleFlow:
+    def __init__(
+        self,
+        schedule="discrete_flow",
+    ):
+        """Create a wrapper class for the forward SDE (EDM type)."""
+        self.T = 1
+        self.t0 = 0.001
+        self.schedule = schedule  # ['continuous', 'discrete_flow']
+        self.total_N = 1000
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        return torch.log(self.marginal_alpha(t))
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return 1 - t
+
+    @staticmethod
+    def marginal_std(t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return t
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = torch.log(self.marginal_std(t))
+        return log_mean_coeff - log_std
+
+    @staticmethod
+    def inverse_lambda(lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        return torch.exp(-lamb)
+
+    def edm_sigma(self, t):
+        return self.marginal_std(t) / self.marginal_alpha(t)
+
+    def edm_inverse_sigma(self, edmsigma):
+        sigma = edmsigma
+        lambda_t = torch.log(1 / sigma)
+        t = self.inverse_lambda(lambda_t)
+        return t
+
+
+def model_wrapper(
+    model,
+    noise_schedule,
+    model_type="noise",
+    model_kwargs={},
+    guidance_type="uncond",
+    condition=None,
+    unconditional_condition=None,
+    guidance_scale=1.0,
+    pag_scale=1.0,
+    pag_applied_layers=[],
+    interval_guidance=[0, 1.0],
+    classifier_fn=None,
+    classifier_kwargs={},
+):
+    """Create a wrapper function for the noise prediction model.
+
+    DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
+    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
+
+    We support four types of the diffusion model by setting `model_type`:
+
+        1. "noise": noise prediction model. (Trained by predicting noise).
+
+        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
+
+        3. "v": velocity prediction model. (Trained by predicting the velocity).
+            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
+
+            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
+                arXiv preprint arXiv:2202.00512 (2022).
+            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
+                arXiv preprint arXiv:2210.02303 (2022).
+
+        4. "score": marginal score function. (Trained by denoising score matching).
+            Note that the score function and the noise prediction model follows a simple relationship:
+            ```
+                noise(x_t, t) = -sigma_t * score(x_t, t)
+            ```
+
+    We support three types of guided sampling by DPMs by setting `guidance_type`:
+        1. "uncond": unconditional sampling by DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+
+        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+
+            The input `classifier_fn` has the following format:
+            ``
+                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
+            ``
+
+            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
+                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
+
+        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
+            ``
+            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
+
+            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
+                arXiv preprint arXiv:2207.12598 (2022).
+
+
+    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
+    or continuous-time labels (i.e. epsilon to T).
+
+    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
+    ``
+        def model_fn(x, t_continuous) -> noise:
+            t_input = get_model_input_time(t_continuous)
+            return noise_pred(model, x, t_input, **model_kwargs)
+    ``
+    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
+
+    ===============================================================
+
+    Args:
+        model: A diffusion model with the corresponding format described above.
+        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+        model_type: A `str`. The parameterization type of the diffusion model.
+                    "noise" or "x_start" or "v" or "score".
+        model_kwargs: A `dict`. A dict for the other inputs of the model function.
+        guidance_type: A `str`. The type of the guidance for sampling.
+                    "uncond" or "classifier" or "classifier-free".
+        condition: A pytorch tensor. The condition for the guided sampling.
+                    Only used for "classifier" or "classifier-free" guidance type.
+        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
+                    Only used for "classifier-free" guidance type.
+        guidance_scale: A `float`. The scale for the guided sampling.
+        classifier_fn: A classifier function. Only used for the classifier guidance.
+        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
+    Returns:
+        A noise prediction model that accepts the noised data and the continuous time as the inputs.
+    """
+
+    def get_model_input_time(t_continuous):
+        """
+        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
+        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
+        For continuous-time DPMs, we just use `t_continuous`.
+        """
+        if noise_schedule.schedule == "discrete":
+            return (
+                t_continuous - 1.0 / noise_schedule.total_N
+            ) * noise_schedule.total_N
+        elif noise_schedule.schedule == "discrete_flow":
+            return t_continuous * noise_schedule.total_N
+        else:
+            return t_continuous
+
+    def noise_pred_fn(x, t_continuous, cond=None):
+        t_input = get_model_input_time(t_continuous)
+        if cond is None:
+            output = model(x, t_input, **model_kwargs)
+        else:
+            output = model(x, t_input, cond, **model_kwargs)
+        if model_type == "noise":
+            return output
+        elif model_type == "x_start":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return (x - expand_dims(alpha_t, x.dim()) * output) / expand_dims(
+                sigma_t, x.dim()
+            )
+        elif model_type == "v":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return (
+                expand_dims(alpha_t, x.dim()) * output
+                + expand_dims(sigma_t, x.dim()) * x
+            )
+        elif model_type == "score":
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            return -expand_dims(sigma_t, x.dim()) * output
+        elif model_type == "flow":
+            _, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            try:
+                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output + x
+            except:
+                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output[0] + x
+            return noise
+
+    def cond_grad_fn(x, t_input):
+        """
+        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
+        """
+        with torch.enable_grad():
+            x_in = x.detach().requires_grad_(True)
+            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
+            return torch.autograd.grad(log_prob.sum(), x_in)[0]
+
+    def model_fn(x, t_continuous):
+        """
+        The noise predicition model function that is used for DPM-Solver.
+        """
+        guidance_tp = guidance_type
+        if guidance_tp == "uncond":
+            return noise_pred_fn(x, t_continuous)
+        elif guidance_tp == "classifier":
+            assert classifier_fn is not None
+            t_input = get_model_input_time(t_continuous)
+            cond_grad = cond_grad_fn(x, t_input)
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            noise = noise_pred_fn(x, t_continuous)
+            return noise - guidance_scale * expand_dims(sigma_t, x.dim()) * cond_grad
+        elif guidance_tp == "classifier-free":
+            if (
+                guidance_scale == 1.0
+                or unconditional_condition is None
+                or not (interval_guidance[0] < t_continuous[0] < interval_guidance[1])
+            ):
+                return noise_pred_fn(x, t_continuous, cond=condition)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t_continuous] * 2)
+                c_in = torch.cat([unconditional_condition, condition])
+                try:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                except:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
+                        2
+                    )
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+        elif guidance_tp == "classifier-free_PAG":
+            for i in pag_applied_layers:
+                if isinstance(model, torch.nn.Module):
+                    model.blocks[i].attn.forward = (
+                        PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
+                        if guidance_scale == 1.0
+                        else PAGCFGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
+                    )
+                else:
+                    model.__self__.blocks[i].attn.forward = (
+                        PAGIdentitySelfAttnProcessorLiteLA(
+                            model.__self__.blocks[i].attn
+                        )
+                        if guidance_scale == 1.0
+                        else PAGCFGIdentitySelfAttnProcessorLiteLA(
+                            model.__self__.blocks[i].attn
+                        )
+                    )
+            num_inputs = 2 if guidance_scale == 1.0 else 3
+            x_in = torch.cat([x] * num_inputs)
+            t_in = torch.cat([t_continuous] * num_inputs)
+            c_in = torch.cat(
+                [condition, condition]
+                if guidance_scale == 1.0
+                else [unconditional_condition, condition, condition]
+            )
+
+            try:
+                chunks = noise_pred_fn(x_in, t_in, cond=c_in).chunk(num_inputs)
+            except:
+                chunks = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(num_inputs)
+
+            if guidance_scale == 1.0:
+                noise, noise_perturb = chunks
+                noise_pred = noise + pag_scale * (noise - noise_perturb)
+            else:
+                noise_uncond, noise, noise_perturb = chunks
+                noise_pred = (
+                    noise_uncond
+                    + guidance_scale * (noise - noise_uncond)
+                    + pag_scale * (noise - noise_perturb)
+                )
+            for i in pag_applied_layers:
+                if isinstance(model, torch.nn.Module):
+                    model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                        model.blocks[i].attn
+                    )
+                else:
+                    model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                        model.__self__.blocks[i].attn
+                    )
+
+            return noise_pred
+        elif guidance_tp == "classifier-free_PAG_seq":
+            num_inputs = 2
+            if t_continuous[0] < 0.5:
+                # cfg
+                if (
+                    guidance_scale == 1.0
+                    or unconditional_condition is None
+                    or not (
+                        interval_guidance[0] < t_continuous[0] < interval_guidance[1]
+                    )
+                ):
+                    return noise_pred_fn(x, t_continuous, cond=condition)
+
+                x_in = torch.cat([x] * num_inputs)
+                t_in = torch.cat([t_continuous] * num_inputs)
+                c_in = torch.cat([unconditional_condition, condition])
+
+                try:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                except:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
+                        num_inputs
+                    )
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+            else:
+                # pag
+                for i in pag_applied_layers:
+                    if isinstance(model, torch.nn.Module):
+                        model.blocks[i].attn.forward = (
+                            PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
+                            if guidance_scale == 1.0
+                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
+                                model.blocks[i].attn
+                            )
+                        )
+                    else:
+                        model.__self__.blocks[i].attn.forward = (
+                            PAGIdentitySelfAttnProcessorLiteLA(
+                                model.__self__.blocks[i].attn
+                            )
+                            if guidance_scale == 1.0
+                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
+                                model.__self__.blocks[i].attn
+                            )
+                        )
+                x_in = torch.cat([x] * 3)
+                t_in = torch.cat([t_continuous] * 3)
+                c_in = torch.cat([unconditional_condition, condition, condition])
+
+                try:
+                    noise_uncond, noise, noise_perturb = noise_pred_fn(
+                        x_in, t_in, cond=c_in
+                    ).chunk(3)
+                except:
+                    noise_uncond, noise, noise_perturb = noise_pred_fn(
+                        x_in, t_in, cond=c_in
+                    )[0].chunk(3)
+
+                for i in pag_applied_layers:
+                    if isinstance(model, torch.nn.Module):
+                        model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                            model.blocks[i].attn
+                        )
+                    else:
+                        model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                            model.__self__.blocks[i].attn
+                        )
+
+                return (
+                    noise_uncond
+                    + guidance_scale * (noise - noise_uncond)
+                    + pag_scale * (noise - noise_perturb)
+                )
+
+    assert model_type in ["noise", "x_start", "v", "score", "flow"]
+    assert guidance_type in [
+        "uncond",
+        "classifier",
+        "classifier-free",
+        "classifier-free_PAG",
+        "classifier-free_PAG_seq",
+    ]
+    return model_fn
+
+
+class DPM_Solver:
+    def __init__(
+        self,
+        model_fn,
+        noise_schedule,
+        algorithm_type="dpmsolver++",
+        correcting_x0_fn=None,
+        correcting_xt_fn=None,
+        thresholding_max_val=1.0,
+        dynamic_thresholding_ratio=0.995,
+    ):
+        """Construct a DPM-Solver.
+
+        We support both DPM-Solver (`algorithm_type="dpmsolver"`) and DPM-Solver++ (`algorithm_type="dpmsolver++"`).
+
+        We also support the "dynamic thresholding" method in Imagen[1]. For pixel-space diffusion models, you
+        can set both `algorithm_type="dpmsolver++"` and `correcting_x0_fn="dynamic_thresholding"` to use the
+        dynamic thresholding. The "dynamic thresholding" can greatly improve the sample quality for pixel-space
+        DPMs with large guidance scales. Note that the thresholding method is **unsuitable** for latent-space
+        DPMs (such as stable-diffusion).
+
+        To support advanced algorithms in image-to-image applications, we also support corrector functions for
+        both x0 and xt.
+
+        Args:
+            model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
+                ``
+                def model_fn(x, t_continuous):
+                    return noise
+                ``
+                The shape of `x` is `(batch_size, **shape)`, and the shape of `t_continuous` is `(batch_size,)`.
+            noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+            algorithm_type: A `str`. Either "dpmsolver" or "dpmsolver++".
+            correcting_x0_fn: A `str` or a function with the following format:
+                ```
+                def correcting_x0_fn(x0, t):
+                    x0_new = ...
+                    return x0_new
+                ```
+                This function is to correct the outputs of the data prediction model at each sampling step. e.g.,
+                ```
+                x0_pred = data_pred_model(xt, t)
+                if correcting_x0_fn is not None:
+                    x0_pred = correcting_x0_fn(x0_pred, t)
+                xt_1 = update(x0_pred, xt, t)
+                ```
+                If `correcting_x0_fn="dynamic_thresholding"`, we use the dynamic thresholding proposed in Imagen[1].
+            correcting_xt_fn: A function with the following format:
+                ```
+                def correcting_xt_fn(xt, t, step):
+                    x_new = ...
+                    return x_new
+                ```
+                This function is to correct the intermediate samples xt at each sampling step. e.g.,
+                ```
+                xt = ...
+                xt = correcting_xt_fn(xt, t, step)
+                ```
+            thresholding_max_val: A `float`. The max value for thresholding.
+                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
+            dynamic_thresholding_ratio: A `float`. The ratio for dynamic thresholding (see Imagen[1] for details).
+                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
+
+        [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour,
+            Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models
+            with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
+        """
+        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
+        self.noise_schedule = noise_schedule
+        assert algorithm_type in ["dpmsolver", "dpmsolver++"]
+        self.algorithm_type = algorithm_type
+        if correcting_x0_fn == "dynamic_thresholding":
+            self.correcting_x0_fn = self.dynamic_thresholding_fn
+        else:
+            self.correcting_x0_fn = correcting_x0_fn
+        self.correcting_xt_fn = correcting_xt_fn
+        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
+        self.thresholding_max_val = thresholding_max_val
+        self.register_progress_bar()
+
+    def register_progress_bar(self, progress_fn=None):
+        """
+        Register a progress bar callback function
+
+        Args:
+            progress_fn: Callback function that takes current step and total steps as parameters
+        """
+        self.progress_fn = (
+            progress_fn if progress_fn is not None else lambda step, total: None
+        )
+
+    def update_progress(self, step, total_steps):
+        """
+        Update sampling progress
+
+        Args:
+            step: Current step number
+            total_steps: Total number of steps
+        """
+        if hasattr(self, "progress_fn"):
+            try:
+                self.progress_fn(
+                    step / total_steps, desc=f"Generating {step}/{total_steps}"
+                )
+            except:
+                self.progress_fn(step, total_steps)
+
+        else:
+            # If no progress_fn registered, use default empty function
+            pass
+
+    def dynamic_thresholding_fn(self, x0, t):
+        """
+        The dynamic thresholding method.
+        """
+        dims = x0.dim()
+        p = self.dynamic_thresholding_ratio
+        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
+        s = expand_dims(
+            torch.maximum(
+                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
+            ),
+            dims,
+        )
+        x0 = torch.clamp(x0, -s, s) / s
+        return x0
+
+    def noise_prediction_fn(self, x, t):
+        """
+        Return the noise prediction model.
+        """
+        return self.model(x, t)
+
+    def data_prediction_fn(self, x, t):
+        """
+        Return the data prediction model (with corrector).
+        """
+        noise = self.noise_prediction_fn(x, t)
+        alpha_t, sigma_t = (
+            self.noise_schedule.marginal_alpha(t),
+            self.noise_schedule.marginal_std(t),
+        )
+        x0 = (x - sigma_t * noise) / alpha_t
+        if self.correcting_x0_fn is not None:
+            x0 = self.correcting_x0_fn(x0, t)
+        return x0
+
+    def model_fn(self, x, t):
+        """
+        Convert the model to the noise prediction model or the data prediction model.
+        """
+        if self.algorithm_type == "dpmsolver++":
+            return self.data_prediction_fn(x, t)
+        else:
+            return self.noise_prediction_fn(x, t)
+
+    def get_time_steps(self, skip_type, t_T, t_0, N, device, shift=1.0):
+        """Compute the intermediate time steps for sampling.
+
+        Args:
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            N: A `int`. The total number of the spacing of the time steps.
+            device: A torch device.
+        Returns:
+            A pytorch tensor of the time steps, with the shape (N + 1,).
+        """
+        if skip_type == "logSNR":
+            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
+            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
+            logSNR_steps = torch.linspace(
+                lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1
+            ).to(device)
+            return self.noise_schedule.inverse_lambda(logSNR_steps)
+        elif skip_type == "time_uniform":
+            return torch.linspace(t_T, t_0, N + 1).to(device)
+        elif skip_type == "time_quadratic":
+            t_order = 2
+            t = (
+                torch.linspace(t_T ** (1.0 / t_order), t_0 ** (1.0 / t_order), N + 1)
+                .pow(t_order)
+                .to(device)
+            )
+            return t
+        elif skip_type == "time_uniform_flow":
+            betas = torch.linspace(t_T, t_0, N + 1).to(device)
+            sigmas = 1.0 - betas
+            sigmas = (shift * sigmas / (1 + (shift - 1) * sigmas)).flip(dims=[0])
+            return sigmas
+        else:
+            raise ValueError(
+                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'"
+            )
+
+    def get_orders_and_timesteps_for_singlestep_solver(
+        self, steps, order, skip_type, t_T, t_0, device
+    ):
+        """
+        Get the order of each step for sampling by the singlestep DPM-Solver.
+
+        We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
+        Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
+            - If order == 1:
+                We take `steps` of DPM-Solver-1 (i.e. DDIM).
+            - If order == 2:
+                - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
+                - If steps % 2 == 0, we use K steps of DPM-Solver-2.
+                - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
+            - If order == 3:
+                - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
+
+        ============================================
+        Args:
+            order: A `int`. The max order for the solver (2 or 3).
+            steps: A `int`. The total number of function evaluations (NFE).
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            device: A torch device.
+        Returns:
+            orders: A list of the solver order of each step.
+        """
+        if order == 3:
+            K = steps // 3 + 1
+            if steps % 3 == 0:
+                orders = [
+                    3,
+                ] * (K - 2) + [2, 1]
+            elif steps % 3 == 1:
+                orders = [
+                    3,
+                ] * (K - 1) + [1]
+            else:
+                orders = [
+                    3,
+                ] * (K - 1) + [2]
+        elif order == 2:
+            if steps % 2 == 0:
+                K = steps // 2
+                orders = [
+                    2,
+                ] * K
+            else:
+                K = steps // 2 + 1
+                orders = [
+                    2,
+                ] * (K - 1) + [1]
+        elif order == 1:
+            K = 1
+            orders = [
+                1,
+            ] * steps
+        else:
+            raise ValueError("'order' must be '1' or '2' or '3'.")
+        if skip_type == "logSNR":
+            # To reproduce the results in DPM-Solver paper
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
+        else:
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
+                torch.cumsum(
+                    torch.tensor(
+                        [
+                            0,
+                        ]
+                        + orders
+                    ),
+                    0,
+                ).to(device)
+            ]
+        return timesteps_outer, orders
+
+    def denoise_to_zero_fn(self, x, s):
+        """
+        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
+        """
+        return self.data_prediction_fn(x, s)
+
+    def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
+        """
+        DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        dims = x.dim()
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        log_alpha_s, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = sigma_t / sigma_s * x - alpha_t * phi_1 * model_s
+            if return_intermediate:
+                return x_t, {"model_s": model_s}
+            else:
+                return x_t
+        else:
+            phi_1 = torch.expm1(h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = torch.exp(log_alpha_t - log_alpha_s) * x - (sigma_t * phi_1) * model_s
+            if return_intermediate:
+                return x_t, {"model_s": model_s}
+            else:
+                return x_t
+
+    def singlestep_dpm_solver_second_update(
+        self,
+        x,
+        s,
+        t,
+        r1=0.5,
+        model_s=None,
+        return_intermediate=False,
+        solver_type="dpmsolver",
+    ):
+        """
+        Singlestep solver DPM-Solver-2 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            r1: A `float`. The hyperparameter of the second-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(
+                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
+            )
+        if r1 is None:
+            r1 = 0.5
+        ns = self.noise_schedule
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        log_alpha_s, log_alpha_s1, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(s1),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_s1, sigma_t = (
+            ns.marginal_std(s),
+            ns.marginal_std(s1),
+            ns.marginal_std(t),
+        )
+        alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_11 = torch.expm1(-r1 * h)
+            phi_1 = torch.expm1(-h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    - (0.5 / r1) * (alpha_t * phi_1) * (model_s1 - model_s)
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (1.0 / r1) * (alpha_t * (phi_1 / h + 1.0)) * (model_s1 - model_s)
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_1 = torch.expm1(h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = (
+                torch.exp(log_alpha_s1 - log_alpha_s) * x
+                - (sigma_s1 * phi_11) * model_s
+            )
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    torch.exp(log_alpha_t - log_alpha_s) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (0.5 / r1) * (sigma_t * phi_1) * (model_s1 - model_s)
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    torch.exp(log_alpha_t - log_alpha_s) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (1.0 / r1) * (sigma_t * (phi_1 / h - 1.0)) * (model_s1 - model_s)
+                )
+        if return_intermediate:
+            return x_t, {"model_s": model_s, "model_s1": model_s1}
+        else:
+            return x_t
+
+    def singlestep_dpm_solver_third_update(
+        self,
+        x,
+        s,
+        t,
+        r1=1.0 / 3.0,
+        r2=2.0 / 3.0,
+        model_s=None,
+        model_s1=None,
+        return_intermediate=False,
+        solver_type="dpmsolver",
+    ):
+        """
+        Singlestep solver DPM-Solver-3 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            r1: A `float`. The hyperparameter of the third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
+                If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(
+                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
+            )
+        if r1 is None:
+            r1 = 1.0 / 3.0
+        if r2 is None:
+            r2 = 2.0 / 3.0
+        ns = self.noise_schedule
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        lambda_s2 = lambda_s + r2 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        s2 = ns.inverse_lambda(lambda_s2)
+        log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(s1),
+            ns.marginal_log_mean_coeff(s2),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_s1, sigma_s2, sigma_t = (
+            ns.marginal_std(s),
+            ns.marginal_std(s1),
+            ns.marginal_std(s2),
+            ns.marginal_std(t),
+        )
+        alpha_s1, alpha_s2, alpha_t = (
+            torch.exp(log_alpha_s1),
+            torch.exp(log_alpha_s2),
+            torch.exp(log_alpha_t),
+        )
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_11 = torch.expm1(-r1 * h)
+            phi_12 = torch.expm1(-r2 * h)
+            phi_1 = torch.expm1(-h)
+            phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.0
+            phi_2 = phi_1 / h + 1.0
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                (sigma_s2 / sigma_s) * x
+                - (alpha_s2 * phi_12) * model_s
+                + r2 / r1 * (alpha_s2 * phi_22) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (1.0 / r2) * (alpha_t * phi_2) * (model_s2 - model_s)
+                )
+            elif solver_type == "taylor":
+                D1_0 = (1.0 / r1) * (model_s1 - model_s)
+                D1_1 = (1.0 / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (alpha_t * phi_2) * D1
+                    - (alpha_t * phi_3) * D2
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_12 = torch.expm1(r2 * h)
+            phi_1 = torch.expm1(h)
+            phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.0
+            phi_2 = phi_1 / h - 1.0
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (torch.exp(log_alpha_s1 - log_alpha_s)) * x - (
+                    sigma_s1 * phi_11
+                ) * model_s
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                (torch.exp(log_alpha_s2 - log_alpha_s)) * x
+                - (sigma_s2 * phi_12) * model_s
+                - r2 / r1 * (sigma_s2 * phi_22) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_s)) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (1.0 / r2) * (sigma_t * phi_2) * (model_s2 - model_s)
+                )
+            elif solver_type == "taylor":
+                D1_0 = (1.0 / r1) * (model_s1 - model_s)
+                D1_1 = (1.0 / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_s)) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (sigma_t * phi_2) * D1
+                    - (sigma_t * phi_3) * D2
+                )
+
+        if return_intermediate:
+            return x_t, {"model_s": model_s, "model_s1": model_s1, "model_s2": model_s2}
+        else:
+            return x_t
+
+    def multistep_dpm_solver_second_update(
+        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
+    ):
+        """
+        Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(
+                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
+            )
+        ns = self.noise_schedule
+        model_prev_1, model_prev_0 = model_prev_list[-2], model_prev_list[-1]
+        t_prev_1, t_prev_0 = t_prev_list[-2], t_prev_list[-1]
+        lambda_prev_1, lambda_prev_0, lambda_t = (
+            ns.marginal_lambda(t_prev_1),
+            ns.marginal_lambda(t_prev_0),
+            ns.marginal_lambda(t),
+        )
+        log_alpha_prev_0, log_alpha_t = (
+            ns.marginal_log_mean_coeff(t_prev_0),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0 = h_0 / h
+        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_prev_0) * x
+                    - (alpha_t * phi_1) * model_prev_0
+                    - 0.5 * (alpha_t * phi_1) * D1_0
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (sigma_t / sigma_prev_0) * x
+                    - (alpha_t * phi_1) * model_prev_0
+                    + (alpha_t * (phi_1 / h + 1.0)) * D1_0
+                )
+        else:
+            phi_1 = torch.expm1(h)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                    - (sigma_t * phi_1) * model_prev_0
+                    - 0.5 * (sigma_t * phi_1) * D1_0
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                    - (sigma_t * phi_1) * model_prev_0
+                    - (sigma_t * (phi_1 / h - 1.0)) * D1_0
+                )
+        return x_t
+
+    def multistep_dpm_solver_third_update(
+        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
+    ):
+        """
+        Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        model_prev_2, model_prev_1, model_prev_0 = model_prev_list
+        t_prev_2, t_prev_1, t_prev_0 = t_prev_list
+        lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = (
+            ns.marginal_lambda(t_prev_2),
+            ns.marginal_lambda(t_prev_1),
+            ns.marginal_lambda(t_prev_0),
+            ns.marginal_lambda(t),
+        )
+        log_alpha_prev_0, log_alpha_t = (
+            ns.marginal_log_mean_coeff(t_prev_0),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_1 = lambda_prev_1 - lambda_prev_2
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0, r1 = h_0 / h, h_1 / h
+        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
+        D1_1 = (1.0 / r1) * (model_prev_1 - model_prev_2)
+        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            phi_2 = phi_1 / h + 1.0
+            phi_3 = phi_2 / h - 0.5
+            x_t = (
+                (sigma_t / sigma_prev_0) * x
+                - (alpha_t * phi_1) * model_prev_0
+                + (alpha_t * phi_2) * D1
+                - (alpha_t * phi_3) * D2
+            )
+        else:
+            phi_1 = torch.expm1(h)
+            phi_2 = phi_1 / h - 1.0
+            phi_3 = phi_2 / h - 0.5
+            x_t = (
+                (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                - (sigma_t * phi_1) * model_prev_0
+                - (sigma_t * phi_2) * D1
+                - (sigma_t * phi_3) * D2
+            )
+        return x_t
+
+    def singlestep_dpm_solver_update(
+        self,
+        x,
+        s,
+        t,
+        order,
+        return_intermediate=False,
+        solver_type="dpmsolver",
+        r1=None,
+        r2=None,
+    ):
+        """
+        Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+            r1: A `float`. The hyperparameter of the second-order or third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(
+                x, s, t, return_intermediate=return_intermediate
+            )
+        elif order == 2:
+            return self.singlestep_dpm_solver_second_update(
+                x,
+                s,
+                t,
+                return_intermediate=return_intermediate,
+                solver_type=solver_type,
+                r1=r1,
+            )
+        elif order == 3:
+            return self.singlestep_dpm_solver_third_update(
+                x,
+                s,
+                t,
+                return_intermediate=return_intermediate,
+                solver_type=solver_type,
+                r1=r1,
+                r2=r2,
+            )
+        else:
+            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
+
+    def multistep_dpm_solver_update(
+        self, x, model_prev_list, t_prev_list, t, order, solver_type="dpmsolver"
+    ):
+        """
+        Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(
+                x, t_prev_list[-1], t, model_s=model_prev_list[-1]
+            )
+        elif order == 2:
+            return self.multistep_dpm_solver_second_update(
+                x, model_prev_list, t_prev_list, t, solver_type=solver_type
+            )
+        elif order == 3:
+            return self.multistep_dpm_solver_third_update(
+                x, model_prev_list, t_prev_list, t, solver_type=solver_type
+            )
+        else:
+            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
+
+    def dpm_solver_adaptive(
+        self,
+        x,
+        order,
+        t_T,
+        t_0,
+        h_init=0.05,
+        atol=0.0078,
+        rtol=0.05,
+        theta=0.9,
+        t_err=1e-5,
+        solver_type="dpmsolver",
+    ):
+        """
+        The adaptive step size solver based on singlestep DPM-Solver.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `t_T`.
+            order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            h_init: A `float`. The initial step size (for logSNR).
+            atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
+            rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
+            theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
+            t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
+                current time and `t_0` is less than `t_err`. The default setting is 1e-5.
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_0: A pytorch tensor. The approximated solution at time `t_0`.
+
+        [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
+        """
+        ns = self.noise_schedule
+        s = t_T * torch.ones((1,)).to(x)
+        lambda_s = ns.marginal_lambda(s)
+        lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
+        h = h_init * torch.ones_like(s).to(x)
+        x_prev = x
+        nfe = 0
+        if order == 2:
+            r1 = 0.5
+            lower_update = lambda x, s, t: self.dpm_solver_first_update(
+                x, s, t, return_intermediate=True
+            )
+            higher_update = lambda x, s, t, **kwargs: (
+                self.singlestep_dpm_solver_second_update(
+                    x, s, t, r1=r1, solver_type=solver_type, **kwargs
+                )
+            )
+        elif order == 3:
+            r1, r2 = 1.0 / 3.0, 2.0 / 3.0
+            lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(
+                x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type
+            )
+            higher_update = lambda x, s, t, **kwargs: (
+                self.singlestep_dpm_solver_third_update(
+                    x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs
+                )
+            )
+        else:
+            raise ValueError(
+                f"For adaptive step size solver, order must be 2 or 3, got {order}"
+            )
+        while torch.abs(s - t_0).mean() > t_err:
+            t = ns.inverse_lambda(lambda_s + h)
+            x_lower, lower_noise_kwargs = lower_update(x, s, t)
+            x_higher = higher_update(x, s, t, **lower_noise_kwargs)
+            delta = torch.max(
+                torch.ones_like(x).to(x) * atol,
+                rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)),
+            )
+            norm_fn = lambda v: torch.sqrt(
+                torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)
+            )
+            E = norm_fn((x_higher - x_lower) / delta).max()
+            if torch.all(E <= 1.0):
+                x = x_higher
+                s = t
+                x_prev = x_lower
+                lambda_s = ns.marginal_lambda(s)
+            h = torch.min(
+                theta * h * torch.float_power(E, -1.0 / order).float(),
+                lambda_0 - lambda_s,
+            )
+            nfe += order
+        print("adaptive solver nfe", nfe)
+        return x
+
+    def add_noise(self, x, t, noise=None):
+        """
+        Compute the noised input xt = alpha_t * x + sigma_t * noise.
+
+        Args:
+            x: A `torch.Tensor` with shape `(batch_size, *shape)`.
+            t: A `torch.Tensor` with shape `(t_size,)`.
+        Returns:
+            xt with shape `(t_size, batch_size, *shape)`.
+        """
+        alpha_t, sigma_t = (
+            self.noise_schedule.marginal_alpha(t),
+            self.noise_schedule.marginal_std(t),
+        )
+        if noise is None:
+            noise = torch.randn((t.shape[0], *x.shape), device=x.device)
+        x = x.reshape((-1, *x.shape))
+        xt = expand_dims(alpha_t, x.dim()) * x + expand_dims(sigma_t, x.dim()) * noise
+        if t.shape[0] == 1:
+            return xt.squeeze(0)
+        else:
+            return xt
+
+    def inverse(
+        self,
+        x,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        order=2,
+        skip_type="time_uniform",
+        method="multistep",
+        lower_order_final=True,
+        denoise_to_zero=False,
+        solver_type="dpmsolver",
+        atol=0.0078,
+        rtol=0.05,
+        return_intermediate=False,
+    ):
+        """
+        Inverse the sample `x` from time `t_start` to `t_end` by DPM-Solver.
+        For discrete-time DPMs, we use `t_start=1/N`, where `N` is the total time steps during training.
+        """
+        t_0 = 1.0 / self.noise_schedule.total_N if t_start is None else t_start
+        t_T = self.noise_schedule.T if t_end is None else t_end
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+        return self.sample(
+            x,
+            steps=steps,
+            t_start=t_0,
+            t_end=t_T,
+            order=order,
+            skip_type=skip_type,
+            method=method,
+            lower_order_final=lower_order_final,
+            denoise_to_zero=denoise_to_zero,
+            solver_type=solver_type,
+            atol=atol,
+            rtol=rtol,
+            return_intermediate=return_intermediate,
+        )
+
+    def sample(
+        self,
+        x,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        order=2,
+        skip_type="time_uniform",
+        method="multistep",
+        lower_order_final=True,
+        denoise_to_zero=False,
+        solver_type="dpmsolver",
+        atol=0.0078,
+        rtol=0.05,
+        return_intermediate=False,
+        flow_shift=1.0,
+    ):
+        """
+        Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
+
+        =====================================================
+
+        We support the following algorithms for both noise prediction model and data prediction model:
+            - 'singlestep':
+                Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
+                We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
+                The total number of function evaluations (NFE) == `steps`.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    - If `order` == 1:
+                        - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
+                        - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
+                        - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                    - If `order` == 3:
+                        - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                        - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
+            - 'multistep':
+                Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
+                We initialize the first `order` values by lower order multistep solvers.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    Denote K = steps.
+                    - If `order` == 1:
+                        - We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
+                    - If `order` == 3:
+                        - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
+            - 'singlestep_fixed':
+                Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
+                We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
+            - 'adaptive':
+                Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
+                We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
+                You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
+                (NFE) and the sample quality.
+                    - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
+                    - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
+
+        =====================================================
+
+        Some advices for choosing the algorithm:
+            - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
+                Use singlestep DPM-Solver or DPM-Solver++ ("DPM-Solver-fast" in the paper) with `order = 3`.
+                e.g., DPM-Solver:
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
+                            skip_type='time_uniform', method='singlestep')
+                e.g., DPM-Solver++:
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
+                            skip_type='time_uniform', method='singlestep')
+            - For **guided sampling with large guidance scale** by DPMs:
+                Use multistep DPM-Solver with `algorithm_type="dpmsolver++"` and `order = 2`.
+                e.g.
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
+                            skip_type='time_uniform', method='multistep')
+
+        We support three types of `skip_type`:
+            - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
+            - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
+            - 'time_quadratic': quadratic time for the time steps.
+
+        =====================================================
+        Args:
+            x: A pytorch tensor. The initial value at time `t_start`
+                e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
+            steps: A `int`. The total number of function evaluations (NFE).
+            t_start: A `float`. The starting time of the sampling.
+                If `T` is None, we use self.noise_schedule.T (default is 1.0).
+            t_end: A `float`. The ending time of the sampling.
+                If `t_end` is None, we use 1. / self.noise_schedule.total_N.
+                e.g. if total_N == 1000, we have `t_end` == 1e-3.
+                For discrete-time DPMs:
+                    - We recommend `t_end` == 1. / self.noise_schedule.total_N.
+                For continuous-time DPMs:
+                    - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
+            order: A `int`. The order of DPM-Solver.
+            skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
+            method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
+            denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
+                Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
+
+                This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
+                score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
+                for diffusion models sampling by diffusion SDEs for low-resolutional images
+                (such as CIFAR-10). However, we observed that such trick does not matter for
+                high-resolutional images. As it needs an additional NFE, we do not recommend
+                it for high-resolutional images.
+            lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
+                Only valid for `method=multistep` and `steps < 15`. We empirically find that
+                this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
+                (especially for steps <= 10). So we recommend to set it to be `True`.
+            solver_type: A `str`. The taylor expansion type for the solver. `dpmsolver` or `taylor`. We recommend `dpmsolver`.
+            atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+            rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+            return_intermediate: A `bool`. Whether to save the xt at each step.
+                When set to `True`, method returns a tuple (x0, intermediates); when set to False, method returns only x0.
+        Returns:
+            x_end: A pytorch tensor. The approximated solution at time `t_end`.
+
+        """
+        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+        if return_intermediate:
+            assert method in [
+                "multistep",
+                "singlestep",
+                "singlestep_fixed",
+            ], "Cannot use adaptive solver when saving intermediate values"
+        if self.correcting_xt_fn is not None:
+            assert method in [
+                "multistep",
+                "singlestep",
+                "singlestep_fixed",
+            ], "Cannot use adaptive solver when correcting_xt_fn is not None"
+        device = x.device
+        intermediates = []
+        with torch.no_grad():
+            if method == "adaptive":
+                x = self.dpm_solver_adaptive(
+                    x,
+                    order=order,
+                    t_T=t_T,
+                    t_0=t_0,
+                    atol=atol,
+                    rtol=rtol,
+                    solver_type=solver_type,
+                )
+            elif method == "multistep":
+                assert steps >= order
+                timesteps = self.get_time_steps(
+                    skip_type=skip_type,
+                    t_T=t_T,
+                    t_0=t_0,
+                    N=steps,
+                    device=device,
+                    shift=flow_shift,
+                )
+                assert timesteps.shape[0] - 1 == steps
+                # Init the initial values.
+                step = 0
+                t = timesteps[step]
+                t_prev_list = [t]
+                model_prev_list = [self.model_fn(x, t)]
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+                self.update_progress(step + 1, len(timesteps))
+                # Init the first `order` values by lower order multistep DPM-Solver.
+                for step in range(1, order):
+                    t = timesteps[step]
+                    x = self.multistep_dpm_solver_update(
+                        x,
+                        model_prev_list,
+                        t_prev_list,
+                        t,
+                        step,
+                        solver_type=solver_type,
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    t_prev_list.append(t)
+                    model_prev_list.append(self.model_fn(x, t))
+                    # update progress bar
+                    self.update_progress(step + 1, len(timesteps))
+                # Compute the remaining values by `order`-th order multistep DPM-Solver.
+                for step in tqdm(
+                    range(order, steps + 1),
+                    disable=os.getenv("DPM_TQDM", "False") == "True",
+                ):
+                    t = timesteps[step]
+                    # We only use lower order for steps < 10
+                    # if lower_order_final and steps < 10:
+                    if lower_order_final:  # recommended by Shuchen Xue
+                        step_order = min(order, steps + 1 - step)
+                    else:
+                        step_order = order
+                    x = self.multistep_dpm_solver_update(
+                        x,
+                        model_prev_list,
+                        t_prev_list,
+                        t,
+                        step_order,
+                        solver_type=solver_type,
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    for i in range(order - 1):
+                        t_prev_list[i] = t_prev_list[i + 1]
+                        model_prev_list[i] = model_prev_list[i + 1]
+                    t_prev_list[-1] = t
+                    # We do not need to evaluate the final model value.
+                    if step < steps:
+                        model_prev_list[-1] = self.model_fn(x, t)
+                    # update progress bar
+                    self.update_progress(step + 1, len(timesteps))
+            elif method in ["singlestep", "singlestep_fixed"]:
+                if method == "singlestep":
+                    timesteps_outer, orders = (
+                        self.get_orders_and_timesteps_for_singlestep_solver(
+                            steps=steps,
+                            order=order,
+                            skip_type=skip_type,
+                            t_T=t_T,
+                            t_0=t_0,
+                            device=device,
+                        )
+                    )
+                elif method == "singlestep_fixed":
+                    K = steps // order
+                    orders = [
+                        order,
+                    ] * K
+                    timesteps_outer = self.get_time_steps(
+                        skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device
+                    )
+                for step, order in enumerate(orders):
+                    s, t = timesteps_outer[step], timesteps_outer[step + 1]
+                    timesteps_inner = self.get_time_steps(
+                        skip_type=skip_type,
+                        t_T=s.item(),
+                        t_0=t.item(),
+                        N=order,
+                        device=device,
+                    )
+                    lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
+                    h = lambda_inner[-1] - lambda_inner[0]
+                    r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
+                    r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
+                    x = self.singlestep_dpm_solver_update(
+                        x, s, t, order, solver_type=solver_type, r1=r1, r2=r2
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    self.update_progress(step + 1, len(timesteps_outer))
+            else:
+                raise ValueError(f"Got wrong method {method}")
+            if denoise_to_zero:
+                t = torch.ones((1,)).to(device) * t_0
+                x = self.denoise_to_zero_fn(x, t)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step + 1)
+                if return_intermediate:
+                    intermediates.append(x)
+        if return_intermediate:
+            return x, intermediates
+        else:
+            return x
+
+
+#############################################################
+# other utility functions
+#############################################################
+
+
+def interpolate_fn(x, xp, yp):
+    """
+    A piecewise linear function y = f(x), using xp and yp as keypoints.
+    We implement f(x) in a differentiable way (i.e. applicable for autograd).
+    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
+
+    Args:
+        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
+        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
+        yp: PyTorch tensor with shape [C, K].
+    Returns:
+        The function values f(x), with shape [N, C].
+    """
+    N, K = x.shape[0], xp.shape[1]
+    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
+    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
+    x_idx = torch.argmin(x_indices, dim=2)
+    cand_start_idx = x_idx - 1
+    start_idx = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(1, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    end_idx = torch.where(
+        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
+    )
+    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
+    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
+    start_idx2 = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(0, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
+    start_y = torch.gather(
+        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
+    ).squeeze(2)
+    end_y = torch.gather(
+        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
+    ).squeeze(2)
+    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
+    return cand
+
+
+def expand_dims(v, dims):
+    """
+    Expand the tensor `v` to the dim `dims`.
+
+    Args:
+        `v`: a PyTorch tensor with shape [N].
+        `dim`: a `int`.
+    Returns:
+        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
+    """
+    return v[(...,) + (None,) * (dims - 1)]
diff --git a/image/sana/sana-1600m/packages/Sana/diffusion/model/edm_sample.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/edm_sample.py
new file mode 100755
index 000000000..77d44a40d
--- /dev/null
+++ b/image/sana/sana-1600m/packages/Sana/diffusion/model/edm_sample.py
@@ -0,0 +1,276 @@
+# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# Modified from OpenAI's diffusion repos
+#     GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py
+#     ADM:   https://github.com/openai/guided-diffusion/blob/main/guided_diffusion
+#     IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+
+
+import numpy as np
+from diffusion.model.utils import *
+from tqdm import tqdm
+
+# ----------------------------------------------------------------------------
+# Proposed EDM sampler (Algorithm 2).
+
+
+def edm_sampler(
+    net,
+    latents,
+    class_labels=None,
+    cfg_scale=None,
+    randn_like=torch.randn_like,
+    num_steps=18,
+    sigma_min=0.002,
+    sigma_max=80,
+    rho=7,
+    S_churn=0,
+    S_min=0,
+    S_max=float("inf"),
+    S_noise=1,
+    **kwargs,
+):
+    # Adjust noise levels based on what's supported by the network.
+    sigma_min = max(sigma_min, net.sigma_min)
+    sigma_max = min(sigma_max, net.sigma_max)
+
+    # Time step discretization.
+    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
+    t_steps = (
+        sigma_max ** (1 / rho)
+        + step_indices
+        / (num_steps - 1)
+        * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
+    ) ** rho
+    t_steps = torch.cat(
+        [net.round_sigma(t_steps), torch.zeros_like(t_steps[:1])]
+    )  # t_N = 0
+
+    # Main sampling loop.
+    x_next = latents.to(torch.float64) * t_steps[0]
+    for i, (t_cur, t_next) in tqdm(
+        list(enumerate(zip(t_steps[:-1], t_steps[1:])))
+    ):  # 0, ..., N-1
+        x_cur = x_next
+
+        # Increase noise temporarily.
+        gamma = (
+            min(S_churn / num_steps, np.sqrt(2) - 1) if S_min <= t_cur <= S_max else 0
+        )
+        t_hat = net.round_sigma(t_cur + gamma * t_cur)
+        x_hat = x_cur + (t_hat**2 - t_cur**2).sqrt() * S_noise * randn_like(x_cur)
+
+        # Euler step.
+        denoised = net(x_hat.float(), t_hat, class_labels, cfg_scale, **kwargs)["x"].to(
+            torch.float64
+        )
+        d_cur = (x_hat - denoised) / t_hat
+        x_next = x_hat + (t_next - t_hat) * d_cur
+
+        # Apply 2nd order correction.
+        if i < num_steps - 1:
+            denoised = net(x_next.float(), t_next, class_labels, cfg_scale, **kwargs)[
+                "x"
+            ].to(torch.float64)
+            d_prime = (x_next - denoised) / t_next
+            x_next = x_hat + (t_next - t_hat) * (0.5 * d_cur + 0.5 * d_prime)
+
+    return x_next
+
+
+# ----------------------------------------------------------------------------
+# Generalized ablation sampler, representing the superset of all sampling
+# methods discussed in the paper.
+
+
+def ablation_sampler(
+    net,
+    latents,
+    class_labels=None,
+    cfg_scale=None,
+    feat=None,
+    randn_like=torch.randn_like,
+    num_steps=18,
+    sigma_min=None,
+    sigma_max=None,
+    rho=7,
+    solver="heun",
+    discretization="edm",
+    schedule="linear",
+    scaling="none",
+    epsilon_s=1e-3,
+    C_1=0.001,
+    C_2=0.008,
+    M=1000,
+    alpha=1,
+    S_churn=0,
+    S_min=0,
+    S_max=float("inf"),
+    S_noise=1,
+):
+    assert solver in ["euler", "heun"]
+    assert discretization in ["vp", "ve", "iddpm", "edm"]
+    assert schedule in ["vp", "ve", "linear"]
+    assert scaling in ["vp", "none"]
+
+    # Helper functions for VP & VE noise level schedules.
+    vp_sigma = lambda beta_d, beta_min: (
+        lambda t: (np.e ** (0.5 * beta_d * (t**2) + beta_min * t) - 1) ** 0.5
+    )
+    vp_sigma_deriv = lambda beta_d, beta_min: (
+        lambda t: 0.5 * (beta_min + beta_d * t) * (sigma(t) + 1 / sigma(t))
+    )
+    vp_sigma_inv = lambda beta_d, beta_min: (
+        lambda sigma: (
+            ((beta_min**2 + 2 * beta_d * (sigma**2 + 1).log()).sqrt() - beta_min)
+            / beta_d
+        )
+    )
+    ve_sigma = lambda t: t.sqrt()
+    ve_sigma_deriv = lambda t: 0.5 / t.sqrt()
+    ve_sigma_inv = lambda sigma: sigma**2
+
+    # Select default noise level range based on the specified time step discretization.
+    if sigma_min is None:
+        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=epsilon_s)
+        sigma_min = {"vp": vp_def, "ve": 0.02, "iddpm": 0.002, "edm": 0.002}[
+            discretization
+        ]
+    if sigma_max is None:
+        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=1)
+        sigma_max = {"vp": vp_def, "ve": 100, "iddpm": 81, "edm": 80}[discretization]
+
+    # Adjust noise levels based on what's supported by the network.
+    sigma_min = max(sigma_min, net.sigma_min)
+    sigma_max = min(sigma_max, net.sigma_max)
+
+    # Compute corresponding betas for VP.
+    vp_beta_d = (
+        2
+        * (np.log(sigma_min**2 + 1) / epsilon_s - np.log(sigma_max**2 + 1))
+        / (epsilon_s - 1)
+    )
+    vp_beta_min = np.log(sigma_max**2 + 1) - 0.5 * vp_beta_d
+
+    # Define time steps in terms of noise level.
+    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
+    if discretization == "vp":
+        orig_t_steps = 1 + step_indices / (num_steps - 1) * (epsilon_s - 1)
+        sigma_steps = vp_sigma(vp_beta_d, vp_beta_min)(orig_t_steps)
+    elif discretization == "ve":
+        orig_t_steps = (sigma_max**2) * (
+            (sigma_min**2 / sigma_max**2) ** (step_indices / (num_steps - 1))
+        )
+        sigma_steps = ve_sigma(orig_t_steps)
+    elif discretization == "iddpm":
+        u = torch.zeros(M + 1, dtype=torch.float64, device=latents.device)
+        alpha_bar = lambda j: (0.5 * np.pi * j / M / (C_2 + 1)).sin() ** 2
+        for j in torch.arange(M, 0, -1, device=latents.device):  # M, ..., 1
+            u[j - 1] = (
+                (u[j] ** 2 + 1) / (alpha_bar(j - 1) / alpha_bar(j)).clip(min=C_1) - 1
+            ).sqrt()
+        u_filtered = u[torch.logical_and(u >= sigma_min, u <= sigma_max)]
+        sigma_steps = u_filtered[
+            ((len(u_filtered) - 1) / (num_steps - 1) * step_indices)
+            .round()
+            .to(torch.int64)
+        ]
+    else:
+        assert discretization == "edm"
+        sigma_steps = (
+            sigma_max ** (1 / rho)
+            + step_indices
+            / (num_steps - 1)
+            * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
+        ) ** rho
+
+    # Define noise level schedule.
+    if schedule == "vp":
+        sigma = vp_sigma(vp_beta_d, vp_beta_min)
+        sigma_deriv = vp_sigma_deriv(vp_beta_d, vp_beta_min)
+        sigma_inv = vp_sigma_inv(vp_beta_d, vp_beta_min)
+    elif schedule == "ve":
+        sigma = ve_sigma
+        sigma_deriv = ve_sigma_deriv
+        sigma_inv = ve_sigma_inv
+    else:
+        assert schedule == "linear"
+        sigma = lambda t: t
+        sigma_deriv = lambda t: 1
+        sigma_inv = lambda sigma: sigma
+
+    # Define scaling schedule.
+    if scaling == "vp":
+        s = lambda t: 1 / (1 + sigma(t) ** 2).sqrt()
+        s_deriv = lambda t: -sigma(t) * sigma_deriv(t) * (s(t) ** 3)
+    else:
+        assert scaling == "none"
+        s = lambda t: 1
+        s_deriv = lambda t: 0
+
+    # Compute final time steps based on the corresponding noise levels.
+    t_steps = sigma_inv(net.round_sigma(sigma_steps))
+    t_steps = torch.cat([t_steps, torch.zeros_like(t_steps[:1])])  # t_N = 0
+
+    # Main sampling loop.
+    t_next = t_steps[0]
+    x_next = latents.to(torch.float64) * (sigma(t_next) * s(t_next))
+    for i, (t_cur, t_next) in enumerate(zip(t_steps[:-1], t_steps[1:])):  # 0, ..., N-1
+        x_cur = x_next
+
+        # Increase noise temporarily.
+        gamma = (
+            min(S_churn / num_steps, np.sqrt(2) - 1)
+            if S_min <= sigma(t_cur) <= S_max
+            else 0
+        )
+        t_hat = sigma_inv(net.round_sigma(sigma(t_cur) + gamma * sigma(t_cur)))
+        x_hat = s(t_hat) / s(t_cur) * x_cur + (
+            sigma(t_hat) ** 2 - sigma(t_cur) ** 2
+        ).clip(min=0).sqrt() * s(t_hat) * S_noise * randn_like(x_cur)
+
+        # Euler step.
+        h = t_next - t_hat
+        denoised = net(
+            x_hat.float() / s(t_hat), sigma(t_hat), class_labels, cfg_scale, feat=feat
+        )["x"].to(torch.float64)
+        d_cur = (
+            sigma_deriv(t_hat) / sigma(t_hat) + s_deriv(t_hat) / s(t_hat)
+        ) * x_hat - sigma_deriv(t_hat) * s(t_hat) / sigma(t_hat) * denoised
+        x_prime = x_hat + alpha * h * d_cur
+        t_prime = t_hat + alpha * h
+
+        # Apply 2nd order correction.
+        if solver == "euler" or i == num_steps - 1:
+            x_next = x_hat + h * d_cur
+        else:
+            assert solver == "heun"
+            denoised = net(
+                x_prime.float() / s(t_prime),
+                sigma(t_prime),
+                class_labels,
+                cfg_scale,
+                feat=feat,
+            )["x"].to(torch.float64)
+            d_prime = (
+                sigma_deriv(t_prime) / sigma(t_prime) + s_deriv(t_prime) / s(t_prime)
+            ) * x_prime - sigma_deriv(t_prime) * s(t_prime) / sigma(t_prime) * denoised
+            x_next = x_hat + h * (
+                (1 - 1 / (2 * alpha)) * d_cur + 1 / (2 * alpha) * d_prime
+            )
+
+    return x_next
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/gaussian_diffusion.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/gaussian_diffusion.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/gaussian_diffusion.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/gaussian_diffusion.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/basic_modules.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/basic_modules.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/basic_modules.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/basic_modules.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/readme.md b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/readme.md
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/fastlinear/readme.md
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/fastlinear/readme.md
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/sana.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/sana.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_U_shape.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_U_shape.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_U_shape.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_U_shape.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_blocks.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_blocks.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_blocks.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_blocks.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_multi_scale.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_multi_scale.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_multi_scale.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_multi_scale.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_others.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_others.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/nets/sana_others.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/nets/sana_others.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/norms.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/norms.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/norms.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/norms.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/respace.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/respace.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/respace.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/respace.py
diff --git a/image/sana/sana-1600m/packages/Sana/diffusion/model/sa_solver.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/sa_solver.py
new file mode 100755
index 000000000..f36ccb4d0
--- /dev/null
+++ b/image/sana/sana-1600m/packages/Sana/diffusion/model/sa_solver.py
@@ -0,0 +1,1616 @@
+# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+import math
+
+import torch
+from tqdm import tqdm
+
+
+class NoiseScheduleVP:
+    def __init__(
+        self,
+        schedule="discrete",
+        betas=None,
+        alphas_cumprod=None,
+        continuous_beta_0=0.1,
+        continuous_beta_1=20.0,
+        dtype=torch.float32,
+    ):
+        """Thanks to DPM-Solver for their code base"""
+        r"""Create a wrapper class for the forward SDE (VP type).
+        ***
+        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
+                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
+        ***
+        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
+        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
+        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
+            log_alpha_t = self.marginal_log_mean_coeff(t)
+            sigma_t = self.marginal_std(t)
+            lambda_t = self.marginal_lambda(t)
+        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
+            t = self.inverse_lambda(lambda_t)
+        ===============================================================
+        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
+        1. For discrete-time DPMs:
+            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
+                t_i = (i + 1) / N
+            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
+            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
+            Args:
+                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
+                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
+            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
+            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
+                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
+                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
+                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
+                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
+                and
+                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
+        2. For continuous-time DPMs:
+            We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
+            schedule are the default settings in DDPM and improved-DDPM:
+            Args:
+                beta_min: A `float` number. The smallest beta for the linear schedule.
+                beta_max: A `float` number. The largest beta for the linear schedule.
+                cosine_s: A `float` number. The hyperparameter in the cosine schedule.
+                cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
+                T: A `float` number. The ending time of the forward process.
+        ===============================================================
+        Args:
+            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
+                    'linear' or 'cosine' for continuous-time DPMs.
+        Returns:
+            A wrapper object of the forward SDE (VP type).
+
+        ===============================================================
+        Example:
+        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', betas=betas)
+        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
+        # For continuous-time DPMs (VPSDE), linear schedule:
+        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
+        """
+
+        if schedule not in ["discrete", "linear", "cosine"]:
+            raise ValueError(
+                "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(
+                    schedule
+                )
+            )
+
+        self.schedule = schedule
+        if schedule == "discrete":
+            if betas is not None:
+                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
+            else:
+                assert alphas_cumprod is not None
+                log_alphas = 0.5 * torch.log(alphas_cumprod)
+            self.total_N = len(log_alphas)
+            self.T = 1.0
+            self.t_array = (
+                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
+                .reshape((1, -1))
+                .to(dtype=dtype)
+            )
+            self.log_alpha_array = log_alphas.reshape(
+                (
+                    1,
+                    -1,
+                )
+            ).to(dtype=dtype)
+        else:
+            self.total_N = 1000
+            self.beta_0 = continuous_beta_0
+            self.beta_1 = continuous_beta_1
+            self.cosine_s = 0.008
+            self.cosine_beta_max = 999.0
+            self.cosine_t_max = (
+                math.atan(self.cosine_beta_max * (1.0 + self.cosine_s) / math.pi)
+                * 2.0
+                * (1.0 + self.cosine_s)
+                / math.pi
+                - self.cosine_s
+            )
+            self.cosine_log_alpha_0 = math.log(
+                math.cos(self.cosine_s / (1.0 + self.cosine_s) * math.pi / 2.0)
+            )
+            self.schedule = schedule
+            if schedule == "cosine":
+                # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
+                # Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
+                self.T = 0.9946
+            else:
+                self.T = 1.0
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        if self.schedule == "discrete":
+            return interpolate_fn(
+                t.reshape((-1, 1)),
+                self.t_array.to(t.device),
+                self.log_alpha_array.to(t.device),
+            ).reshape(-1)
+        elif self.schedule == "linear":
+            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
+        elif self.schedule == "cosine":
+            log_alpha_fn = lambda s: torch.log(
+                torch.cos((s + self.cosine_s) / (1.0 + self.cosine_s) * math.pi / 2.0)
+            )
+            log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
+            return log_alpha_t
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return torch.exp(self.marginal_log_mean_coeff(t))
+
+    def marginal_std(self, t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
+        return log_mean_coeff - log_std
+
+    def inverse_lambda(self, lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        if self.schedule == "linear":
+            tmp = (
+                2.0
+                * (self.beta_1 - self.beta_0)
+                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
+            )
+            Delta = self.beta_0**2 + tmp
+            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
+        elif self.schedule == "discrete":
+            log_alpha = -0.5 * torch.logaddexp(
+                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
+            )
+            t = interpolate_fn(
+                log_alpha.reshape((-1, 1)),
+                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
+                torch.flip(self.t_array.to(lamb.device), [1]),
+            )
+            return t.reshape((-1,))
+        else:
+            log_alpha = -0.5 * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
+            t_fn = lambda log_alpha_t: (
+                torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0))
+                * 2.0
+                * (1.0 + self.cosine_s)
+                / math.pi
+                - self.cosine_s
+            )
+            t = t_fn(log_alpha)
+            return t
+
+    def edm_sigma(self, t):
+        return self.marginal_std(t) / self.marginal_alpha(t)
+
+    def edm_inverse_sigma(self, edmsigma):
+        alpha = 1 / (edmsigma**2 + 1).sqrt()
+        sigma = alpha * edmsigma
+        lambda_t = torch.log(alpha / sigma)
+        t = self.inverse_lambda(lambda_t)
+        return t
+
+
+def model_wrapper(
+    model,
+    noise_schedule,
+    model_type="noise",
+    model_kwargs={},
+    guidance_type="uncond",
+    condition=None,
+    unconditional_condition=None,
+    guidance_scale=1.0,
+    classifier_fn=None,
+    classifier_kwargs={},
+):
+    """Thanks to DPM-Solver for their code base"""
+    """Create a wrapper function for the noise prediction model.
+    SA-Solver needs to solve the continuous-time diffusion SDEs. For DPMs trained on discrete-time labels, we need to
+    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
+    We support four types of the diffusion model by setting `model_type`:
+        1. "noise": noise prediction model. (Trained by predicting noise).
+        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
+        3. "v": velocity prediction model. (Trained by predicting the velocity).
+            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
+            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
+                arXiv preprint arXiv:2202.00512 (2022).
+            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
+                arXiv preprint arXiv:2210.02303 (2022).
+
+        4. "score": marginal score function. (Trained by denoising score matching).
+            Note that the score function and the noise prediction model follows a simple relationship:
+            ```
+                noise(x_t, t) = -sigma_t * score(x_t, t)
+            ```
+    We support three types of guided sampling by DPMs by setting `guidance_type`:
+        1. "uncond": unconditional sampling by DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+            The input `classifier_fn` has the following format:
+            ``
+                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
+            ``
+            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
+                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
+        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
+            ``
+            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
+            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
+                arXiv preprint arXiv:2207.12598 (2022).
+
+    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
+    or continuous-time labels (i.e. epsilon to T).
+    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
+    ``
+        def model_fn(x, t_continuous) -> noise:
+            t_input = get_model_input_time(t_continuous)
+            return noise_pred(model, x, t_input, **model_kwargs)
+    ``
+    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for SA-Solver.
+    ===============================================================
+    Args:
+        model: A diffusion model with the corresponding format described above.
+        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+        model_type: A `str`. The parameterization type of the diffusion model.
+                    "noise" or "x_start" or "v" or "score".
+        model_kwargs: A `dict`. A dict for the other inputs of the model function.
+        guidance_type: A `str`. The type of the guidance for sampling.
+                    "uncond" or "classifier" or "classifier-free".
+        condition: A pytorch tensor. The condition for the guided sampling.
+                    Only used for "classifier" or "classifier-free" guidance type.
+        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
+                    Only used for "classifier-free" guidance type.
+        guidance_scale: A `float`. The scale for the guided sampling.
+        classifier_fn: A classifier function. Only used for the classifier guidance.
+        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
+    Returns:
+        A noise prediction model that accepts the noised data and the continuous time as the inputs.
+    """
+
+    def get_model_input_time(t_continuous):
+        """
+        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
+        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
+        For continuous-time DPMs, we just use `t_continuous`.
+        """
+        if noise_schedule.schedule == "discrete":
+            return (t_continuous - 1.0 / noise_schedule.total_N) * 1000.0
+        else:
+            return t_continuous
+
+    def noise_pred_fn(x, t_continuous, cond=None):
+        t_input = get_model_input_time(t_continuous)
+        if cond is None:
+            output = model(x, t_input, **model_kwargs)
+        else:
+            output = model(x, t_input, cond, **model_kwargs)
+        if model_type == "noise":
+            return output
+        elif model_type == "x_start":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return (x - alpha_t[0] * output) / sigma_t[0]
+        elif model_type == "v":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return alpha_t[0] * output + sigma_t[0] * x
+        elif model_type == "score":
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            return -sigma_t[0] * output
+
+    def cond_grad_fn(x, t_input):
+        """
+        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
+        """
+        with torch.enable_grad():
+            x_in = x.detach().requires_grad_(True)
+            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
+            return torch.autograd.grad(log_prob.sum(), x_in)[0]
+
+    def model_fn(x, t_continuous):
+        """
+        The noise predicition model function that is used for DPM-Solver.
+        """
+        if guidance_type == "uncond":
+            return noise_pred_fn(x, t_continuous)
+        elif guidance_type == "classifier":
+            assert classifier_fn is not None
+            t_input = get_model_input_time(t_continuous)
+            cond_grad = cond_grad_fn(x, t_input)
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            noise = noise_pred_fn(x, t_continuous)
+            return noise - guidance_scale * sigma_t * cond_grad
+        elif guidance_type == "classifier-free":
+            if guidance_scale == 1.0 or unconditional_condition is None:
+                return noise_pred_fn(x, t_continuous, cond=condition)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t_continuous] * 2)
+                c_in = torch.cat([unconditional_condition, condition])
+                noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+
+    assert model_type in ["noise", "x_start", "v", "score"]
+    assert guidance_type in ["uncond", "classifier", "classifier-free"]
+    return model_fn
+
+
+class SASolver:
+    def __init__(
+        self,
+        model_fn,
+        noise_schedule,
+        algorithm_type="data_prediction",
+        correcting_x0_fn=None,
+        correcting_xt_fn=None,
+        thresholding_max_val=1.0,
+        dynamic_thresholding_ratio=0.995,
+    ):
+        """
+        Construct a SA-Solver
+        The default value for algorithm_type is "data_prediction" and we recommend not to change it to
+        "noise_prediction". For details, please see Appendix A.2.4 in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
+        self.noise_schedule = noise_schedule
+        assert algorithm_type in ["data_prediction", "noise_prediction"]
+
+        if correcting_x0_fn == "dynamic_thresholding":
+            self.correcting_x0_fn = self.dynamic_thresholding_fn
+        else:
+            self.correcting_x0_fn = correcting_x0_fn
+
+        self.correcting_xt_fn = correcting_xt_fn
+        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
+        self.thresholding_max_val = thresholding_max_val
+
+        self.predict_x0 = algorithm_type == "data_prediction"
+
+        self.sigma_min = float(self.noise_schedule.edm_sigma(torch.tensor([1e-3])))
+        self.sigma_max = float(self.noise_schedule.edm_sigma(torch.tensor([1])))
+
+    def dynamic_thresholding_fn(self, x0, t=None):
+        """
+        The dynamic thresholding method.
+        """
+        dims = x0.dim()
+        p = self.dynamic_thresholding_ratio
+        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
+        s = expand_dims(
+            torch.maximum(
+                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
+            ),
+            dims,
+        )
+        x0 = torch.clamp(x0, -s, s) / s
+        return x0
+
+    def noise_prediction_fn(self, x, t):
+        """
+        Return the noise prediction model.
+        """
+        return self.model(x, t)
+
+    def data_prediction_fn(self, x, t):
+        """
+        Return the data prediction model (with corrector).
+        """
+        noise = self.noise_prediction_fn(x, t)
+        alpha_t, sigma_t = (
+            self.noise_schedule.marginal_alpha(t),
+            self.noise_schedule.marginal_std(t),
+        )
+        x0 = (x - sigma_t * noise) / alpha_t
+        if self.correcting_x0_fn is not None:
+            x0 = self.correcting_x0_fn(x0)
+        return x0
+
+    def model_fn(self, x, t):
+        """
+        Convert the model to the noise prediction model or the data prediction model.
+        """
+
+        if self.predict_x0:
+            return self.data_prediction_fn(x, t)
+        else:
+            return self.noise_prediction_fn(x, t)
+
+    def get_time_steps(self, skip_type, t_T, t_0, N, order, device):
+        """Compute the intermediate time steps for sampling."""
+        if skip_type == "logSNR":
+            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
+            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
+            logSNR_steps = lambda_T + torch.linspace(
+                torch.tensor(0.0).cpu().item(),
+                (lambda_0 - lambda_T).cpu().item() ** (1.0 / order),
+                N + 1,
+            ).pow(order).to(device)
+            return self.noise_schedule.inverse_lambda(logSNR_steps)
+        elif skip_type == "time":
+            t = (
+                torch.linspace(t_T ** (1.0 / order), t_0 ** (1.0 / order), N + 1)
+                .pow(order)
+                .to(device)
+            )
+            return t
+        elif skip_type == "karras":
+            sigma_min = max(0.002, self.sigma_min)
+            sigma_max = min(80, self.sigma_max)
+            sigma_steps = (
+                torch.linspace(sigma_max ** (1.0 / 7), sigma_min ** (1.0 / 7), N + 1)
+                .pow(7)
+                .to(device)
+            )
+            t = self.noise_schedule.edm_inverse_sigma(sigma_steps)
+            return t
+        else:
+            raise ValueError(
+                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time' or 'karras'"
+            )
+
+    def denoise_to_zero_fn(self, x, s):
+        """
+        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
+        """
+        return self.data_prediction_fn(x, s)
+
+    def get_coefficients_exponential_negative(
+        self, order, interval_start, interval_end
+    ):
+        """
+        Calculate the integral of exp(-x) * x^order dx from interval_start to interval_end
+        For calculating the coefficient of gradient terms after the lagrange interpolation,
+        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        For noise_prediction formula.
+        """
+        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
+
+        if order == 0:
+            return torch.exp(-interval_end) * (
+                torch.exp(interval_end - interval_start) - 1
+            )
+        elif order == 1:
+            return torch.exp(-interval_end) * (
+                (interval_start + 1) * torch.exp(interval_end - interval_start)
+                - (interval_end + 1)
+            )
+        elif order == 2:
+            return torch.exp(-interval_end) * (
+                (interval_start**2 + 2 * interval_start + 2)
+                * torch.exp(interval_end - interval_start)
+                - (interval_end**2 + 2 * interval_end + 2)
+            )
+        elif order == 3:
+            return torch.exp(-interval_end) * (
+                (interval_start**3 + 3 * interval_start**2 + 6 * interval_start + 6)
+                * torch.exp(interval_end - interval_start)
+                - (interval_end**3 + 3 * interval_end**2 + 6 * interval_end + 6)
+            )
+
+    def get_coefficients_exponential_positive(
+        self, order, interval_start, interval_end, tau
+    ):
+        """
+        Calculate the integral of exp(x(1+tau^2)) * x^order dx from interval_start to interval_end
+        For calculating the coefficient of gradient terms after the lagrange interpolation,
+        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        For data_prediction formula.
+        """
+        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
+
+        # after change of variable(cov)
+        interval_end_cov = (1 + tau**2) * interval_end
+        interval_start_cov = (1 + tau**2) * interval_start
+
+        if order == 0:
+            return (
+                torch.exp(interval_end_cov)
+                * (1 - torch.exp(-(interval_end_cov - interval_start_cov)))
+                / (1 + tau**2)
+            )
+        elif order == 1:
+            return (
+                torch.exp(interval_end_cov)
+                * (
+                    (interval_end_cov - 1)
+                    - (interval_start_cov - 1)
+                    * torch.exp(-(interval_end_cov - interval_start_cov))
+                )
+                / ((1 + tau**2) ** 2)
+            )
+        elif order == 2:
+            return (
+                torch.exp(interval_end_cov)
+                * (
+                    (interval_end_cov**2 - 2 * interval_end_cov + 2)
+                    - (interval_start_cov**2 - 2 * interval_start_cov + 2)
+                    * torch.exp(-(interval_end_cov - interval_start_cov))
+                )
+                / ((1 + tau**2) ** 3)
+            )
+        elif order == 3:
+            return (
+                torch.exp(interval_end_cov)
+                * (
+                    (
+                        interval_end_cov**3
+                        - 3 * interval_end_cov**2
+                        + 6 * interval_end_cov
+                        - 6
+                    )
+                    - (
+                        interval_start_cov**3
+                        - 3 * interval_start_cov**2
+                        + 6 * interval_start_cov
+                        - 6
+                    )
+                    * torch.exp(-(interval_end_cov - interval_start_cov))
+                )
+                / ((1 + tau**2) ** 4)
+            )
+
+    def lagrange_polynomial_coefficient(self, order, lambda_list):
+        """
+        Calculate the coefficient of lagrange polynomial
+        For lagrange interpolation
+        """
+        assert order in [0, 1, 2, 3]
+        assert order == len(lambda_list) - 1
+        if order == 0:
+            return [[1]]
+        elif order == 1:
+            return [
+                [
+                    1 / (lambda_list[0] - lambda_list[1]),
+                    -lambda_list[1] / (lambda_list[0] - lambda_list[1]),
+                ],
+                [
+                    1 / (lambda_list[1] - lambda_list[0]),
+                    -lambda_list[0] / (lambda_list[1] - lambda_list[0]),
+                ],
+            ]
+        elif order == 2:
+            denominator1 = (lambda_list[0] - lambda_list[1]) * (
+                lambda_list[0] - lambda_list[2]
+            )
+            denominator2 = (lambda_list[1] - lambda_list[0]) * (
+                lambda_list[1] - lambda_list[2]
+            )
+            denominator3 = (lambda_list[2] - lambda_list[0]) * (
+                lambda_list[2] - lambda_list[1]
+            )
+            return [
+                [
+                    1 / denominator1,
+                    (-lambda_list[1] - lambda_list[2]) / denominator1,
+                    lambda_list[1] * lambda_list[2] / denominator1,
+                ],
+                [
+                    1 / denominator2,
+                    (-lambda_list[0] - lambda_list[2]) / denominator2,
+                    lambda_list[0] * lambda_list[2] / denominator2,
+                ],
+                [
+                    1 / denominator3,
+                    (-lambda_list[0] - lambda_list[1]) / denominator3,
+                    lambda_list[0] * lambda_list[1] / denominator3,
+                ],
+            ]
+        elif order == 3:
+            denominator1 = (
+                (lambda_list[0] - lambda_list[1])
+                * (lambda_list[0] - lambda_list[2])
+                * (lambda_list[0] - lambda_list[3])
+            )
+            denominator2 = (
+                (lambda_list[1] - lambda_list[0])
+                * (lambda_list[1] - lambda_list[2])
+                * (lambda_list[1] - lambda_list[3])
+            )
+            denominator3 = (
+                (lambda_list[2] - lambda_list[0])
+                * (lambda_list[2] - lambda_list[1])
+                * (lambda_list[2] - lambda_list[3])
+            )
+            denominator4 = (
+                (lambda_list[3] - lambda_list[0])
+                * (lambda_list[3] - lambda_list[1])
+                * (lambda_list[3] - lambda_list[2])
+            )
+            return [
+                [
+                    1 / denominator1,
+                    (-lambda_list[1] - lambda_list[2] - lambda_list[3]) / denominator1,
+                    (
+                        lambda_list[1] * lambda_list[2]
+                        + lambda_list[1] * lambda_list[3]
+                        + lambda_list[2] * lambda_list[3]
+                    )
+                    / denominator1,
+                    (-lambda_list[1] * lambda_list[2] * lambda_list[3]) / denominator1,
+                ],
+                [
+                    1 / denominator2,
+                    (-lambda_list[0] - lambda_list[2] - lambda_list[3]) / denominator2,
+                    (
+                        lambda_list[0] * lambda_list[2]
+                        + lambda_list[0] * lambda_list[3]
+                        + lambda_list[2] * lambda_list[3]
+                    )
+                    / denominator2,
+                    (-lambda_list[0] * lambda_list[2] * lambda_list[3]) / denominator2,
+                ],
+                [
+                    1 / denominator3,
+                    (-lambda_list[0] - lambda_list[1] - lambda_list[3]) / denominator3,
+                    (
+                        lambda_list[0] * lambda_list[1]
+                        + lambda_list[0] * lambda_list[3]
+                        + lambda_list[1] * lambda_list[3]
+                    )
+                    / denominator3,
+                    (-lambda_list[0] * lambda_list[1] * lambda_list[3]) / denominator3,
+                ],
+                [
+                    1 / denominator4,
+                    (-lambda_list[0] - lambda_list[1] - lambda_list[2]) / denominator4,
+                    (
+                        lambda_list[0] * lambda_list[1]
+                        + lambda_list[0] * lambda_list[2]
+                        + lambda_list[1] * lambda_list[2]
+                    )
+                    / denominator4,
+                    (-lambda_list[0] * lambda_list[1] * lambda_list[2]) / denominator4,
+                ],
+            ]
+
+    def get_coefficients_fn(
+        self, order, interval_start, interval_end, lambda_list, tau
+    ):
+        """
+        Calculate the coefficient of gradients.
+        """
+        assert order in [1, 2, 3, 4]
+        assert order == len(lambda_list), (
+            "the length of lambda list must be equal to the order"
+        )
+        coefficients = []
+        lagrange_coefficient = self.lagrange_polynomial_coefficient(
+            order - 1, lambda_list
+        )
+        for i in range(order):
+            coefficient = 0
+            for j in range(order):
+                if self.predict_x0:
+                    coefficient += lagrange_coefficient[i][
+                        j
+                    ] * self.get_coefficients_exponential_positive(
+                        order - 1 - j, interval_start, interval_end, tau
+                    )
+                else:
+                    coefficient += lagrange_coefficient[i][
+                        j
+                    ] * self.get_coefficients_exponential_negative(
+                        order - 1 - j, interval_start, interval_end
+                    )
+            coefficients.append(coefficient)
+        assert len(coefficients) == order, (
+            "the length of coefficients does not match the order"
+        )
+        return coefficients
+
+    def adams_bashforth_update(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Predictor, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def adams_moulton_update(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        t_list = t_prev_list + [t]
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def adams_bashforth_update_few_steps(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Predictor, with the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        if self.predict_x0:
+            if (
+                order == 2
+            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to unipc. Note: This is used only for few steps sampling.
+                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
+                # ODE case
+                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
+                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
+                gradient_coefficients[0] += (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h**2 / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2)
+                    )
+                    / (
+                        ns.marginal_lambda(t_prev_list[-1])
+                        - ns.marginal_lambda(t_prev_list[-2])
+                    )
+                )
+                gradient_coefficients[1] -= (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h**2 / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2)
+                    )
+                    / (
+                        ns.marginal_lambda(t_prev_list[-1])
+                        - ns.marginal_lambda(t_prev_list[-2])
+                    )
+                )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def adams_moulton_update_few_steps(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        t_list = t_prev_list + [t]
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        if self.predict_x0:
+            if (
+                order == 2
+            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to UniPC. Note: This is used only for few steps sampling.
+                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
+                # ODE case
+                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
+                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
+                gradient_coefficients[0] += (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2 * h)
+                    )
+                )
+                gradient_coefficients[1] -= (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2 * h)
+                    )
+                )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def sample_few_steps(
+        self,
+        x,
+        tau,
+        steps=5,
+        t_start=None,
+        t_end=None,
+        skip_type="time",
+        skip_order=1,
+        predictor_order=3,
+        corrector_order=4,
+        pc_mode="PEC",
+        return_intermediate=False,
+    ):
+        """
+        For the PC-mode, please refer to the wiki page
+        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
+        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
+        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
+        """
+
+        skip_first_step = False
+        skip_final_step = True
+        lower_order_final = True
+        denoise_to_zero = False
+
+        assert pc_mode in [
+            "PEC",
+            "PECE",
+        ], "Predictor-corrector mode only supports PEC and PECE"
+        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+
+        device = x.device
+        intermediates = []
+        with torch.no_grad():
+            assert steps >= max(predictor_order, corrector_order - 1)
+            timesteps = self.get_time_steps(
+                skip_type=skip_type,
+                t_T=t_T,
+                t_0=t_0,
+                N=steps,
+                order=skip_order,
+                device=device,
+            )
+            assert timesteps.shape[0] - 1 == steps
+            # Init the initial values.
+            step = 0
+            t = timesteps[step]
+            noise = torch.randn_like(x)
+            t_prev_list = [t]
+            # do not evaluate if skip_first_step
+            if skip_first_step:
+                if self.predict_x0:
+                    alpha_t = self.noise_schedule.marginal_alpha(t)
+                    sigma_t = self.noise_schedule.marginal_std(t)
+                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
+                else:
+                    model_prev_list = [x]
+            else:
+                model_prev_list = [self.model_fn(x, t)]
+
+            if self.correcting_xt_fn is not None:
+                x = self.correcting_xt_fn(x, t, step)
+            if return_intermediate:
+                intermediates.append(x)
+
+            # determine the first several values
+            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
+                t = timesteps[step]
+                predictor_order_used = min(predictor_order, step)
+                corrector_order_used = min(corrector_order, step + 1)
+                noise = torch.randn_like(x)
+                # predictor step
+                x_p = self.adams_bashforth_update_few_steps(
+                    order=predictor_order_used,
+                    x=x,
+                    tau=tau(t),
+                    model_prev_list=model_prev_list,
+                    t_prev_list=t_prev_list,
+                    noise=noise,
+                    t=t,
+                )
+                # evaluation step
+                model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                model_prev_list.append(model_x)
+                # corrector step
+                if corrector_order > 0:
+                    x = self.adams_moulton_update_few_steps(
+                        order=corrector_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x = x_p
+
+                # evaluation step if correction and mode = pece
+                if corrector_order > 0:
+                    if pc_mode == "PECE":
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+
+            for step in tqdm(
+                range(max(predictor_order, corrector_order - 1), steps + 1)
+            ):
+                if lower_order_final:
+                    predictor_order_used = min(predictor_order, steps - step + 1)
+                    corrector_order_used = min(corrector_order, steps - step + 2)
+
+                else:
+                    predictor_order_used = predictor_order
+                    corrector_order_used = corrector_order
+                t = timesteps[step]
+                noise = torch.randn_like(x)
+
+                # predictor step
+                if skip_final_step and step == steps and not denoise_to_zero:
+                    x_p = self.adams_bashforth_update_few_steps(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=0,
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x_p = self.adams_bashforth_update_few_steps(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+
+                # evaluation step
+                # do not evaluate if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                # do not update if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_prev_list.append(model_x)
+
+                # corrector step
+                # do not correct if skip_final_step and step = steps
+                if corrector_order > 0:
+                    if not skip_final_step or step < steps:
+                        x = self.adams_moulton_update_few_steps(
+                            order=corrector_order_used,
+                            x=x,
+                            tau=tau(t),
+                            model_prev_list=model_prev_list,
+                            t_prev_list=t_prev_list,
+                            noise=noise,
+                            t=t,
+                        )
+                    else:
+                        x = x_p
+                else:
+                    x = x_p
+
+                # evaluation step if mode = pece and step != steps
+                if corrector_order > 0:
+                    if pc_mode == "PECE" and step < steps:
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+                del model_prev_list[0]
+
+            if denoise_to_zero:
+                t = torch.ones((1,)).to(device) * t_0
+                x = self.denoise_to_zero_fn(x, t)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step + 1)
+                if return_intermediate:
+                    intermediates.append(x)
+        if return_intermediate:
+            return x, intermediates
+        else:
+            return x
+
+    def sample_more_steps(
+        self,
+        x,
+        tau,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        skip_type="time",
+        skip_order=1,
+        predictor_order=3,
+        corrector_order=4,
+        pc_mode="PEC",
+        return_intermediate=False,
+    ):
+        """
+        For the PC-mode, please refer to the wiki page
+        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
+        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
+        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
+        """
+
+        skip_first_step = False
+        skip_final_step = False
+        lower_order_final = True
+        denoise_to_zero = True
+
+        assert pc_mode in [
+            "PEC",
+            "PECE",
+        ], "Predictor-corrector mode only supports PEC and PECE"
+        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+
+        device = x.device
+        intermediates = []
+        with torch.no_grad():
+            assert steps >= max(predictor_order, corrector_order - 1)
+            timesteps = self.get_time_steps(
+                skip_type=skip_type,
+                t_T=t_T,
+                t_0=t_0,
+                N=steps,
+                order=skip_order,
+                device=device,
+            )
+            assert timesteps.shape[0] - 1 == steps
+            # Init the initial values.
+            step = 0
+            t = timesteps[step]
+            noise = torch.randn_like(x)
+            t_prev_list = [t]
+            # do not evaluate if skip_first_step
+            if skip_first_step:
+                if self.predict_x0:
+                    alpha_t = self.noise_schedule.marginal_alpha(t)
+                    sigma_t = self.noise_schedule.marginal_std(t)
+                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
+                else:
+                    model_prev_list = [x]
+            else:
+                model_prev_list = [self.model_fn(x, t)]
+
+            if self.correcting_xt_fn is not None:
+                x = self.correcting_xt_fn(x, t, step)
+            if return_intermediate:
+                intermediates.append(x)
+
+            # determine the first several values
+            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
+                t = timesteps[step]
+                predictor_order_used = min(predictor_order, step)
+                corrector_order_used = min(corrector_order, step + 1)
+                noise = torch.randn_like(x)
+                # predictor step
+                x_p = self.adams_bashforth_update(
+                    order=predictor_order_used,
+                    x=x,
+                    tau=tau(t),
+                    model_prev_list=model_prev_list,
+                    t_prev_list=t_prev_list,
+                    noise=noise,
+                    t=t,
+                )
+                # evaluation step
+                model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                model_prev_list.append(model_x)
+                # corrector step
+                if corrector_order > 0:
+                    x = self.adams_moulton_update(
+                        order=corrector_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x = x_p
+
+                # evaluation step if mode = pece
+                if corrector_order > 0:
+                    if pc_mode == "PECE":
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+
+            for step in tqdm(
+                range(max(predictor_order, corrector_order - 1), steps + 1)
+            ):
+                if lower_order_final:
+                    predictor_order_used = min(predictor_order, steps - step + 1)
+                    corrector_order_used = min(corrector_order, steps - step + 2)
+
+                else:
+                    predictor_order_used = predictor_order
+                    corrector_order_used = corrector_order
+                t = timesteps[step]
+                noise = torch.randn_like(x)
+
+                # predictor step
+                if skip_final_step and step == steps and not denoise_to_zero:
+                    x_p = self.adams_bashforth_update(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=0,
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x_p = self.adams_bashforth_update(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+
+                # evaluation step
+                # do not evaluate if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                # do not update if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_prev_list.append(model_x)
+
+                # corrector step
+                # do not correct if skip_final_step and step = steps
+                if corrector_order > 0:
+                    if not skip_final_step or step < steps:
+                        x = self.adams_moulton_update(
+                            order=corrector_order_used,
+                            x=x,
+                            tau=tau(t),
+                            model_prev_list=model_prev_list,
+                            t_prev_list=t_prev_list,
+                            noise=noise,
+                            t=t,
+                        )
+                    else:
+                        x = x_p
+                else:
+                    x = x_p
+
+                # evaluation step if mode = pece and step != steps
+                if corrector_order > 0:
+                    if pc_mode == "PECE" and step < steps:
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+                del model_prev_list[0]
+
+            if denoise_to_zero:
+                t = torch.ones((1,)).to(device) * t_0
+                x = self.denoise_to_zero_fn(x, t)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step + 1)
+                if return_intermediate:
+                    intermediates.append(x)
+        if return_intermediate:
+            return x, intermediates
+        else:
+            return x
+
+    def sample(
+        self,
+        mode,
+        x,
+        tau,
+        steps,
+        t_start=None,
+        t_end=None,
+        skip_type="time",
+        skip_order=1,
+        predictor_order=3,
+        corrector_order=4,
+        pc_mode="PEC",
+        return_intermediate=False,
+    ):
+        """
+        For the PC-mode, please refer to the wiki page
+        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
+        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
+        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
+
+        'few_steps' mode is recommended. The differences between 'few_steps' and 'more_steps' are as below:
+        1) 'few_steps' do not correct at final step and do not denoise to zero, while 'more_steps' do these two.
+        Thus the NFEs for 'few_steps' = steps, NFEs for 'more_steps' = steps + 2
+        For most of the experiments and tasks, we find these two operations do not have much help to sample quality.
+        2) 'few_steps' use a rescaling trick as in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        We find it will slightly improve the sample quality especially in few steps.
+        """
+        assert mode in [
+            "few_steps",
+            "more_steps",
+        ], "mode must be either 'few_steps' or 'more_steps'"
+        if mode == "few_steps":
+            return self.sample_few_steps(
+                x=x,
+                tau=tau,
+                steps=steps,
+                t_start=t_start,
+                t_end=t_end,
+                skip_type=skip_type,
+                skip_order=skip_order,
+                predictor_order=predictor_order,
+                corrector_order=corrector_order,
+                pc_mode=pc_mode,
+                return_intermediate=return_intermediate,
+            )
+        else:
+            return self.sample_more_steps(
+                x=x,
+                tau=tau,
+                steps=steps,
+                t_start=t_start,
+                t_end=t_end,
+                skip_type=skip_type,
+                skip_order=skip_order,
+                predictor_order=predictor_order,
+                corrector_order=corrector_order,
+                pc_mode=pc_mode,
+                return_intermediate=return_intermediate,
+            )
+
+
+#############################################################
+# other utility functions
+#############################################################
+
+
+def interpolate_fn(x, xp, yp):
+    """
+    A piecewise linear function y = f(x), using xp and yp as keypoints.
+    We implement f(x) in a differentiable way (i.e. applicable for autograd).
+    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
+    Args:
+        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
+        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
+        yp: PyTorch tensor with shape [C, K].
+    Returns:
+        The function values f(x), with shape [N, C].
+    """
+    N, K = x.shape[0], xp.shape[1]
+    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
+    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
+    x_idx = torch.argmin(x_indices, dim=2)
+    cand_start_idx = x_idx - 1
+    start_idx = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(1, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    end_idx = torch.where(
+        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
+    )
+    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
+    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
+    start_idx2 = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(0, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
+    start_y = torch.gather(
+        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
+    ).squeeze(2)
+    end_y = torch.gather(
+        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
+    ).squeeze(2)
+    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
+    return cand
+
+
+def expand_dims(v, dims):
+    """
+    Expand the tensor `v` to the dim `dims`.
+    Args:
+        `v`: a PyTorch tensor with shape [N].
+        `dim`: a `int`.
+    Returns:
+        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
+    """
+    return v[(...,) + (None,) * (dims - 1)]
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/timestep_sampler.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/timestep_sampler.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/timestep_sampler.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/timestep_sampler.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/utils.py b/image/sana/sana-1600m/packages/Sana/diffusion/model/utils.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/utils.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/model/utils.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/sa_sampler.py b/image/sana/sana-1600m/packages/Sana/diffusion/sa_sampler.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/sa_sampler.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/sa_sampler.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/sa_solver_diffusers.py b/image/sana/sana-1600m/packages/Sana/diffusion/sa_solver_diffusers.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/sa_solver_diffusers.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/sa_solver_diffusers.py
diff --git a/llama/llama-2-70b-chat/model/__init__.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/__init__.py
old mode 100644
new mode 100755
similarity index 100%
rename from llama/llama-2-70b-chat/model/__init__.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/checkpoint.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/checkpoint.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/checkpoint.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/checkpoint.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/config.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/config.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/config.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/config.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/data_sampler.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/data_sampler.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/data_sampler.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/data_sampler.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/dist_utils.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/dist_utils.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/dist_utils.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/dist_utils.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/import_utils.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/import_utils.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/import_utils.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/import_utils.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/logger.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/logger.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/logger.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/logger.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/lr_scheduler.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/lr_scheduler.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/lr_scheduler.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/lr_scheduler.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/misc.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/misc.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/misc.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/misc.py
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/optimizer.py b/image/sana/sana-1600m/packages/Sana/diffusion/utils/optimizer.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/optimizer.py
rename to image/sana/sana-1600m/packages/Sana/diffusion/utils/optimizer.py
diff --git a/sana/sana_1600M/packages/Sana/environment_setup.sh b/image/sana/sana-1600m/packages/Sana/environment_setup.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/environment_setup.sh
rename to image/sana/sana-1600m/packages/Sana/environment_setup.sh
diff --git a/sana/sana_1600M/packages/Sana/pyproject.toml b/image/sana/sana-1600m/packages/Sana/pyproject.toml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/pyproject.toml
rename to image/sana/sana-1600m/packages/Sana/pyproject.toml
diff --git a/sana/sana_1600M/packages/Sana/sana/cli/run.py b/image/sana/sana-1600m/packages/Sana/sana/cli/run.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/sana/cli/run.py
rename to image/sana/sana-1600m/packages/Sana/sana/cli/run.py
diff --git a/sana/sana_1600M/packages/Sana/sana/cli/upload2hf.py b/image/sana/sana-1600m/packages/Sana/sana/cli/upload2hf.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/sana/cli/upload2hf.py
rename to image/sana/sana-1600m/packages/Sana/sana/cli/upload2hf.py
diff --git a/sana/sana_1600M/packages/Sana/sana/tools/__init__.py b/image/sana/sana-1600m/packages/Sana/sana/tools/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/sana/tools/__init__.py
rename to image/sana/sana-1600m/packages/Sana/sana/tools/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/sana/tools/download.py b/image/sana/sana-1600m/packages/Sana/sana/tools/download.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/sana/tools/download.py
rename to image/sana/sana-1600m/packages/Sana/sana/tools/download.py
diff --git a/sana/sana_1600M/packages/Sana/sana/tools/hf_utils.py b/image/sana/sana-1600m/packages/Sana/sana/tools/hf_utils.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/sana/tools/hf_utils.py
rename to image/sana/sana-1600m/packages/Sana/sana/tools/hf_utils.py
diff --git a/sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric.sh b/image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric_dpg.sh b/image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric_dpg.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric_dpg.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric_dpg.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric_geneval.sh b/image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric_geneval.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric_geneval.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric_geneval.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh b/image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/infer_metric_run_inference_metric.sh b/image/sana/sana-1600m/packages/Sana/scripts/infer_metric_run_inference_metric.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/infer_metric_run_inference_metric.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/infer_metric_run_inference_metric.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh b/image/sana/sana-1600m/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/infer_run_inference.sh b/image/sana/sana-1600m/packages/Sana/scripts/infer_run_inference.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/infer_run_inference.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/infer_run_inference.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/infer_run_inference_geneval.sh b/image/sana/sana-1600m/packages/Sana/scripts/infer_run_inference_geneval.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/infer_run_inference_geneval.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/infer_run_inference_geneval.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh b/image/sana/sana-1600m/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh
rename to image/sana/sana-1600m/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh
diff --git a/sana/sana_1600M/packages/Sana/scripts/inference.py b/image/sana/sana-1600m/packages/Sana/scripts/inference.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/inference.py
rename to image/sana/sana-1600m/packages/Sana/scripts/inference.py
diff --git a/sana/sana_1600M/packages/Sana/scripts/inference_dpg.py b/image/sana/sana-1600m/packages/Sana/scripts/inference_dpg.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/inference_dpg.py
rename to image/sana/sana-1600m/packages/Sana/scripts/inference_dpg.py
diff --git a/sana/sana_1600M/packages/Sana/scripts/inference_geneval.py b/image/sana/sana-1600m/packages/Sana/scripts/inference_geneval.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/inference_geneval.py
rename to image/sana/sana-1600m/packages/Sana/scripts/inference_geneval.py
diff --git a/sana/sana_1600M/packages/Sana/scripts/inference_geneval_diffusers.py b/image/sana/sana-1600m/packages/Sana/scripts/inference_geneval_diffusers.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/inference_geneval_diffusers.py
rename to image/sana/sana-1600m/packages/Sana/scripts/inference_geneval_diffusers.py
diff --git a/sana/sana_1600M/packages/Sana/scripts/inference_image_reward.py b/image/sana/sana-1600m/packages/Sana/scripts/inference_image_reward.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/inference_image_reward.py
rename to image/sana/sana-1600m/packages/Sana/scripts/inference_image_reward.py
diff --git a/sana/sana_1600M/packages/Sana/scripts/interface.py b/image/sana/sana-1600m/packages/Sana/scripts/interface.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/interface.py
rename to image/sana/sana-1600m/packages/Sana/scripts/interface.py
diff --git a/sana/sana_1600M/packages/Sana/scripts/style.css b/image/sana/sana-1600m/packages/Sana/scripts/style.css
similarity index 100%
rename from sana/sana_1600M/packages/Sana/scripts/style.css
rename to image/sana/sana-1600m/packages/Sana/scripts/style.css
diff --git a/sana/sana_1600M/packages/Sana/tests/bash/entry.sh b/image/sana/sana-1600m/packages/Sana/tests/bash/entry.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tests/bash/entry.sh
rename to image/sana/sana-1600m/packages/Sana/tests/bash/entry.sh
diff --git a/sana/sana_1600M/packages/Sana/tests/bash/test_inference.sh b/image/sana/sana-1600m/packages/Sana/tests/bash/test_inference.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tests/bash/test_inference.sh
rename to image/sana/sana-1600m/packages/Sana/tests/bash/test_inference.sh
diff --git a/sana/sana_1600M/packages/Sana/tests/bash/test_training_1epoch.sh b/image/sana/sana-1600m/packages/Sana/tests/bash/test_training_1epoch.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tests/bash/test_training_1epoch.sh
rename to image/sana/sana-1600m/packages/Sana/tests/bash/test_training_1epoch.sh
diff --git a/llama/llama-2-70b/model/__init__.py b/image/sana/sana-1600m/packages/Sana/tools/__init__.py
similarity index 100%
rename from llama/llama-2-70b/model/__init__.py
rename to image/sana/sana-1600m/packages/Sana/tools/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/tools/convert_py_to_yaml.py b/image/sana/sana-1600m/packages/Sana/tools/convert_py_to_yaml.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/convert_py_to_yaml.py
rename to image/sana/sana-1600m/packages/Sana/tools/convert_py_to_yaml.py
diff --git a/sana/sana_1600M/packages/Sana/tools/convert_sana_pag_to_diffusers.py b/image/sana/sana-1600m/packages/Sana/tools/convert_sana_pag_to_diffusers.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/convert_sana_pag_to_diffusers.py
rename to image/sana/sana-1600m/packages/Sana/tools/convert_sana_pag_to_diffusers.py
diff --git a/sana/sana_1600M/packages/Sana/tools/convert_sana_to_diffusers.py b/image/sana/sana-1600m/packages/Sana/tools/convert_sana_to_diffusers.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/convert_sana_to_diffusers.py
rename to image/sana/sana-1600m/packages/Sana/tools/convert_sana_to_diffusers.py
diff --git a/sana/sana_1600M/packages/Sana/tools/download.py b/image/sana/sana-1600m/packages/Sana/tools/download.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/download.py
rename to image/sana/sana-1600m/packages/Sana/tools/download.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/.gitignore b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/.gitignore
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/.gitignore
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/.gitignore
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/LICENSE b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/LICENSE
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/LICENSE
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/LICENSE
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/README.md b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/README.md
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/README.md
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/README.md
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/clip_score.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/clip_score.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/clip_score.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/clip_score.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/setup.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/setup.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/setup.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/setup.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/compute_clipscore.sh b/image/sana/sana-1600m/packages/Sana/tools/metrics/compute_clipscore.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/compute_clipscore.sh
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/compute_clipscore.sh
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/compute_dpg.sh b/image/sana/sana-1600m/packages/Sana/tools/metrics/compute_dpg.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/compute_dpg.sh
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/compute_dpg.sh
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/compute_fid_embedding.sh b/image/sana/sana-1600m/packages/Sana/tools/metrics/compute_fid_embedding.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/compute_fid_embedding.sh
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/compute_fid_embedding.sh
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/compute_geneval.sh b/image/sana/sana-1600m/packages/Sana/tools/metrics/compute_geneval.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/compute_geneval.sh
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/compute_geneval.sh
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/compute_imagereward.sh b/image/sana/sana-1600m/packages/Sana/tools/metrics/compute_imagereward.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/compute_imagereward.sh
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/compute_imagereward.sh
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv b/image/sana/sana-1600m/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/dpg_bench/requirements.txt b/image/sana/sana-1600m/packages/Sana/tools/metrics/dpg_bench/requirements.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/dpg_bench/requirements.txt
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/dpg_bench/requirements.txt
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/LICENSE b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/LICENSE
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/LICENSE
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/LICENSE
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/README.md b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/README.md
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/README.md
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/README.md
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/environment.yml b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/environment.yml
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/environment.yml
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/environment.yml
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/object_names.txt b/image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/object_names.txt
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/geneval/prompts/object_names.txt
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/geneval/prompts/object_names.txt
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json b/image/sana/sana-1600m/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/image_reward/compute_image_reward.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/image_reward/compute_image_reward.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/image_reward/compute_image_reward.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/image_reward/compute_image_reward.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/.gitignore b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/.gitignore
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/.gitignore
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/.gitignore
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/LICENSE b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/LICENSE
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/LICENSE
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/LICENSE
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/README.md b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/README.md
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/README.md
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/README.md
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/noxfile.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/noxfile.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/noxfile.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/noxfile.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/setup.cfg b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/setup.cfg
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/setup.cfg
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/setup.cfg
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/setup.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/setup.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/setup.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/setup.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py
diff --git a/sana/sana_1600M/packages/Sana/tools/metrics/utils.py b/image/sana/sana-1600m/packages/Sana/tools/metrics/utils.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/metrics/utils.py
rename to image/sana/sana-1600m/packages/Sana/tools/metrics/utils.py
diff --git a/sana/sana_1600M/packages/Sana/train_scripts/train.py b/image/sana/sana-1600m/packages/Sana/train_scripts/train.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/train_scripts/train.py
rename to image/sana/sana-1600m/packages/Sana/train_scripts/train.py
diff --git a/sana/sana_1600M/packages/Sana/train_scripts/train.sh b/image/sana/sana-1600m/packages/Sana/train_scripts/train.sh
similarity index 100%
rename from sana/sana_1600M/packages/Sana/train_scripts/train.sh
rename to image/sana/sana-1600m/packages/Sana/train_scripts/train.sh
diff --git a/image/sana/sana-600m/README.md b/image/sana/sana-600m/README.md
new file mode 100644
index 000000000..2f83d6f71
--- /dev/null
+++ b/image/sana/sana-600m/README.md
@@ -0,0 +1,40 @@
+# Sana 600M
+
+Deploy Sana 600M for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | H100_40GB |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "a photo of an astronaut riding a horse on mars",
+  "height": 1024,
+  "width": 1024,
+  "guidance_scale": 5.0,
+  "pag_guidance_scale": 2.0,
+  "num_inference_steps": 18,
+  "seed": 4096
+}'
+```
+
+## Configuration highlights
+
+- Base image: `alphatozeta/cuda-python:12.1.1-cudnn8-devel-ubuntu22.04`
+- System packages: `ffmpeg, libsm6, libxext6, python3.10-venv`
diff --git a/image/sana/sana-600m/config.yaml b/image/sana/sana-600m/config.yaml
new file mode 100644
index 000000000..eb0174851
--- /dev/null
+++ b/image/sana/sana-600m/config.yaml
@@ -0,0 +1,33 @@
+description: "Sana 600M for image generation"
+build_commands: []
+base_image:
+  image: alphatozeta/cuda-python:12.1.1-cudnn8-devel-ubuntu22.04
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "Efficient-Large-Model/Sana_600M_1024px_diffusers"
+  example_model_input: {
+    "prompt": "a photo of an astronaut riding a horse on mars",
+    "height": 1024,
+    "width": 1024,
+    "guidance_scale": 5.0,
+    "pag_guidance_scale": 2.0,
+    "num_inference_steps": 18,
+    "seed": 4096,
+  }
+model_name: Sana 600M
+python_version: py311
+requirements:
+- git+https://github.com/NVlabs/Sana.git@d7945026d8d85008aca1d1e6db5717a1069f5c84
+- huggingface-hub==0.26.3
+- hf-transfer==0.1.8
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+secrets:
+  hf_access_token: "null"
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
+- python3.10-venv
diff --git a/llama/llama-2-7b-chat/model/__init__.py b/image/sana/sana-600m/model/__init__.py
similarity index 100%
rename from llama/llama-2-7b-chat/model/__init__.py
rename to image/sana/sana-600m/model/__init__.py
diff --git a/sana/sana_600M/model/model.py b/image/sana/sana-600m/model/model.py
similarity index 100%
rename from sana/sana_600M/model/model.py
rename to image/sana/sana-600m/model/model.py
diff --git a/sana/sana_600M/packages/Sana/CITATION.bib b/image/sana/sana-600m/packages/Sana/CITATION.bib
similarity index 100%
rename from sana/sana_600M/packages/Sana/CITATION.bib
rename to image/sana/sana-600m/packages/Sana/CITATION.bib
diff --git a/sana/sana_600M/packages/Sana/CIs/add_license_all.sh b/image/sana/sana-600m/packages/Sana/CIs/add_license_all.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/CIs/add_license_all.sh
rename to image/sana/sana-600m/packages/Sana/CIs/add_license_all.sh
diff --git a/sana/sana_600M/packages/Sana/Dockerfile b/image/sana/sana-600m/packages/Sana/Dockerfile
similarity index 100%
rename from sana/sana_600M/packages/Sana/Dockerfile
rename to image/sana/sana-600m/packages/Sana/Dockerfile
diff --git a/sana/sana_600M/packages/Sana/LICENSE b/image/sana/sana-600m/packages/Sana/LICENSE
similarity index 100%
rename from sana/sana_600M/packages/Sana/LICENSE
rename to image/sana/sana-600m/packages/Sana/LICENSE
diff --git a/sana/sana_600M/packages/Sana/README.md b/image/sana/sana-600m/packages/Sana/README.md
similarity index 100%
rename from sana/sana_600M/packages/Sana/README.md
rename to image/sana/sana-600m/packages/Sana/README.md
diff --git a/sana/sana_600M/packages/Sana/app/app_sana.py b/image/sana/sana-600m/packages/Sana/app/app_sana.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/app/app_sana.py
rename to image/sana/sana-600m/packages/Sana/app/app_sana.py
diff --git a/sana/sana_600M/packages/Sana/app/app_sana_multithread.py b/image/sana/sana-600m/packages/Sana/app/app_sana_multithread.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/app/app_sana_multithread.py
rename to image/sana/sana-600m/packages/Sana/app/app_sana_multithread.py
diff --git a/sana/sana_600M/packages/Sana/app/safety_check.py b/image/sana/sana-600m/packages/Sana/app/safety_check.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/app/safety_check.py
rename to image/sana/sana-600m/packages/Sana/app/safety_check.py
diff --git a/sana/sana_600M/packages/Sana/app/sana_pipeline.py b/image/sana/sana-600m/packages/Sana/app/sana_pipeline.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/app/sana_pipeline.py
rename to image/sana/sana-600m/packages/Sana/app/sana_pipeline.py
diff --git a/sana/sana_600M/packages/Sana/asset/Sana.jpg b/image/sana/sana-600m/packages/Sana/asset/Sana.jpg
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/Sana.jpg
rename to image/sana/sana-600m/packages/Sana/asset/Sana.jpg
diff --git a/sana/sana_600M/packages/Sana/asset/docs/metrics_toolkit.md b/image/sana/sana-600m/packages/Sana/asset/docs/metrics_toolkit.md
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/docs/metrics_toolkit.md
rename to image/sana/sana-600m/packages/Sana/asset/docs/metrics_toolkit.md
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/00000000.png b/image/sana/sana-600m/packages/Sana/asset/example_data/00000000.png
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/00000000.png
rename to image/sana/sana-600m/packages/Sana/asset/example_data/00000000.png
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/00000000.txt b/image/sana/sana-600m/packages/Sana/asset/example_data/00000000.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/00000000.txt
rename to image/sana/sana-600m/packages/Sana/asset/example_data/00000000.txt
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/00000000_InternVL2-26B.json b/image/sana/sana-600m/packages/Sana/asset/example_data/00000000_InternVL2-26B.json
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/00000000_InternVL2-26B.json
rename to image/sana/sana-600m/packages/Sana/asset/example_data/00000000_InternVL2-26B.json
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json b/image/sana/sana-600m/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json
rename to image/sana/sana-600m/packages/Sana/asset/example_data/00000000_InternVL2-26B_clip_score.json
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json b/image/sana/sana-600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json
rename to image/sana/sana-600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B.json
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json b/image/sana/sana-600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json
rename to image/sana/sana-600m/packages/Sana/asset/example_data/00000000_VILA1-5-13B_clip_score.json
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/00000000_prompt_clip_score.json b/image/sana/sana-600m/packages/Sana/asset/example_data/00000000_prompt_clip_score.json
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/00000000_prompt_clip_score.json
rename to image/sana/sana-600m/packages/Sana/asset/example_data/00000000_prompt_clip_score.json
diff --git a/sana/sana_600M/packages/Sana/asset/example_data/meta_data.json b/image/sana/sana-600m/packages/Sana/asset/example_data/meta_data.json
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/example_data/meta_data.json
rename to image/sana/sana-600m/packages/Sana/asset/example_data/meta_data.json
diff --git a/sana/sana_600M/packages/Sana/asset/examples.py b/image/sana/sana-600m/packages/Sana/asset/examples.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/examples.py
rename to image/sana/sana-600m/packages/Sana/asset/examples.py
diff --git a/sana/sana_600M/packages/Sana/asset/logo.png b/image/sana/sana-600m/packages/Sana/asset/logo.png
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/logo.png
rename to image/sana/sana-600m/packages/Sana/asset/logo.png
diff --git a/sana/sana_600M/packages/Sana/asset/model-incremental.jpg b/image/sana/sana-600m/packages/Sana/asset/model-incremental.jpg
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/model-incremental.jpg
rename to image/sana/sana-600m/packages/Sana/asset/model-incremental.jpg
diff --git a/sana/sana_600M/packages/Sana/asset/model_paths.txt b/image/sana/sana-600m/packages/Sana/asset/model_paths.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/model_paths.txt
rename to image/sana/sana-600m/packages/Sana/asset/model_paths.txt
diff --git a/sana/sana_600M/packages/Sana/asset/samples.txt b/image/sana/sana-600m/packages/Sana/asset/samples.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/samples.txt
rename to image/sana/sana-600m/packages/Sana/asset/samples.txt
diff --git a/sana/sana_600M/packages/Sana/asset/samples_mini.txt b/image/sana/sana-600m/packages/Sana/asset/samples_mini.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/asset/samples_mini.txt
rename to image/sana/sana-600m/packages/Sana/asset/samples_mini.txt
diff --git a/sana/sana_600M/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_app_config/Sana_1600M_app.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_app_config/Sana_600M_app.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_base.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_base.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_base.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_base.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_config/1024ms/Sana_1600M_img1024_AdamW.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_config/1024ms/Sana_600M_img1024.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/Sana_1600M_img512.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/Sana_600M_img512.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/ci_Sana_600M_img512.yaml
diff --git a/sana/sana_600M/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml b/image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml
similarity index 100%
rename from sana/sana_600M/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml
rename to image/sana/sana-600m/packages/Sana/configs/sana_config/512ms/sample_dataset.yaml
diff --git a/sana/sana_600M/packages/Sana/diffusion/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/data/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/builder.py b/image/sana/sana-600m/packages/Sana/diffusion/data/builder.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/builder.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/builder.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/datasets/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/data/datasets/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/datasets/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/datasets/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/datasets/sana_data.py b/image/sana/sana-600m/packages/Sana/diffusion/data/datasets/sana_data.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/datasets/sana_data.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/datasets/sana_data.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py b/image/sana/sana-600m/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/datasets/sana_data_multi_scale.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/datasets/utils.py b/image/sana/sana-600m/packages/Sana/diffusion/data/datasets/utils.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/datasets/utils.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/datasets/utils.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/transforms.py b/image/sana/sana-600m/packages/Sana/diffusion/data/transforms.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/transforms.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/transforms.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/wids/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/data/wids/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/wids/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/wids/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/wids/wids.py b/image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/wids/wids.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/wids/wids_dl.py b/image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_dl.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/wids/wids_dl.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_dl.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/wids/wids_lru.py b/image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_lru.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/wids/wids_lru.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_lru.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/wids/wids_mmtar.py b/image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_mmtar.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/wids/wids_mmtar.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_mmtar.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/wids/wids_specs.py b/image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_specs.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/wids/wids_specs.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_specs.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/data/wids/wids_tar.py b/image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_tar.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/data/wids/wids_tar.py
rename to image/sana/sana-600m/packages/Sana/diffusion/data/wids/wids_tar.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/dpm_solver.py b/image/sana/sana-600m/packages/Sana/diffusion/dpm_solver.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/dpm_solver.py
rename to image/sana/sana-600m/packages/Sana/diffusion/dpm_solver.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/flow_euler_sampler.py b/image/sana/sana-600m/packages/Sana/diffusion/flow_euler_sampler.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/flow_euler_sampler.py
rename to image/sana/sana-600m/packages/Sana/diffusion/flow_euler_sampler.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/iddpm.py b/image/sana/sana-600m/packages/Sana/diffusion/iddpm.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/iddpm.py
rename to image/sana/sana-600m/packages/Sana/diffusion/iddpm.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/lcm_scheduler.py b/image/sana/sana-600m/packages/Sana/diffusion/lcm_scheduler.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/lcm_scheduler.py
rename to image/sana/sana-600m/packages/Sana/diffusion/lcm_scheduler.py
diff --git a/llama/llama-2-7b/model/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/__init__.py
old mode 100644
new mode 100755
similarity index 100%
rename from llama/llama-2-7b/model/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/act.py b/image/sana/sana-600m/packages/Sana/diffusion/model/act.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/act.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/act.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/builder.py b/image/sana/sana-600m/packages/Sana/diffusion/model/builder.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/builder.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/builder.py
diff --git a/llama/llama-3-70b-instruct/model/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py
similarity index 100%
rename from llama/llama-3-70b-instruct/model/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/ae_model_zoo.py
diff --git a/llama/llama-3-8b-instruct/model/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py
similarity index 100%
rename from llama/llama-3-8b-instruct/model/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/setup.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/trainer/run_config.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/dist.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/ema.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/export.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/image.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/init.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/lr.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/metric.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/misc.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/utils/opt.py
diff --git a/llama/llama-3_1-405b-instruct/model/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py
similarity index 100%
rename from llama/llama-3_1-405b-instruct/model/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/efficientvit/dc_ae.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/act.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/drop.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/norm.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/ops.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/nn/triton_rms_norm.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/list.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/network.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/dc_ae/efficientvit/models/utils/random.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/diffusion_utils.py b/image/sana/sana-600m/packages/Sana/diffusion/model/diffusion_utils.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/diffusion_utils.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/diffusion_utils.py
diff --git a/image/sana/sana-600m/packages/Sana/diffusion/model/dpm_solver.py b/image/sana/sana-600m/packages/Sana/diffusion/model/dpm_solver.py
new file mode 100755
index 000000000..a791dcd38
--- /dev/null
+++ b/image/sana/sana-600m/packages/Sana/diffusion/model/dpm_solver.py
@@ -0,0 +1,1908 @@
+# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# This file is modified from https://github.com/PixArt-alpha/PixArt-sigma
+import os
+
+import torch
+from tqdm import tqdm
+
+from .nets.sana_blocks import (
+    PAGCFGIdentitySelfAttnProcessorLiteLA,
+    PAGIdentitySelfAttnProcessorLiteLA,
+    SelfAttnProcessorLiteLA,
+)
+
+
+class NoiseScheduleVP:
+    def __init__(
+        self,
+        schedule="discrete",
+        betas=None,
+        alphas_cumprod=None,
+        continuous_beta_0=0.1,
+        continuous_beta_1=20.0,
+        dtype=torch.float32,
+    ):
+        r"""Create a wrapper class for the forward SDE (VP type).
+
+        ***
+        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
+                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
+        ***
+
+        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
+        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
+        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
+
+            log_alpha_t = self.marginal_log_mean_coeff(t)
+            sigma_t = self.marginal_std(t)
+            lambda_t = self.marginal_lambda(t)
+
+        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
+
+            t = self.inverse_lambda(lambda_t)
+
+        ===============================================================
+
+        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
+
+        1. For discrete-time DPMs:
+
+            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
+                t_i = (i + 1) / N
+            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
+            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
+
+            Args:
+                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
+                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
+
+            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
+
+            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
+                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
+                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
+                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
+                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
+                and
+                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
+
+
+        2. For continuous-time DPMs:
+
+            We support the linear VPSDE for the continuous time setting. The hyperparameters for the noise
+            schedule are the default settings in Yang Song's ScoreSDE:
+
+            Args:
+                beta_min: A `float` number. The smallest beta for the linear schedule.
+                beta_max: A `float` number. The largest beta for the linear schedule.
+                T: A `float` number. The ending time of the forward process.
+
+        ===============================================================
+
+        Args:
+            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
+                    'linear' for continuous-time DPMs.
+        Returns:
+            A wrapper object of the forward SDE (VP type).
+
+        ===============================================================
+
+        Example:
+
+        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', betas=betas)
+
+        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
+
+        # For continuous-time DPMs (VPSDE), linear schedule:
+        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
+
+        """
+
+        if schedule not in ["discrete", "linear"]:
+            raise ValueError(
+                f"Unsupported noise schedule {schedule}. The schedule needs to be 'discrete' or 'linear'"
+            )
+
+        self.schedule = schedule
+        if schedule == "discrete":
+            if betas is not None:
+                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
+            else:
+                assert alphas_cumprod is not None
+                log_alphas = 0.5 * torch.log(alphas_cumprod)
+            self.T = 1.0
+            self.log_alpha_array = (
+                self.numerical_clip_alpha(log_alphas)
+                .reshape(
+                    (
+                        1,
+                        -1,
+                    )
+                )
+                .to(dtype=dtype)
+            )
+            self.total_N = self.log_alpha_array.shape[1]
+            self.t_array = (
+                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
+                .reshape((1, -1))
+                .to(dtype=dtype)
+            )
+        else:
+            self.T = 1.0
+            self.total_N = 1000
+            self.beta_0 = continuous_beta_0
+            self.beta_1 = continuous_beta_1
+
+    def numerical_clip_alpha(self, log_alphas, clipped_lambda=-5.1):
+        """
+        For some beta schedules such as cosine schedule, the log-SNR has numerical isssues.
+        We clip the log-SNR near t=T within -5.1 to ensure the stability.
+        Such a trick is very useful for diffusion models with the cosine schedule, such as i-DDPM, guided-diffusion and GLIDE.
+        """
+        log_sigmas = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_alphas))
+        lambs = log_alphas - log_sigmas
+        idx = torch.searchsorted(torch.flip(lambs, [0]), clipped_lambda)
+        if idx > 0:
+            log_alphas = log_alphas[:-idx]
+        return log_alphas
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        if self.schedule == "discrete":
+            return interpolate_fn(
+                t.reshape((-1, 1)),
+                self.t_array.to(t.device),
+                self.log_alpha_array.to(t.device),
+            ).reshape(-1)
+        elif self.schedule == "linear":
+            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return torch.exp(self.marginal_log_mean_coeff(t))
+
+    def marginal_std(self, t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
+        return log_mean_coeff - log_std
+
+    def inverse_lambda(self, lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        if self.schedule == "linear":
+            tmp = (
+                2.0
+                * (self.beta_1 - self.beta_0)
+                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
+            )
+            Delta = self.beta_0**2 + tmp
+            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
+        elif self.schedule == "discrete":
+            log_alpha = -0.5 * torch.logaddexp(
+                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
+            )
+            t = interpolate_fn(
+                log_alpha.reshape((-1, 1)),
+                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
+                torch.flip(self.t_array.to(lamb.device), [1]),
+            )
+            return t.reshape((-1,))
+
+
+class NoiseScheduleFlow:
+    def __init__(
+        self,
+        schedule="discrete_flow",
+    ):
+        """Create a wrapper class for the forward SDE (EDM type)."""
+        self.T = 1
+        self.t0 = 0.001
+        self.schedule = schedule  # ['continuous', 'discrete_flow']
+        self.total_N = 1000
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        return torch.log(self.marginal_alpha(t))
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return 1 - t
+
+    @staticmethod
+    def marginal_std(t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return t
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = torch.log(self.marginal_std(t))
+        return log_mean_coeff - log_std
+
+    @staticmethod
+    def inverse_lambda(lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        return torch.exp(-lamb)
+
+    def edm_sigma(self, t):
+        return self.marginal_std(t) / self.marginal_alpha(t)
+
+    def edm_inverse_sigma(self, edmsigma):
+        sigma = edmsigma
+        lambda_t = torch.log(1 / sigma)
+        t = self.inverse_lambda(lambda_t)
+        return t
+
+
+def model_wrapper(
+    model,
+    noise_schedule,
+    model_type="noise",
+    model_kwargs={},
+    guidance_type="uncond",
+    condition=None,
+    unconditional_condition=None,
+    guidance_scale=1.0,
+    pag_scale=1.0,
+    pag_applied_layers=[],
+    interval_guidance=[0, 1.0],
+    classifier_fn=None,
+    classifier_kwargs={},
+):
+    """Create a wrapper function for the noise prediction model.
+
+    DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
+    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
+
+    We support four types of the diffusion model by setting `model_type`:
+
+        1. "noise": noise prediction model. (Trained by predicting noise).
+
+        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
+
+        3. "v": velocity prediction model. (Trained by predicting the velocity).
+            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
+
+            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
+                arXiv preprint arXiv:2202.00512 (2022).
+            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
+                arXiv preprint arXiv:2210.02303 (2022).
+
+        4. "score": marginal score function. (Trained by denoising score matching).
+            Note that the score function and the noise prediction model follows a simple relationship:
+            ```
+                noise(x_t, t) = -sigma_t * score(x_t, t)
+            ```
+
+    We support three types of guided sampling by DPMs by setting `guidance_type`:
+        1. "uncond": unconditional sampling by DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+
+        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+
+            The input `classifier_fn` has the following format:
+            ``
+                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
+            ``
+
+            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
+                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
+
+        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
+            ``
+            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
+
+            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
+                arXiv preprint arXiv:2207.12598 (2022).
+
+
+    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
+    or continuous-time labels (i.e. epsilon to T).
+
+    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
+    ``
+        def model_fn(x, t_continuous) -> noise:
+            t_input = get_model_input_time(t_continuous)
+            return noise_pred(model, x, t_input, **model_kwargs)
+    ``
+    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
+
+    ===============================================================
+
+    Args:
+        model: A diffusion model with the corresponding format described above.
+        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+        model_type: A `str`. The parameterization type of the diffusion model.
+                    "noise" or "x_start" or "v" or "score".
+        model_kwargs: A `dict`. A dict for the other inputs of the model function.
+        guidance_type: A `str`. The type of the guidance for sampling.
+                    "uncond" or "classifier" or "classifier-free".
+        condition: A pytorch tensor. The condition for the guided sampling.
+                    Only used for "classifier" or "classifier-free" guidance type.
+        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
+                    Only used for "classifier-free" guidance type.
+        guidance_scale: A `float`. The scale for the guided sampling.
+        classifier_fn: A classifier function. Only used for the classifier guidance.
+        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
+    Returns:
+        A noise prediction model that accepts the noised data and the continuous time as the inputs.
+    """
+
+    def get_model_input_time(t_continuous):
+        """
+        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
+        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
+        For continuous-time DPMs, we just use `t_continuous`.
+        """
+        if noise_schedule.schedule == "discrete":
+            return (
+                t_continuous - 1.0 / noise_schedule.total_N
+            ) * noise_schedule.total_N
+        elif noise_schedule.schedule == "discrete_flow":
+            return t_continuous * noise_schedule.total_N
+        else:
+            return t_continuous
+
+    def noise_pred_fn(x, t_continuous, cond=None):
+        t_input = get_model_input_time(t_continuous)
+        if cond is None:
+            output = model(x, t_input, **model_kwargs)
+        else:
+            output = model(x, t_input, cond, **model_kwargs)
+        if model_type == "noise":
+            return output
+        elif model_type == "x_start":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return (x - expand_dims(alpha_t, x.dim()) * output) / expand_dims(
+                sigma_t, x.dim()
+            )
+        elif model_type == "v":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return (
+                expand_dims(alpha_t, x.dim()) * output
+                + expand_dims(sigma_t, x.dim()) * x
+            )
+        elif model_type == "score":
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            return -expand_dims(sigma_t, x.dim()) * output
+        elif model_type == "flow":
+            _, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            try:
+                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output + x
+            except:
+                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output[0] + x
+            return noise
+
+    def cond_grad_fn(x, t_input):
+        """
+        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
+        """
+        with torch.enable_grad():
+            x_in = x.detach().requires_grad_(True)
+            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
+            return torch.autograd.grad(log_prob.sum(), x_in)[0]
+
+    def model_fn(x, t_continuous):
+        """
+        The noise predicition model function that is used for DPM-Solver.
+        """
+        guidance_tp = guidance_type
+        if guidance_tp == "uncond":
+            return noise_pred_fn(x, t_continuous)
+        elif guidance_tp == "classifier":
+            assert classifier_fn is not None
+            t_input = get_model_input_time(t_continuous)
+            cond_grad = cond_grad_fn(x, t_input)
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            noise = noise_pred_fn(x, t_continuous)
+            return noise - guidance_scale * expand_dims(sigma_t, x.dim()) * cond_grad
+        elif guidance_tp == "classifier-free":
+            if (
+                guidance_scale == 1.0
+                or unconditional_condition is None
+                or not (interval_guidance[0] < t_continuous[0] < interval_guidance[1])
+            ):
+                return noise_pred_fn(x, t_continuous, cond=condition)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t_continuous] * 2)
+                c_in = torch.cat([unconditional_condition, condition])
+                try:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                except:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
+                        2
+                    )
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+        elif guidance_tp == "classifier-free_PAG":
+            for i in pag_applied_layers:
+                if isinstance(model, torch.nn.Module):
+                    model.blocks[i].attn.forward = (
+                        PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
+                        if guidance_scale == 1.0
+                        else PAGCFGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
+                    )
+                else:
+                    model.__self__.blocks[i].attn.forward = (
+                        PAGIdentitySelfAttnProcessorLiteLA(
+                            model.__self__.blocks[i].attn
+                        )
+                        if guidance_scale == 1.0
+                        else PAGCFGIdentitySelfAttnProcessorLiteLA(
+                            model.__self__.blocks[i].attn
+                        )
+                    )
+            num_inputs = 2 if guidance_scale == 1.0 else 3
+            x_in = torch.cat([x] * num_inputs)
+            t_in = torch.cat([t_continuous] * num_inputs)
+            c_in = torch.cat(
+                [condition, condition]
+                if guidance_scale == 1.0
+                else [unconditional_condition, condition, condition]
+            )
+
+            try:
+                chunks = noise_pred_fn(x_in, t_in, cond=c_in).chunk(num_inputs)
+            except:
+                chunks = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(num_inputs)
+
+            if guidance_scale == 1.0:
+                noise, noise_perturb = chunks
+                noise_pred = noise + pag_scale * (noise - noise_perturb)
+            else:
+                noise_uncond, noise, noise_perturb = chunks
+                noise_pred = (
+                    noise_uncond
+                    + guidance_scale * (noise - noise_uncond)
+                    + pag_scale * (noise - noise_perturb)
+                )
+            for i in pag_applied_layers:
+                if isinstance(model, torch.nn.Module):
+                    model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                        model.blocks[i].attn
+                    )
+                else:
+                    model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                        model.__self__.blocks[i].attn
+                    )
+
+            return noise_pred
+        elif guidance_tp == "classifier-free_PAG_seq":
+            num_inputs = 2
+            if t_continuous[0] < 0.5:
+                # cfg
+                if (
+                    guidance_scale == 1.0
+                    or unconditional_condition is None
+                    or not (
+                        interval_guidance[0] < t_continuous[0] < interval_guidance[1]
+                    )
+                ):
+                    return noise_pred_fn(x, t_continuous, cond=condition)
+
+                x_in = torch.cat([x] * num_inputs)
+                t_in = torch.cat([t_continuous] * num_inputs)
+                c_in = torch.cat([unconditional_condition, condition])
+
+                try:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                except:
+                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
+                        num_inputs
+                    )
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+            else:
+                # pag
+                for i in pag_applied_layers:
+                    if isinstance(model, torch.nn.Module):
+                        model.blocks[i].attn.forward = (
+                            PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
+                            if guidance_scale == 1.0
+                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
+                                model.blocks[i].attn
+                            )
+                        )
+                    else:
+                        model.__self__.blocks[i].attn.forward = (
+                            PAGIdentitySelfAttnProcessorLiteLA(
+                                model.__self__.blocks[i].attn
+                            )
+                            if guidance_scale == 1.0
+                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
+                                model.__self__.blocks[i].attn
+                            )
+                        )
+                x_in = torch.cat([x] * 3)
+                t_in = torch.cat([t_continuous] * 3)
+                c_in = torch.cat([unconditional_condition, condition, condition])
+
+                try:
+                    noise_uncond, noise, noise_perturb = noise_pred_fn(
+                        x_in, t_in, cond=c_in
+                    ).chunk(3)
+                except:
+                    noise_uncond, noise, noise_perturb = noise_pred_fn(
+                        x_in, t_in, cond=c_in
+                    )[0].chunk(3)
+
+                for i in pag_applied_layers:
+                    if isinstance(model, torch.nn.Module):
+                        model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                            model.blocks[i].attn
+                        )
+                    else:
+                        model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
+                            model.__self__.blocks[i].attn
+                        )
+
+                return (
+                    noise_uncond
+                    + guidance_scale * (noise - noise_uncond)
+                    + pag_scale * (noise - noise_perturb)
+                )
+
+    assert model_type in ["noise", "x_start", "v", "score", "flow"]
+    assert guidance_type in [
+        "uncond",
+        "classifier",
+        "classifier-free",
+        "classifier-free_PAG",
+        "classifier-free_PAG_seq",
+    ]
+    return model_fn
+
+
+class DPM_Solver:
+    def __init__(
+        self,
+        model_fn,
+        noise_schedule,
+        algorithm_type="dpmsolver++",
+        correcting_x0_fn=None,
+        correcting_xt_fn=None,
+        thresholding_max_val=1.0,
+        dynamic_thresholding_ratio=0.995,
+    ):
+        """Construct a DPM-Solver.
+
+        We support both DPM-Solver (`algorithm_type="dpmsolver"`) and DPM-Solver++ (`algorithm_type="dpmsolver++"`).
+
+        We also support the "dynamic thresholding" method in Imagen[1]. For pixel-space diffusion models, you
+        can set both `algorithm_type="dpmsolver++"` and `correcting_x0_fn="dynamic_thresholding"` to use the
+        dynamic thresholding. The "dynamic thresholding" can greatly improve the sample quality for pixel-space
+        DPMs with large guidance scales. Note that the thresholding method is **unsuitable** for latent-space
+        DPMs (such as stable-diffusion).
+
+        To support advanced algorithms in image-to-image applications, we also support corrector functions for
+        both x0 and xt.
+
+        Args:
+            model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
+                ``
+                def model_fn(x, t_continuous):
+                    return noise
+                ``
+                The shape of `x` is `(batch_size, **shape)`, and the shape of `t_continuous` is `(batch_size,)`.
+            noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+            algorithm_type: A `str`. Either "dpmsolver" or "dpmsolver++".
+            correcting_x0_fn: A `str` or a function with the following format:
+                ```
+                def correcting_x0_fn(x0, t):
+                    x0_new = ...
+                    return x0_new
+                ```
+                This function is to correct the outputs of the data prediction model at each sampling step. e.g.,
+                ```
+                x0_pred = data_pred_model(xt, t)
+                if correcting_x0_fn is not None:
+                    x0_pred = correcting_x0_fn(x0_pred, t)
+                xt_1 = update(x0_pred, xt, t)
+                ```
+                If `correcting_x0_fn="dynamic_thresholding"`, we use the dynamic thresholding proposed in Imagen[1].
+            correcting_xt_fn: A function with the following format:
+                ```
+                def correcting_xt_fn(xt, t, step):
+                    x_new = ...
+                    return x_new
+                ```
+                This function is to correct the intermediate samples xt at each sampling step. e.g.,
+                ```
+                xt = ...
+                xt = correcting_xt_fn(xt, t, step)
+                ```
+            thresholding_max_val: A `float`. The max value for thresholding.
+                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
+            dynamic_thresholding_ratio: A `float`. The ratio for dynamic thresholding (see Imagen[1] for details).
+                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
+
+        [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour,
+            Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models
+            with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
+        """
+        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
+        self.noise_schedule = noise_schedule
+        assert algorithm_type in ["dpmsolver", "dpmsolver++"]
+        self.algorithm_type = algorithm_type
+        if correcting_x0_fn == "dynamic_thresholding":
+            self.correcting_x0_fn = self.dynamic_thresholding_fn
+        else:
+            self.correcting_x0_fn = correcting_x0_fn
+        self.correcting_xt_fn = correcting_xt_fn
+        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
+        self.thresholding_max_val = thresholding_max_val
+        self.register_progress_bar()
+
+    def register_progress_bar(self, progress_fn=None):
+        """
+        Register a progress bar callback function
+
+        Args:
+            progress_fn: Callback function that takes current step and total steps as parameters
+        """
+        self.progress_fn = (
+            progress_fn if progress_fn is not None else lambda step, total: None
+        )
+
+    def update_progress(self, step, total_steps):
+        """
+        Update sampling progress
+
+        Args:
+            step: Current step number
+            total_steps: Total number of steps
+        """
+        if hasattr(self, "progress_fn"):
+            try:
+                self.progress_fn(
+                    step / total_steps, desc=f"Generating {step}/{total_steps}"
+                )
+            except:
+                self.progress_fn(step, total_steps)
+
+        else:
+            # If no progress_fn registered, use default empty function
+            pass
+
+    def dynamic_thresholding_fn(self, x0, t):
+        """
+        The dynamic thresholding method.
+        """
+        dims = x0.dim()
+        p = self.dynamic_thresholding_ratio
+        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
+        s = expand_dims(
+            torch.maximum(
+                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
+            ),
+            dims,
+        )
+        x0 = torch.clamp(x0, -s, s) / s
+        return x0
+
+    def noise_prediction_fn(self, x, t):
+        """
+        Return the noise prediction model.
+        """
+        return self.model(x, t)
+
+    def data_prediction_fn(self, x, t):
+        """
+        Return the data prediction model (with corrector).
+        """
+        noise = self.noise_prediction_fn(x, t)
+        alpha_t, sigma_t = (
+            self.noise_schedule.marginal_alpha(t),
+            self.noise_schedule.marginal_std(t),
+        )
+        x0 = (x - sigma_t * noise) / alpha_t
+        if self.correcting_x0_fn is not None:
+            x0 = self.correcting_x0_fn(x0, t)
+        return x0
+
+    def model_fn(self, x, t):
+        """
+        Convert the model to the noise prediction model or the data prediction model.
+        """
+        if self.algorithm_type == "dpmsolver++":
+            return self.data_prediction_fn(x, t)
+        else:
+            return self.noise_prediction_fn(x, t)
+
+    def get_time_steps(self, skip_type, t_T, t_0, N, device, shift=1.0):
+        """Compute the intermediate time steps for sampling.
+
+        Args:
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            N: A `int`. The total number of the spacing of the time steps.
+            device: A torch device.
+        Returns:
+            A pytorch tensor of the time steps, with the shape (N + 1,).
+        """
+        if skip_type == "logSNR":
+            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
+            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
+            logSNR_steps = torch.linspace(
+                lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1
+            ).to(device)
+            return self.noise_schedule.inverse_lambda(logSNR_steps)
+        elif skip_type == "time_uniform":
+            return torch.linspace(t_T, t_0, N + 1).to(device)
+        elif skip_type == "time_quadratic":
+            t_order = 2
+            t = (
+                torch.linspace(t_T ** (1.0 / t_order), t_0 ** (1.0 / t_order), N + 1)
+                .pow(t_order)
+                .to(device)
+            )
+            return t
+        elif skip_type == "time_uniform_flow":
+            betas = torch.linspace(t_T, t_0, N + 1).to(device)
+            sigmas = 1.0 - betas
+            sigmas = (shift * sigmas / (1 + (shift - 1) * sigmas)).flip(dims=[0])
+            return sigmas
+        else:
+            raise ValueError(
+                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'"
+            )
+
+    def get_orders_and_timesteps_for_singlestep_solver(
+        self, steps, order, skip_type, t_T, t_0, device
+    ):
+        """
+        Get the order of each step for sampling by the singlestep DPM-Solver.
+
+        We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
+        Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
+            - If order == 1:
+                We take `steps` of DPM-Solver-1 (i.e. DDIM).
+            - If order == 2:
+                - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
+                - If steps % 2 == 0, we use K steps of DPM-Solver-2.
+                - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
+            - If order == 3:
+                - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
+                - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
+
+        ============================================
+        Args:
+            order: A `int`. The max order for the solver (2 or 3).
+            steps: A `int`. The total number of function evaluations (NFE).
+            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
+                - 'logSNR': uniform logSNR for the time steps.
+                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
+                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            device: A torch device.
+        Returns:
+            orders: A list of the solver order of each step.
+        """
+        if order == 3:
+            K = steps // 3 + 1
+            if steps % 3 == 0:
+                orders = [
+                    3,
+                ] * (K - 2) + [2, 1]
+            elif steps % 3 == 1:
+                orders = [
+                    3,
+                ] * (K - 1) + [1]
+            else:
+                orders = [
+                    3,
+                ] * (K - 1) + [2]
+        elif order == 2:
+            if steps % 2 == 0:
+                K = steps // 2
+                orders = [
+                    2,
+                ] * K
+            else:
+                K = steps // 2 + 1
+                orders = [
+                    2,
+                ] * (K - 1) + [1]
+        elif order == 1:
+            K = 1
+            orders = [
+                1,
+            ] * steps
+        else:
+            raise ValueError("'order' must be '1' or '2' or '3'.")
+        if skip_type == "logSNR":
+            # To reproduce the results in DPM-Solver paper
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
+        else:
+            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
+                torch.cumsum(
+                    torch.tensor(
+                        [
+                            0,
+                        ]
+                        + orders
+                    ),
+                    0,
+                ).to(device)
+            ]
+        return timesteps_outer, orders
+
+    def denoise_to_zero_fn(self, x, s):
+        """
+        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
+        """
+        return self.data_prediction_fn(x, s)
+
+    def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
+        """
+        DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        dims = x.dim()
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        log_alpha_s, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = sigma_t / sigma_s * x - alpha_t * phi_1 * model_s
+            if return_intermediate:
+                return x_t, {"model_s": model_s}
+            else:
+                return x_t
+        else:
+            phi_1 = torch.expm1(h)
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_t = torch.exp(log_alpha_t - log_alpha_s) * x - (sigma_t * phi_1) * model_s
+            if return_intermediate:
+                return x_t, {"model_s": model_s}
+            else:
+                return x_t
+
+    def singlestep_dpm_solver_second_update(
+        self,
+        x,
+        s,
+        t,
+        r1=0.5,
+        model_s=None,
+        return_intermediate=False,
+        solver_type="dpmsolver",
+    ):
+        """
+        Singlestep solver DPM-Solver-2 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            r1: A `float`. The hyperparameter of the second-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(
+                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
+            )
+        if r1 is None:
+            r1 = 0.5
+        ns = self.noise_schedule
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        log_alpha_s, log_alpha_s1, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(s1),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_s1, sigma_t = (
+            ns.marginal_std(s),
+            ns.marginal_std(s1),
+            ns.marginal_std(t),
+        )
+        alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_11 = torch.expm1(-r1 * h)
+            phi_1 = torch.expm1(-h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    - (0.5 / r1) * (alpha_t * phi_1) * (model_s1 - model_s)
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (1.0 / r1) * (alpha_t * (phi_1 / h + 1.0)) * (model_s1 - model_s)
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_1 = torch.expm1(h)
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            x_s1 = (
+                torch.exp(log_alpha_s1 - log_alpha_s) * x
+                - (sigma_s1 * phi_11) * model_s
+            )
+            model_s1 = self.model_fn(x_s1, s1)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    torch.exp(log_alpha_t - log_alpha_s) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (0.5 / r1) * (sigma_t * phi_1) * (model_s1 - model_s)
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    torch.exp(log_alpha_t - log_alpha_s) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (1.0 / r1) * (sigma_t * (phi_1 / h - 1.0)) * (model_s1 - model_s)
+                )
+        if return_intermediate:
+            return x_t, {"model_s": model_s, "model_s1": model_s1}
+        else:
+            return x_t
+
+    def singlestep_dpm_solver_third_update(
+        self,
+        x,
+        s,
+        t,
+        r1=1.0 / 3.0,
+        r2=2.0 / 3.0,
+        model_s=None,
+        model_s1=None,
+        return_intermediate=False,
+        solver_type="dpmsolver",
+    ):
+        """
+        Singlestep solver DPM-Solver-3 from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            r1: A `float`. The hyperparameter of the third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+            model_s: A pytorch tensor. The model function evaluated at time `s`.
+                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
+            model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
+                If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(
+                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
+            )
+        if r1 is None:
+            r1 = 1.0 / 3.0
+        if r2 is None:
+            r2 = 2.0 / 3.0
+        ns = self.noise_schedule
+        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
+        h = lambda_t - lambda_s
+        lambda_s1 = lambda_s + r1 * h
+        lambda_s2 = lambda_s + r2 * h
+        s1 = ns.inverse_lambda(lambda_s1)
+        s2 = ns.inverse_lambda(lambda_s2)
+        log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = (
+            ns.marginal_log_mean_coeff(s),
+            ns.marginal_log_mean_coeff(s1),
+            ns.marginal_log_mean_coeff(s2),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_s, sigma_s1, sigma_s2, sigma_t = (
+            ns.marginal_std(s),
+            ns.marginal_std(s1),
+            ns.marginal_std(s2),
+            ns.marginal_std(t),
+        )
+        alpha_s1, alpha_s2, alpha_t = (
+            torch.exp(log_alpha_s1),
+            torch.exp(log_alpha_s2),
+            torch.exp(log_alpha_t),
+        )
+
+        if self.algorithm_type == "dpmsolver++":
+            phi_11 = torch.expm1(-r1 * h)
+            phi_12 = torch.expm1(-r2 * h)
+            phi_1 = torch.expm1(-h)
+            phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.0
+            phi_2 = phi_1 / h + 1.0
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                (sigma_s2 / sigma_s) * x
+                - (alpha_s2 * phi_12) * model_s
+                + r2 / r1 * (alpha_s2 * phi_22) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (1.0 / r2) * (alpha_t * phi_2) * (model_s2 - model_s)
+                )
+            elif solver_type == "taylor":
+                D1_0 = (1.0 / r1) * (model_s1 - model_s)
+                D1_1 = (1.0 / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    (sigma_t / sigma_s) * x
+                    - (alpha_t * phi_1) * model_s
+                    + (alpha_t * phi_2) * D1
+                    - (alpha_t * phi_3) * D2
+                )
+        else:
+            phi_11 = torch.expm1(r1 * h)
+            phi_12 = torch.expm1(r2 * h)
+            phi_1 = torch.expm1(h)
+            phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.0
+            phi_2 = phi_1 / h - 1.0
+            phi_3 = phi_2 / h - 0.5
+
+            if model_s is None:
+                model_s = self.model_fn(x, s)
+            if model_s1 is None:
+                x_s1 = (torch.exp(log_alpha_s1 - log_alpha_s)) * x - (
+                    sigma_s1 * phi_11
+                ) * model_s
+                model_s1 = self.model_fn(x_s1, s1)
+            x_s2 = (
+                (torch.exp(log_alpha_s2 - log_alpha_s)) * x
+                - (sigma_s2 * phi_12) * model_s
+                - r2 / r1 * (sigma_s2 * phi_22) * (model_s1 - model_s)
+            )
+            model_s2 = self.model_fn(x_s2, s2)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_s)) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (1.0 / r2) * (sigma_t * phi_2) * (model_s2 - model_s)
+                )
+            elif solver_type == "taylor":
+                D1_0 = (1.0 / r1) * (model_s1 - model_s)
+                D1_1 = (1.0 / r2) * (model_s2 - model_s)
+                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
+                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_s)) * x
+                    - (sigma_t * phi_1) * model_s
+                    - (sigma_t * phi_2) * D1
+                    - (sigma_t * phi_3) * D2
+                )
+
+        if return_intermediate:
+            return x_t, {"model_s": model_s, "model_s1": model_s1, "model_s2": model_s2}
+        else:
+            return x_t
+
+    def multistep_dpm_solver_second_update(
+        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
+    ):
+        """
+        Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if solver_type not in ["dpmsolver", "taylor"]:
+            raise ValueError(
+                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
+            )
+        ns = self.noise_schedule
+        model_prev_1, model_prev_0 = model_prev_list[-2], model_prev_list[-1]
+        t_prev_1, t_prev_0 = t_prev_list[-2], t_prev_list[-1]
+        lambda_prev_1, lambda_prev_0, lambda_t = (
+            ns.marginal_lambda(t_prev_1),
+            ns.marginal_lambda(t_prev_0),
+            ns.marginal_lambda(t),
+        )
+        log_alpha_prev_0, log_alpha_t = (
+            ns.marginal_log_mean_coeff(t_prev_0),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0 = h_0 / h
+        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (sigma_t / sigma_prev_0) * x
+                    - (alpha_t * phi_1) * model_prev_0
+                    - 0.5 * (alpha_t * phi_1) * D1_0
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (sigma_t / sigma_prev_0) * x
+                    - (alpha_t * phi_1) * model_prev_0
+                    + (alpha_t * (phi_1 / h + 1.0)) * D1_0
+                )
+        else:
+            phi_1 = torch.expm1(h)
+            if solver_type == "dpmsolver":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                    - (sigma_t * phi_1) * model_prev_0
+                    - 0.5 * (sigma_t * phi_1) * D1_0
+                )
+            elif solver_type == "taylor":
+                x_t = (
+                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                    - (sigma_t * phi_1) * model_prev_0
+                    - (sigma_t * (phi_1 / h - 1.0)) * D1_0
+                )
+        return x_t
+
+    def multistep_dpm_solver_third_update(
+        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
+    ):
+        """
+        Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        ns = self.noise_schedule
+        model_prev_2, model_prev_1, model_prev_0 = model_prev_list
+        t_prev_2, t_prev_1, t_prev_0 = t_prev_list
+        lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = (
+            ns.marginal_lambda(t_prev_2),
+            ns.marginal_lambda(t_prev_1),
+            ns.marginal_lambda(t_prev_0),
+            ns.marginal_lambda(t),
+        )
+        log_alpha_prev_0, log_alpha_t = (
+            ns.marginal_log_mean_coeff(t_prev_0),
+            ns.marginal_log_mean_coeff(t),
+        )
+        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
+        alpha_t = torch.exp(log_alpha_t)
+
+        h_1 = lambda_prev_1 - lambda_prev_2
+        h_0 = lambda_prev_0 - lambda_prev_1
+        h = lambda_t - lambda_prev_0
+        r0, r1 = h_0 / h, h_1 / h
+        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
+        D1_1 = (1.0 / r1) * (model_prev_1 - model_prev_2)
+        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
+        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
+        if self.algorithm_type == "dpmsolver++":
+            phi_1 = torch.expm1(-h)
+            phi_2 = phi_1 / h + 1.0
+            phi_3 = phi_2 / h - 0.5
+            x_t = (
+                (sigma_t / sigma_prev_0) * x
+                - (alpha_t * phi_1) * model_prev_0
+                + (alpha_t * phi_2) * D1
+                - (alpha_t * phi_3) * D2
+            )
+        else:
+            phi_1 = torch.expm1(h)
+            phi_2 = phi_1 / h - 1.0
+            phi_3 = phi_2 / h - 0.5
+            x_t = (
+                (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
+                - (sigma_t * phi_1) * model_prev_0
+                - (sigma_t * phi_2) * D1
+                - (sigma_t * phi_3) * D2
+            )
+        return x_t
+
+    def singlestep_dpm_solver_update(
+        self,
+        x,
+        s,
+        t,
+        order,
+        return_intermediate=False,
+        solver_type="dpmsolver",
+        r1=None,
+        r2=None,
+    ):
+        """
+        Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            s: A pytorch tensor. The starting time, with the shape (1,).
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+            r1: A `float`. The hyperparameter of the second-order or third-order solver.
+            r2: A `float`. The hyperparameter of the third-order solver.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(
+                x, s, t, return_intermediate=return_intermediate
+            )
+        elif order == 2:
+            return self.singlestep_dpm_solver_second_update(
+                x,
+                s,
+                t,
+                return_intermediate=return_intermediate,
+                solver_type=solver_type,
+                r1=r1,
+            )
+        elif order == 3:
+            return self.singlestep_dpm_solver_third_update(
+                x,
+                s,
+                t,
+                return_intermediate=return_intermediate,
+                solver_type=solver_type,
+                r1=r1,
+                r2=r2,
+            )
+        else:
+            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
+
+    def multistep_dpm_solver_update(
+        self, x, model_prev_list, t_prev_list, t, order, solver_type="dpmsolver"
+    ):
+        """
+        Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `s`.
+            model_prev_list: A list of pytorch tensor. The previous computed model values.
+            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
+            t: A pytorch tensor. The ending time, with the shape (1,).
+            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_t: A pytorch tensor. The approximated solution at time `t`.
+        """
+        if order == 1:
+            return self.dpm_solver_first_update(
+                x, t_prev_list[-1], t, model_s=model_prev_list[-1]
+            )
+        elif order == 2:
+            return self.multistep_dpm_solver_second_update(
+                x, model_prev_list, t_prev_list, t, solver_type=solver_type
+            )
+        elif order == 3:
+            return self.multistep_dpm_solver_third_update(
+                x, model_prev_list, t_prev_list, t, solver_type=solver_type
+            )
+        else:
+            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
+
+    def dpm_solver_adaptive(
+        self,
+        x,
+        order,
+        t_T,
+        t_0,
+        h_init=0.05,
+        atol=0.0078,
+        rtol=0.05,
+        theta=0.9,
+        t_err=1e-5,
+        solver_type="dpmsolver",
+    ):
+        """
+        The adaptive step size solver based on singlestep DPM-Solver.
+
+        Args:
+            x: A pytorch tensor. The initial value at time `t_T`.
+            order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
+            t_T: A `float`. The starting time of the sampling (default is T).
+            t_0: A `float`. The ending time of the sampling (default is epsilon).
+            h_init: A `float`. The initial step size (for logSNR).
+            atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
+            rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
+            theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
+            t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
+                current time and `t_0` is less than `t_err`. The default setting is 1e-5.
+            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
+                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
+        Returns:
+            x_0: A pytorch tensor. The approximated solution at time `t_0`.
+
+        [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
+        """
+        ns = self.noise_schedule
+        s = t_T * torch.ones((1,)).to(x)
+        lambda_s = ns.marginal_lambda(s)
+        lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
+        h = h_init * torch.ones_like(s).to(x)
+        x_prev = x
+        nfe = 0
+        if order == 2:
+            r1 = 0.5
+            lower_update = lambda x, s, t: self.dpm_solver_first_update(
+                x, s, t, return_intermediate=True
+            )
+            higher_update = lambda x, s, t, **kwargs: (
+                self.singlestep_dpm_solver_second_update(
+                    x, s, t, r1=r1, solver_type=solver_type, **kwargs
+                )
+            )
+        elif order == 3:
+            r1, r2 = 1.0 / 3.0, 2.0 / 3.0
+            lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(
+                x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type
+            )
+            higher_update = lambda x, s, t, **kwargs: (
+                self.singlestep_dpm_solver_third_update(
+                    x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs
+                )
+            )
+        else:
+            raise ValueError(
+                f"For adaptive step size solver, order must be 2 or 3, got {order}"
+            )
+        while torch.abs(s - t_0).mean() > t_err:
+            t = ns.inverse_lambda(lambda_s + h)
+            x_lower, lower_noise_kwargs = lower_update(x, s, t)
+            x_higher = higher_update(x, s, t, **lower_noise_kwargs)
+            delta = torch.max(
+                torch.ones_like(x).to(x) * atol,
+                rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)),
+            )
+            norm_fn = lambda v: torch.sqrt(
+                torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)
+            )
+            E = norm_fn((x_higher - x_lower) / delta).max()
+            if torch.all(E <= 1.0):
+                x = x_higher
+                s = t
+                x_prev = x_lower
+                lambda_s = ns.marginal_lambda(s)
+            h = torch.min(
+                theta * h * torch.float_power(E, -1.0 / order).float(),
+                lambda_0 - lambda_s,
+            )
+            nfe += order
+        print("adaptive solver nfe", nfe)
+        return x
+
+    def add_noise(self, x, t, noise=None):
+        """
+        Compute the noised input xt = alpha_t * x + sigma_t * noise.
+
+        Args:
+            x: A `torch.Tensor` with shape `(batch_size, *shape)`.
+            t: A `torch.Tensor` with shape `(t_size,)`.
+        Returns:
+            xt with shape `(t_size, batch_size, *shape)`.
+        """
+        alpha_t, sigma_t = (
+            self.noise_schedule.marginal_alpha(t),
+            self.noise_schedule.marginal_std(t),
+        )
+        if noise is None:
+            noise = torch.randn((t.shape[0], *x.shape), device=x.device)
+        x = x.reshape((-1, *x.shape))
+        xt = expand_dims(alpha_t, x.dim()) * x + expand_dims(sigma_t, x.dim()) * noise
+        if t.shape[0] == 1:
+            return xt.squeeze(0)
+        else:
+            return xt
+
+    def inverse(
+        self,
+        x,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        order=2,
+        skip_type="time_uniform",
+        method="multistep",
+        lower_order_final=True,
+        denoise_to_zero=False,
+        solver_type="dpmsolver",
+        atol=0.0078,
+        rtol=0.05,
+        return_intermediate=False,
+    ):
+        """
+        Inverse the sample `x` from time `t_start` to `t_end` by DPM-Solver.
+        For discrete-time DPMs, we use `t_start=1/N`, where `N` is the total time steps during training.
+        """
+        t_0 = 1.0 / self.noise_schedule.total_N if t_start is None else t_start
+        t_T = self.noise_schedule.T if t_end is None else t_end
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+        return self.sample(
+            x,
+            steps=steps,
+            t_start=t_0,
+            t_end=t_T,
+            order=order,
+            skip_type=skip_type,
+            method=method,
+            lower_order_final=lower_order_final,
+            denoise_to_zero=denoise_to_zero,
+            solver_type=solver_type,
+            atol=atol,
+            rtol=rtol,
+            return_intermediate=return_intermediate,
+        )
+
+    def sample(
+        self,
+        x,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        order=2,
+        skip_type="time_uniform",
+        method="multistep",
+        lower_order_final=True,
+        denoise_to_zero=False,
+        solver_type="dpmsolver",
+        atol=0.0078,
+        rtol=0.05,
+        return_intermediate=False,
+        flow_shift=1.0,
+    ):
+        """
+        Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
+
+        =====================================================
+
+        We support the following algorithms for both noise prediction model and data prediction model:
+            - 'singlestep':
+                Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
+                We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
+                The total number of function evaluations (NFE) == `steps`.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    - If `order` == 1:
+                        - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
+                        - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
+                        - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                    - If `order` == 3:
+                        - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
+                        - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
+                        - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
+            - 'multistep':
+                Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
+                We initialize the first `order` values by lower order multistep solvers.
+                Given a fixed NFE == `steps`, the sampling procedure is:
+                    Denote K = steps.
+                    - If `order` == 1:
+                        - We use K steps of DPM-Solver-1 (i.e. DDIM).
+                    - If `order` == 2:
+                        - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
+                    - If `order` == 3:
+                        - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
+            - 'singlestep_fixed':
+                Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
+                We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
+            - 'adaptive':
+                Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
+                We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
+                You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
+                (NFE) and the sample quality.
+                    - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
+                    - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
+
+        =====================================================
+
+        Some advices for choosing the algorithm:
+            - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
+                Use singlestep DPM-Solver or DPM-Solver++ ("DPM-Solver-fast" in the paper) with `order = 3`.
+                e.g., DPM-Solver:
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
+                            skip_type='time_uniform', method='singlestep')
+                e.g., DPM-Solver++:
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
+                            skip_type='time_uniform', method='singlestep')
+            - For **guided sampling with large guidance scale** by DPMs:
+                Use multistep DPM-Solver with `algorithm_type="dpmsolver++"` and `order = 2`.
+                e.g.
+                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
+                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
+                            skip_type='time_uniform', method='multistep')
+
+        We support three types of `skip_type`:
+            - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
+            - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
+            - 'time_quadratic': quadratic time for the time steps.
+
+        =====================================================
+        Args:
+            x: A pytorch tensor. The initial value at time `t_start`
+                e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
+            steps: A `int`. The total number of function evaluations (NFE).
+            t_start: A `float`. The starting time of the sampling.
+                If `T` is None, we use self.noise_schedule.T (default is 1.0).
+            t_end: A `float`. The ending time of the sampling.
+                If `t_end` is None, we use 1. / self.noise_schedule.total_N.
+                e.g. if total_N == 1000, we have `t_end` == 1e-3.
+                For discrete-time DPMs:
+                    - We recommend `t_end` == 1. / self.noise_schedule.total_N.
+                For continuous-time DPMs:
+                    - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
+            order: A `int`. The order of DPM-Solver.
+            skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
+            method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
+            denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
+                Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
+
+                This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
+                score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
+                for diffusion models sampling by diffusion SDEs for low-resolutional images
+                (such as CIFAR-10). However, we observed that such trick does not matter for
+                high-resolutional images. As it needs an additional NFE, we do not recommend
+                it for high-resolutional images.
+            lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
+                Only valid for `method=multistep` and `steps < 15`. We empirically find that
+                this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
+                (especially for steps <= 10). So we recommend to set it to be `True`.
+            solver_type: A `str`. The taylor expansion type for the solver. `dpmsolver` or `taylor`. We recommend `dpmsolver`.
+            atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+            rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
+            return_intermediate: A `bool`. Whether to save the xt at each step.
+                When set to `True`, method returns a tuple (x0, intermediates); when set to False, method returns only x0.
+        Returns:
+            x_end: A pytorch tensor. The approximated solution at time `t_end`.
+
+        """
+        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+        if return_intermediate:
+            assert method in [
+                "multistep",
+                "singlestep",
+                "singlestep_fixed",
+            ], "Cannot use adaptive solver when saving intermediate values"
+        if self.correcting_xt_fn is not None:
+            assert method in [
+                "multistep",
+                "singlestep",
+                "singlestep_fixed",
+            ], "Cannot use adaptive solver when correcting_xt_fn is not None"
+        device = x.device
+        intermediates = []
+        with torch.no_grad():
+            if method == "adaptive":
+                x = self.dpm_solver_adaptive(
+                    x,
+                    order=order,
+                    t_T=t_T,
+                    t_0=t_0,
+                    atol=atol,
+                    rtol=rtol,
+                    solver_type=solver_type,
+                )
+            elif method == "multistep":
+                assert steps >= order
+                timesteps = self.get_time_steps(
+                    skip_type=skip_type,
+                    t_T=t_T,
+                    t_0=t_0,
+                    N=steps,
+                    device=device,
+                    shift=flow_shift,
+                )
+                assert timesteps.shape[0] - 1 == steps
+                # Init the initial values.
+                step = 0
+                t = timesteps[step]
+                t_prev_list = [t]
+                model_prev_list = [self.model_fn(x, t)]
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+                self.update_progress(step + 1, len(timesteps))
+                # Init the first `order` values by lower order multistep DPM-Solver.
+                for step in range(1, order):
+                    t = timesteps[step]
+                    x = self.multistep_dpm_solver_update(
+                        x,
+                        model_prev_list,
+                        t_prev_list,
+                        t,
+                        step,
+                        solver_type=solver_type,
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    t_prev_list.append(t)
+                    model_prev_list.append(self.model_fn(x, t))
+                    # update progress bar
+                    self.update_progress(step + 1, len(timesteps))
+                # Compute the remaining values by `order`-th order multistep DPM-Solver.
+                for step in tqdm(
+                    range(order, steps + 1),
+                    disable=os.getenv("DPM_TQDM", "False") == "True",
+                ):
+                    t = timesteps[step]
+                    # We only use lower order for steps < 10
+                    # if lower_order_final and steps < 10:
+                    if lower_order_final:  # recommended by Shuchen Xue
+                        step_order = min(order, steps + 1 - step)
+                    else:
+                        step_order = order
+                    x = self.multistep_dpm_solver_update(
+                        x,
+                        model_prev_list,
+                        t_prev_list,
+                        t,
+                        step_order,
+                        solver_type=solver_type,
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    for i in range(order - 1):
+                        t_prev_list[i] = t_prev_list[i + 1]
+                        model_prev_list[i] = model_prev_list[i + 1]
+                    t_prev_list[-1] = t
+                    # We do not need to evaluate the final model value.
+                    if step < steps:
+                        model_prev_list[-1] = self.model_fn(x, t)
+                    # update progress bar
+                    self.update_progress(step + 1, len(timesteps))
+            elif method in ["singlestep", "singlestep_fixed"]:
+                if method == "singlestep":
+                    timesteps_outer, orders = (
+                        self.get_orders_and_timesteps_for_singlestep_solver(
+                            steps=steps,
+                            order=order,
+                            skip_type=skip_type,
+                            t_T=t_T,
+                            t_0=t_0,
+                            device=device,
+                        )
+                    )
+                elif method == "singlestep_fixed":
+                    K = steps // order
+                    orders = [
+                        order,
+                    ] * K
+                    timesteps_outer = self.get_time_steps(
+                        skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device
+                    )
+                for step, order in enumerate(orders):
+                    s, t = timesteps_outer[step], timesteps_outer[step + 1]
+                    timesteps_inner = self.get_time_steps(
+                        skip_type=skip_type,
+                        t_T=s.item(),
+                        t_0=t.item(),
+                        N=order,
+                        device=device,
+                    )
+                    lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
+                    h = lambda_inner[-1] - lambda_inner[0]
+                    r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
+                    r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
+                    x = self.singlestep_dpm_solver_update(
+                        x, s, t, order, solver_type=solver_type, r1=r1, r2=r2
+                    )
+                    if self.correcting_xt_fn is not None:
+                        x = self.correcting_xt_fn(x, t, step)
+                    if return_intermediate:
+                        intermediates.append(x)
+                    self.update_progress(step + 1, len(timesteps_outer))
+            else:
+                raise ValueError(f"Got wrong method {method}")
+            if denoise_to_zero:
+                t = torch.ones((1,)).to(device) * t_0
+                x = self.denoise_to_zero_fn(x, t)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step + 1)
+                if return_intermediate:
+                    intermediates.append(x)
+        if return_intermediate:
+            return x, intermediates
+        else:
+            return x
+
+
+#############################################################
+# other utility functions
+#############################################################
+
+
+def interpolate_fn(x, xp, yp):
+    """
+    A piecewise linear function y = f(x), using xp and yp as keypoints.
+    We implement f(x) in a differentiable way (i.e. applicable for autograd).
+    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
+
+    Args:
+        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
+        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
+        yp: PyTorch tensor with shape [C, K].
+    Returns:
+        The function values f(x), with shape [N, C].
+    """
+    N, K = x.shape[0], xp.shape[1]
+    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
+    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
+    x_idx = torch.argmin(x_indices, dim=2)
+    cand_start_idx = x_idx - 1
+    start_idx = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(1, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    end_idx = torch.where(
+        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
+    )
+    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
+    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
+    start_idx2 = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(0, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
+    start_y = torch.gather(
+        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
+    ).squeeze(2)
+    end_y = torch.gather(
+        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
+    ).squeeze(2)
+    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
+    return cand
+
+
+def expand_dims(v, dims):
+    """
+    Expand the tensor `v` to the dim `dims`.
+
+    Args:
+        `v`: a PyTorch tensor with shape [N].
+        `dim`: a `int`.
+    Returns:
+        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
+    """
+    return v[(...,) + (None,) * (dims - 1)]
diff --git a/image/sana/sana-600m/packages/Sana/diffusion/model/edm_sample.py b/image/sana/sana-600m/packages/Sana/diffusion/model/edm_sample.py
new file mode 100755
index 000000000..77d44a40d
--- /dev/null
+++ b/image/sana/sana-600m/packages/Sana/diffusion/model/edm_sample.py
@@ -0,0 +1,276 @@
+# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+# Modified from OpenAI's diffusion repos
+#     GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py
+#     ADM:   https://github.com/openai/guided-diffusion/blob/main/guided_diffusion
+#     IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
+
+
+import numpy as np
+from diffusion.model.utils import *
+from tqdm import tqdm
+
+# ----------------------------------------------------------------------------
+# Proposed EDM sampler (Algorithm 2).
+
+
+def edm_sampler(
+    net,
+    latents,
+    class_labels=None,
+    cfg_scale=None,
+    randn_like=torch.randn_like,
+    num_steps=18,
+    sigma_min=0.002,
+    sigma_max=80,
+    rho=7,
+    S_churn=0,
+    S_min=0,
+    S_max=float("inf"),
+    S_noise=1,
+    **kwargs,
+):
+    # Adjust noise levels based on what's supported by the network.
+    sigma_min = max(sigma_min, net.sigma_min)
+    sigma_max = min(sigma_max, net.sigma_max)
+
+    # Time step discretization.
+    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
+    t_steps = (
+        sigma_max ** (1 / rho)
+        + step_indices
+        / (num_steps - 1)
+        * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
+    ) ** rho
+    t_steps = torch.cat(
+        [net.round_sigma(t_steps), torch.zeros_like(t_steps[:1])]
+    )  # t_N = 0
+
+    # Main sampling loop.
+    x_next = latents.to(torch.float64) * t_steps[0]
+    for i, (t_cur, t_next) in tqdm(
+        list(enumerate(zip(t_steps[:-1], t_steps[1:])))
+    ):  # 0, ..., N-1
+        x_cur = x_next
+
+        # Increase noise temporarily.
+        gamma = (
+            min(S_churn / num_steps, np.sqrt(2) - 1) if S_min <= t_cur <= S_max else 0
+        )
+        t_hat = net.round_sigma(t_cur + gamma * t_cur)
+        x_hat = x_cur + (t_hat**2 - t_cur**2).sqrt() * S_noise * randn_like(x_cur)
+
+        # Euler step.
+        denoised = net(x_hat.float(), t_hat, class_labels, cfg_scale, **kwargs)["x"].to(
+            torch.float64
+        )
+        d_cur = (x_hat - denoised) / t_hat
+        x_next = x_hat + (t_next - t_hat) * d_cur
+
+        # Apply 2nd order correction.
+        if i < num_steps - 1:
+            denoised = net(x_next.float(), t_next, class_labels, cfg_scale, **kwargs)[
+                "x"
+            ].to(torch.float64)
+            d_prime = (x_next - denoised) / t_next
+            x_next = x_hat + (t_next - t_hat) * (0.5 * d_cur + 0.5 * d_prime)
+
+    return x_next
+
+
+# ----------------------------------------------------------------------------
+# Generalized ablation sampler, representing the superset of all sampling
+# methods discussed in the paper.
+
+
+def ablation_sampler(
+    net,
+    latents,
+    class_labels=None,
+    cfg_scale=None,
+    feat=None,
+    randn_like=torch.randn_like,
+    num_steps=18,
+    sigma_min=None,
+    sigma_max=None,
+    rho=7,
+    solver="heun",
+    discretization="edm",
+    schedule="linear",
+    scaling="none",
+    epsilon_s=1e-3,
+    C_1=0.001,
+    C_2=0.008,
+    M=1000,
+    alpha=1,
+    S_churn=0,
+    S_min=0,
+    S_max=float("inf"),
+    S_noise=1,
+):
+    assert solver in ["euler", "heun"]
+    assert discretization in ["vp", "ve", "iddpm", "edm"]
+    assert schedule in ["vp", "ve", "linear"]
+    assert scaling in ["vp", "none"]
+
+    # Helper functions for VP & VE noise level schedules.
+    vp_sigma = lambda beta_d, beta_min: (
+        lambda t: (np.e ** (0.5 * beta_d * (t**2) + beta_min * t) - 1) ** 0.5
+    )
+    vp_sigma_deriv = lambda beta_d, beta_min: (
+        lambda t: 0.5 * (beta_min + beta_d * t) * (sigma(t) + 1 / sigma(t))
+    )
+    vp_sigma_inv = lambda beta_d, beta_min: (
+        lambda sigma: (
+            ((beta_min**2 + 2 * beta_d * (sigma**2 + 1).log()).sqrt() - beta_min)
+            / beta_d
+        )
+    )
+    ve_sigma = lambda t: t.sqrt()
+    ve_sigma_deriv = lambda t: 0.5 / t.sqrt()
+    ve_sigma_inv = lambda sigma: sigma**2
+
+    # Select default noise level range based on the specified time step discretization.
+    if sigma_min is None:
+        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=epsilon_s)
+        sigma_min = {"vp": vp_def, "ve": 0.02, "iddpm": 0.002, "edm": 0.002}[
+            discretization
+        ]
+    if sigma_max is None:
+        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=1)
+        sigma_max = {"vp": vp_def, "ve": 100, "iddpm": 81, "edm": 80}[discretization]
+
+    # Adjust noise levels based on what's supported by the network.
+    sigma_min = max(sigma_min, net.sigma_min)
+    sigma_max = min(sigma_max, net.sigma_max)
+
+    # Compute corresponding betas for VP.
+    vp_beta_d = (
+        2
+        * (np.log(sigma_min**2 + 1) / epsilon_s - np.log(sigma_max**2 + 1))
+        / (epsilon_s - 1)
+    )
+    vp_beta_min = np.log(sigma_max**2 + 1) - 0.5 * vp_beta_d
+
+    # Define time steps in terms of noise level.
+    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
+    if discretization == "vp":
+        orig_t_steps = 1 + step_indices / (num_steps - 1) * (epsilon_s - 1)
+        sigma_steps = vp_sigma(vp_beta_d, vp_beta_min)(orig_t_steps)
+    elif discretization == "ve":
+        orig_t_steps = (sigma_max**2) * (
+            (sigma_min**2 / sigma_max**2) ** (step_indices / (num_steps - 1))
+        )
+        sigma_steps = ve_sigma(orig_t_steps)
+    elif discretization == "iddpm":
+        u = torch.zeros(M + 1, dtype=torch.float64, device=latents.device)
+        alpha_bar = lambda j: (0.5 * np.pi * j / M / (C_2 + 1)).sin() ** 2
+        for j in torch.arange(M, 0, -1, device=latents.device):  # M, ..., 1
+            u[j - 1] = (
+                (u[j] ** 2 + 1) / (alpha_bar(j - 1) / alpha_bar(j)).clip(min=C_1) - 1
+            ).sqrt()
+        u_filtered = u[torch.logical_and(u >= sigma_min, u <= sigma_max)]
+        sigma_steps = u_filtered[
+            ((len(u_filtered) - 1) / (num_steps - 1) * step_indices)
+            .round()
+            .to(torch.int64)
+        ]
+    else:
+        assert discretization == "edm"
+        sigma_steps = (
+            sigma_max ** (1 / rho)
+            + step_indices
+            / (num_steps - 1)
+            * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
+        ) ** rho
+
+    # Define noise level schedule.
+    if schedule == "vp":
+        sigma = vp_sigma(vp_beta_d, vp_beta_min)
+        sigma_deriv = vp_sigma_deriv(vp_beta_d, vp_beta_min)
+        sigma_inv = vp_sigma_inv(vp_beta_d, vp_beta_min)
+    elif schedule == "ve":
+        sigma = ve_sigma
+        sigma_deriv = ve_sigma_deriv
+        sigma_inv = ve_sigma_inv
+    else:
+        assert schedule == "linear"
+        sigma = lambda t: t
+        sigma_deriv = lambda t: 1
+        sigma_inv = lambda sigma: sigma
+
+    # Define scaling schedule.
+    if scaling == "vp":
+        s = lambda t: 1 / (1 + sigma(t) ** 2).sqrt()
+        s_deriv = lambda t: -sigma(t) * sigma_deriv(t) * (s(t) ** 3)
+    else:
+        assert scaling == "none"
+        s = lambda t: 1
+        s_deriv = lambda t: 0
+
+    # Compute final time steps based on the corresponding noise levels.
+    t_steps = sigma_inv(net.round_sigma(sigma_steps))
+    t_steps = torch.cat([t_steps, torch.zeros_like(t_steps[:1])])  # t_N = 0
+
+    # Main sampling loop.
+    t_next = t_steps[0]
+    x_next = latents.to(torch.float64) * (sigma(t_next) * s(t_next))
+    for i, (t_cur, t_next) in enumerate(zip(t_steps[:-1], t_steps[1:])):  # 0, ..., N-1
+        x_cur = x_next
+
+        # Increase noise temporarily.
+        gamma = (
+            min(S_churn / num_steps, np.sqrt(2) - 1)
+            if S_min <= sigma(t_cur) <= S_max
+            else 0
+        )
+        t_hat = sigma_inv(net.round_sigma(sigma(t_cur) + gamma * sigma(t_cur)))
+        x_hat = s(t_hat) / s(t_cur) * x_cur + (
+            sigma(t_hat) ** 2 - sigma(t_cur) ** 2
+        ).clip(min=0).sqrt() * s(t_hat) * S_noise * randn_like(x_cur)
+
+        # Euler step.
+        h = t_next - t_hat
+        denoised = net(
+            x_hat.float() / s(t_hat), sigma(t_hat), class_labels, cfg_scale, feat=feat
+        )["x"].to(torch.float64)
+        d_cur = (
+            sigma_deriv(t_hat) / sigma(t_hat) + s_deriv(t_hat) / s(t_hat)
+        ) * x_hat - sigma_deriv(t_hat) * s(t_hat) / sigma(t_hat) * denoised
+        x_prime = x_hat + alpha * h * d_cur
+        t_prime = t_hat + alpha * h
+
+        # Apply 2nd order correction.
+        if solver == "euler" or i == num_steps - 1:
+            x_next = x_hat + h * d_cur
+        else:
+            assert solver == "heun"
+            denoised = net(
+                x_prime.float() / s(t_prime),
+                sigma(t_prime),
+                class_labels,
+                cfg_scale,
+                feat=feat,
+            )["x"].to(torch.float64)
+            d_prime = (
+                sigma_deriv(t_prime) / sigma(t_prime) + s_deriv(t_prime) / s(t_prime)
+            ) * x_prime - sigma_deriv(t_prime) * s(t_prime) / sigma(t_prime) * denoised
+            x_next = x_hat + h * (
+                (1 - 1 / (2 * alpha)) * d_cur + 1 / (2 * alpha) * d_prime
+            )
+
+    return x_next
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/gaussian_diffusion.py b/image/sana/sana-600m/packages/Sana/diffusion/model/gaussian_diffusion.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/gaussian_diffusion.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/gaussian_diffusion.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/basic_modules.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/basic_modules.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/basic_modules.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/basic_modules.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_ffn.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/develop_triton_litemla.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/flash_attn.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/lite_mla.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/mb_conv_pre_glu.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/act.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/conv.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/nn/norm.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_fwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/custom_autotune.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/linear_relu_fwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/mm.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/pad_vk_mm_fwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/proj_divide_bwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_mm_relu_bwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_divide_fwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_lite_mla_kernels/vk_q_mm_relu_bwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/depthwise_conv_fwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/triton_mb_conv_pre_glu_kernels/linear_glu_fwd.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/compare_results.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/custom_autotune.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/dtype.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/export_onnx.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/modules/utils/model.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/readme.md b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/readme.md
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/fastlinear/readme.md
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/fastlinear/readme.md
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/sana.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/sana.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/sana_U_shape.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_U_shape.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/sana_U_shape.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_U_shape.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_U_shape_multi_scale.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/sana_blocks.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_blocks.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/sana_blocks.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_blocks.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/sana_multi_scale.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_multi_scale.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/sana_multi_scale.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_multi_scale.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_multi_scale_adaln.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/nets/sana_others.py b/image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_others.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/nets/sana_others.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/nets/sana_others.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/norms.py b/image/sana/sana-600m/packages/Sana/diffusion/model/norms.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/norms.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/norms.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/respace.py b/image/sana/sana-600m/packages/Sana/diffusion/model/respace.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/respace.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/respace.py
diff --git a/image/sana/sana-600m/packages/Sana/diffusion/model/sa_solver.py b/image/sana/sana-600m/packages/Sana/diffusion/model/sa_solver.py
new file mode 100755
index 000000000..f36ccb4d0
--- /dev/null
+++ b/image/sana/sana-600m/packages/Sana/diffusion/model/sa_solver.py
@@ -0,0 +1,1616 @@
+# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+# SPDX-License-Identifier: Apache-2.0
+
+import math
+
+import torch
+from tqdm import tqdm
+
+
+class NoiseScheduleVP:
+    def __init__(
+        self,
+        schedule="discrete",
+        betas=None,
+        alphas_cumprod=None,
+        continuous_beta_0=0.1,
+        continuous_beta_1=20.0,
+        dtype=torch.float32,
+    ):
+        """Thanks to DPM-Solver for their code base"""
+        r"""Create a wrapper class for the forward SDE (VP type).
+        ***
+        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
+                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
+        ***
+        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
+        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
+        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
+            log_alpha_t = self.marginal_log_mean_coeff(t)
+            sigma_t = self.marginal_std(t)
+            lambda_t = self.marginal_lambda(t)
+        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
+            t = self.inverse_lambda(lambda_t)
+        ===============================================================
+        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
+        1. For discrete-time DPMs:
+            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
+                t_i = (i + 1) / N
+            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
+            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
+            Args:
+                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
+                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
+            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
+            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
+                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
+                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
+                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
+                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
+                and
+                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
+        2. For continuous-time DPMs:
+            We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
+            schedule are the default settings in DDPM and improved-DDPM:
+            Args:
+                beta_min: A `float` number. The smallest beta for the linear schedule.
+                beta_max: A `float` number. The largest beta for the linear schedule.
+                cosine_s: A `float` number. The hyperparameter in the cosine schedule.
+                cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
+                T: A `float` number. The ending time of the forward process.
+        ===============================================================
+        Args:
+            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
+                    'linear' or 'cosine' for continuous-time DPMs.
+        Returns:
+            A wrapper object of the forward SDE (VP type).
+
+        ===============================================================
+        Example:
+        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', betas=betas)
+        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
+        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
+        # For continuous-time DPMs (VPSDE), linear schedule:
+        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
+        """
+
+        if schedule not in ["discrete", "linear", "cosine"]:
+            raise ValueError(
+                "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(
+                    schedule
+                )
+            )
+
+        self.schedule = schedule
+        if schedule == "discrete":
+            if betas is not None:
+                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
+            else:
+                assert alphas_cumprod is not None
+                log_alphas = 0.5 * torch.log(alphas_cumprod)
+            self.total_N = len(log_alphas)
+            self.T = 1.0
+            self.t_array = (
+                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
+                .reshape((1, -1))
+                .to(dtype=dtype)
+            )
+            self.log_alpha_array = log_alphas.reshape(
+                (
+                    1,
+                    -1,
+                )
+            ).to(dtype=dtype)
+        else:
+            self.total_N = 1000
+            self.beta_0 = continuous_beta_0
+            self.beta_1 = continuous_beta_1
+            self.cosine_s = 0.008
+            self.cosine_beta_max = 999.0
+            self.cosine_t_max = (
+                math.atan(self.cosine_beta_max * (1.0 + self.cosine_s) / math.pi)
+                * 2.0
+                * (1.0 + self.cosine_s)
+                / math.pi
+                - self.cosine_s
+            )
+            self.cosine_log_alpha_0 = math.log(
+                math.cos(self.cosine_s / (1.0 + self.cosine_s) * math.pi / 2.0)
+            )
+            self.schedule = schedule
+            if schedule == "cosine":
+                # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
+                # Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
+                self.T = 0.9946
+            else:
+                self.T = 1.0
+
+    def marginal_log_mean_coeff(self, t):
+        """
+        Compute log(alpha_t) of a given continuous-time label t in [0, T].
+        """
+        if self.schedule == "discrete":
+            return interpolate_fn(
+                t.reshape((-1, 1)),
+                self.t_array.to(t.device),
+                self.log_alpha_array.to(t.device),
+            ).reshape(-1)
+        elif self.schedule == "linear":
+            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
+        elif self.schedule == "cosine":
+            log_alpha_fn = lambda s: torch.log(
+                torch.cos((s + self.cosine_s) / (1.0 + self.cosine_s) * math.pi / 2.0)
+            )
+            log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
+            return log_alpha_t
+
+    def marginal_alpha(self, t):
+        """
+        Compute alpha_t of a given continuous-time label t in [0, T].
+        """
+        return torch.exp(self.marginal_log_mean_coeff(t))
+
+    def marginal_std(self, t):
+        """
+        Compute sigma_t of a given continuous-time label t in [0, T].
+        """
+        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
+
+    def marginal_lambda(self, t):
+        """
+        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
+        """
+        log_mean_coeff = self.marginal_log_mean_coeff(t)
+        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
+        return log_mean_coeff - log_std
+
+    def inverse_lambda(self, lamb):
+        """
+        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
+        """
+        if self.schedule == "linear":
+            tmp = (
+                2.0
+                * (self.beta_1 - self.beta_0)
+                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
+            )
+            Delta = self.beta_0**2 + tmp
+            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
+        elif self.schedule == "discrete":
+            log_alpha = -0.5 * torch.logaddexp(
+                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
+            )
+            t = interpolate_fn(
+                log_alpha.reshape((-1, 1)),
+                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
+                torch.flip(self.t_array.to(lamb.device), [1]),
+            )
+            return t.reshape((-1,))
+        else:
+            log_alpha = -0.5 * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
+            t_fn = lambda log_alpha_t: (
+                torch.arccos(torch.exp(log_alpha_t + self.cosine_log_alpha_0))
+                * 2.0
+                * (1.0 + self.cosine_s)
+                / math.pi
+                - self.cosine_s
+            )
+            t = t_fn(log_alpha)
+            return t
+
+    def edm_sigma(self, t):
+        return self.marginal_std(t) / self.marginal_alpha(t)
+
+    def edm_inverse_sigma(self, edmsigma):
+        alpha = 1 / (edmsigma**2 + 1).sqrt()
+        sigma = alpha * edmsigma
+        lambda_t = torch.log(alpha / sigma)
+        t = self.inverse_lambda(lambda_t)
+        return t
+
+
+def model_wrapper(
+    model,
+    noise_schedule,
+    model_type="noise",
+    model_kwargs={},
+    guidance_type="uncond",
+    condition=None,
+    unconditional_condition=None,
+    guidance_scale=1.0,
+    classifier_fn=None,
+    classifier_kwargs={},
+):
+    """Thanks to DPM-Solver for their code base"""
+    """Create a wrapper function for the noise prediction model.
+    SA-Solver needs to solve the continuous-time diffusion SDEs. For DPMs trained on discrete-time labels, we need to
+    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
+    We support four types of the diffusion model by setting `model_type`:
+        1. "noise": noise prediction model. (Trained by predicting noise).
+        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
+        3. "v": velocity prediction model. (Trained by predicting the velocity).
+            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
+            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
+                arXiv preprint arXiv:2202.00512 (2022).
+            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
+                arXiv preprint arXiv:2210.02303 (2022).
+
+        4. "score": marginal score function. (Trained by denoising score matching).
+            Note that the score function and the noise prediction model follows a simple relationship:
+            ```
+                noise(x_t, t) = -sigma_t * score(x_t, t)
+            ```
+    We support three types of guided sampling by DPMs by setting `guidance_type`:
+        1. "uncond": unconditional sampling by DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
+            ``
+            The input `classifier_fn` has the following format:
+            ``
+                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
+            ``
+            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
+                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
+        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
+            The input `model` has the following format:
+            ``
+                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
+            ``
+            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
+            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
+                arXiv preprint arXiv:2207.12598 (2022).
+
+    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
+    or continuous-time labels (i.e. epsilon to T).
+    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
+    ``
+        def model_fn(x, t_continuous) -> noise:
+            t_input = get_model_input_time(t_continuous)
+            return noise_pred(model, x, t_input, **model_kwargs)
+    ``
+    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for SA-Solver.
+    ===============================================================
+    Args:
+        model: A diffusion model with the corresponding format described above.
+        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
+        model_type: A `str`. The parameterization type of the diffusion model.
+                    "noise" or "x_start" or "v" or "score".
+        model_kwargs: A `dict`. A dict for the other inputs of the model function.
+        guidance_type: A `str`. The type of the guidance for sampling.
+                    "uncond" or "classifier" or "classifier-free".
+        condition: A pytorch tensor. The condition for the guided sampling.
+                    Only used for "classifier" or "classifier-free" guidance type.
+        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
+                    Only used for "classifier-free" guidance type.
+        guidance_scale: A `float`. The scale for the guided sampling.
+        classifier_fn: A classifier function. Only used for the classifier guidance.
+        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
+    Returns:
+        A noise prediction model that accepts the noised data and the continuous time as the inputs.
+    """
+
+    def get_model_input_time(t_continuous):
+        """
+        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
+        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
+        For continuous-time DPMs, we just use `t_continuous`.
+        """
+        if noise_schedule.schedule == "discrete":
+            return (t_continuous - 1.0 / noise_schedule.total_N) * 1000.0
+        else:
+            return t_continuous
+
+    def noise_pred_fn(x, t_continuous, cond=None):
+        t_input = get_model_input_time(t_continuous)
+        if cond is None:
+            output = model(x, t_input, **model_kwargs)
+        else:
+            output = model(x, t_input, cond, **model_kwargs)
+        if model_type == "noise":
+            return output
+        elif model_type == "x_start":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return (x - alpha_t[0] * output) / sigma_t[0]
+        elif model_type == "v":
+            alpha_t, sigma_t = (
+                noise_schedule.marginal_alpha(t_continuous),
+                noise_schedule.marginal_std(t_continuous),
+            )
+            return alpha_t[0] * output + sigma_t[0] * x
+        elif model_type == "score":
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            return -sigma_t[0] * output
+
+    def cond_grad_fn(x, t_input):
+        """
+        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
+        """
+        with torch.enable_grad():
+            x_in = x.detach().requires_grad_(True)
+            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
+            return torch.autograd.grad(log_prob.sum(), x_in)[0]
+
+    def model_fn(x, t_continuous):
+        """
+        The noise predicition model function that is used for DPM-Solver.
+        """
+        if guidance_type == "uncond":
+            return noise_pred_fn(x, t_continuous)
+        elif guidance_type == "classifier":
+            assert classifier_fn is not None
+            t_input = get_model_input_time(t_continuous)
+            cond_grad = cond_grad_fn(x, t_input)
+            sigma_t = noise_schedule.marginal_std(t_continuous)
+            noise = noise_pred_fn(x, t_continuous)
+            return noise - guidance_scale * sigma_t * cond_grad
+        elif guidance_type == "classifier-free":
+            if guidance_scale == 1.0 or unconditional_condition is None:
+                return noise_pred_fn(x, t_continuous, cond=condition)
+            else:
+                x_in = torch.cat([x] * 2)
+                t_in = torch.cat([t_continuous] * 2)
+                c_in = torch.cat([unconditional_condition, condition])
+                noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
+                return noise_uncond + guidance_scale * (noise - noise_uncond)
+
+    assert model_type in ["noise", "x_start", "v", "score"]
+    assert guidance_type in ["uncond", "classifier", "classifier-free"]
+    return model_fn
+
+
+class SASolver:
+    def __init__(
+        self,
+        model_fn,
+        noise_schedule,
+        algorithm_type="data_prediction",
+        correcting_x0_fn=None,
+        correcting_xt_fn=None,
+        thresholding_max_val=1.0,
+        dynamic_thresholding_ratio=0.995,
+    ):
+        """
+        Construct a SA-Solver
+        The default value for algorithm_type is "data_prediction" and we recommend not to change it to
+        "noise_prediction". For details, please see Appendix A.2.4 in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
+        self.noise_schedule = noise_schedule
+        assert algorithm_type in ["data_prediction", "noise_prediction"]
+
+        if correcting_x0_fn == "dynamic_thresholding":
+            self.correcting_x0_fn = self.dynamic_thresholding_fn
+        else:
+            self.correcting_x0_fn = correcting_x0_fn
+
+        self.correcting_xt_fn = correcting_xt_fn
+        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
+        self.thresholding_max_val = thresholding_max_val
+
+        self.predict_x0 = algorithm_type == "data_prediction"
+
+        self.sigma_min = float(self.noise_schedule.edm_sigma(torch.tensor([1e-3])))
+        self.sigma_max = float(self.noise_schedule.edm_sigma(torch.tensor([1])))
+
+    def dynamic_thresholding_fn(self, x0, t=None):
+        """
+        The dynamic thresholding method.
+        """
+        dims = x0.dim()
+        p = self.dynamic_thresholding_ratio
+        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
+        s = expand_dims(
+            torch.maximum(
+                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
+            ),
+            dims,
+        )
+        x0 = torch.clamp(x0, -s, s) / s
+        return x0
+
+    def noise_prediction_fn(self, x, t):
+        """
+        Return the noise prediction model.
+        """
+        return self.model(x, t)
+
+    def data_prediction_fn(self, x, t):
+        """
+        Return the data prediction model (with corrector).
+        """
+        noise = self.noise_prediction_fn(x, t)
+        alpha_t, sigma_t = (
+            self.noise_schedule.marginal_alpha(t),
+            self.noise_schedule.marginal_std(t),
+        )
+        x0 = (x - sigma_t * noise) / alpha_t
+        if self.correcting_x0_fn is not None:
+            x0 = self.correcting_x0_fn(x0)
+        return x0
+
+    def model_fn(self, x, t):
+        """
+        Convert the model to the noise prediction model or the data prediction model.
+        """
+
+        if self.predict_x0:
+            return self.data_prediction_fn(x, t)
+        else:
+            return self.noise_prediction_fn(x, t)
+
+    def get_time_steps(self, skip_type, t_T, t_0, N, order, device):
+        """Compute the intermediate time steps for sampling."""
+        if skip_type == "logSNR":
+            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
+            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
+            logSNR_steps = lambda_T + torch.linspace(
+                torch.tensor(0.0).cpu().item(),
+                (lambda_0 - lambda_T).cpu().item() ** (1.0 / order),
+                N + 1,
+            ).pow(order).to(device)
+            return self.noise_schedule.inverse_lambda(logSNR_steps)
+        elif skip_type == "time":
+            t = (
+                torch.linspace(t_T ** (1.0 / order), t_0 ** (1.0 / order), N + 1)
+                .pow(order)
+                .to(device)
+            )
+            return t
+        elif skip_type == "karras":
+            sigma_min = max(0.002, self.sigma_min)
+            sigma_max = min(80, self.sigma_max)
+            sigma_steps = (
+                torch.linspace(sigma_max ** (1.0 / 7), sigma_min ** (1.0 / 7), N + 1)
+                .pow(7)
+                .to(device)
+            )
+            t = self.noise_schedule.edm_inverse_sigma(sigma_steps)
+            return t
+        else:
+            raise ValueError(
+                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time' or 'karras'"
+            )
+
+    def denoise_to_zero_fn(self, x, s):
+        """
+        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
+        """
+        return self.data_prediction_fn(x, s)
+
+    def get_coefficients_exponential_negative(
+        self, order, interval_start, interval_end
+    ):
+        """
+        Calculate the integral of exp(-x) * x^order dx from interval_start to interval_end
+        For calculating the coefficient of gradient terms after the lagrange interpolation,
+        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        For noise_prediction formula.
+        """
+        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
+
+        if order == 0:
+            return torch.exp(-interval_end) * (
+                torch.exp(interval_end - interval_start) - 1
+            )
+        elif order == 1:
+            return torch.exp(-interval_end) * (
+                (interval_start + 1) * torch.exp(interval_end - interval_start)
+                - (interval_end + 1)
+            )
+        elif order == 2:
+            return torch.exp(-interval_end) * (
+                (interval_start**2 + 2 * interval_start + 2)
+                * torch.exp(interval_end - interval_start)
+                - (interval_end**2 + 2 * interval_end + 2)
+            )
+        elif order == 3:
+            return torch.exp(-interval_end) * (
+                (interval_start**3 + 3 * interval_start**2 + 6 * interval_start + 6)
+                * torch.exp(interval_end - interval_start)
+                - (interval_end**3 + 3 * interval_end**2 + 6 * interval_end + 6)
+            )
+
+    def get_coefficients_exponential_positive(
+        self, order, interval_start, interval_end, tau
+    ):
+        """
+        Calculate the integral of exp(x(1+tau^2)) * x^order dx from interval_start to interval_end
+        For calculating the coefficient of gradient terms after the lagrange interpolation,
+        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        For data_prediction formula.
+        """
+        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
+
+        # after change of variable(cov)
+        interval_end_cov = (1 + tau**2) * interval_end
+        interval_start_cov = (1 + tau**2) * interval_start
+
+        if order == 0:
+            return (
+                torch.exp(interval_end_cov)
+                * (1 - torch.exp(-(interval_end_cov - interval_start_cov)))
+                / (1 + tau**2)
+            )
+        elif order == 1:
+            return (
+                torch.exp(interval_end_cov)
+                * (
+                    (interval_end_cov - 1)
+                    - (interval_start_cov - 1)
+                    * torch.exp(-(interval_end_cov - interval_start_cov))
+                )
+                / ((1 + tau**2) ** 2)
+            )
+        elif order == 2:
+            return (
+                torch.exp(interval_end_cov)
+                * (
+                    (interval_end_cov**2 - 2 * interval_end_cov + 2)
+                    - (interval_start_cov**2 - 2 * interval_start_cov + 2)
+                    * torch.exp(-(interval_end_cov - interval_start_cov))
+                )
+                / ((1 + tau**2) ** 3)
+            )
+        elif order == 3:
+            return (
+                torch.exp(interval_end_cov)
+                * (
+                    (
+                        interval_end_cov**3
+                        - 3 * interval_end_cov**2
+                        + 6 * interval_end_cov
+                        - 6
+                    )
+                    - (
+                        interval_start_cov**3
+                        - 3 * interval_start_cov**2
+                        + 6 * interval_start_cov
+                        - 6
+                    )
+                    * torch.exp(-(interval_end_cov - interval_start_cov))
+                )
+                / ((1 + tau**2) ** 4)
+            )
+
+    def lagrange_polynomial_coefficient(self, order, lambda_list):
+        """
+        Calculate the coefficient of lagrange polynomial
+        For lagrange interpolation
+        """
+        assert order in [0, 1, 2, 3]
+        assert order == len(lambda_list) - 1
+        if order == 0:
+            return [[1]]
+        elif order == 1:
+            return [
+                [
+                    1 / (lambda_list[0] - lambda_list[1]),
+                    -lambda_list[1] / (lambda_list[0] - lambda_list[1]),
+                ],
+                [
+                    1 / (lambda_list[1] - lambda_list[0]),
+                    -lambda_list[0] / (lambda_list[1] - lambda_list[0]),
+                ],
+            ]
+        elif order == 2:
+            denominator1 = (lambda_list[0] - lambda_list[1]) * (
+                lambda_list[0] - lambda_list[2]
+            )
+            denominator2 = (lambda_list[1] - lambda_list[0]) * (
+                lambda_list[1] - lambda_list[2]
+            )
+            denominator3 = (lambda_list[2] - lambda_list[0]) * (
+                lambda_list[2] - lambda_list[1]
+            )
+            return [
+                [
+                    1 / denominator1,
+                    (-lambda_list[1] - lambda_list[2]) / denominator1,
+                    lambda_list[1] * lambda_list[2] / denominator1,
+                ],
+                [
+                    1 / denominator2,
+                    (-lambda_list[0] - lambda_list[2]) / denominator2,
+                    lambda_list[0] * lambda_list[2] / denominator2,
+                ],
+                [
+                    1 / denominator3,
+                    (-lambda_list[0] - lambda_list[1]) / denominator3,
+                    lambda_list[0] * lambda_list[1] / denominator3,
+                ],
+            ]
+        elif order == 3:
+            denominator1 = (
+                (lambda_list[0] - lambda_list[1])
+                * (lambda_list[0] - lambda_list[2])
+                * (lambda_list[0] - lambda_list[3])
+            )
+            denominator2 = (
+                (lambda_list[1] - lambda_list[0])
+                * (lambda_list[1] - lambda_list[2])
+                * (lambda_list[1] - lambda_list[3])
+            )
+            denominator3 = (
+                (lambda_list[2] - lambda_list[0])
+                * (lambda_list[2] - lambda_list[1])
+                * (lambda_list[2] - lambda_list[3])
+            )
+            denominator4 = (
+                (lambda_list[3] - lambda_list[0])
+                * (lambda_list[3] - lambda_list[1])
+                * (lambda_list[3] - lambda_list[2])
+            )
+            return [
+                [
+                    1 / denominator1,
+                    (-lambda_list[1] - lambda_list[2] - lambda_list[3]) / denominator1,
+                    (
+                        lambda_list[1] * lambda_list[2]
+                        + lambda_list[1] * lambda_list[3]
+                        + lambda_list[2] * lambda_list[3]
+                    )
+                    / denominator1,
+                    (-lambda_list[1] * lambda_list[2] * lambda_list[3]) / denominator1,
+                ],
+                [
+                    1 / denominator2,
+                    (-lambda_list[0] - lambda_list[2] - lambda_list[3]) / denominator2,
+                    (
+                        lambda_list[0] * lambda_list[2]
+                        + lambda_list[0] * lambda_list[3]
+                        + lambda_list[2] * lambda_list[3]
+                    )
+                    / denominator2,
+                    (-lambda_list[0] * lambda_list[2] * lambda_list[3]) / denominator2,
+                ],
+                [
+                    1 / denominator3,
+                    (-lambda_list[0] - lambda_list[1] - lambda_list[3]) / denominator3,
+                    (
+                        lambda_list[0] * lambda_list[1]
+                        + lambda_list[0] * lambda_list[3]
+                        + lambda_list[1] * lambda_list[3]
+                    )
+                    / denominator3,
+                    (-lambda_list[0] * lambda_list[1] * lambda_list[3]) / denominator3,
+                ],
+                [
+                    1 / denominator4,
+                    (-lambda_list[0] - lambda_list[1] - lambda_list[2]) / denominator4,
+                    (
+                        lambda_list[0] * lambda_list[1]
+                        + lambda_list[0] * lambda_list[2]
+                        + lambda_list[1] * lambda_list[2]
+                    )
+                    / denominator4,
+                    (-lambda_list[0] * lambda_list[1] * lambda_list[2]) / denominator4,
+                ],
+            ]
+
+    def get_coefficients_fn(
+        self, order, interval_start, interval_end, lambda_list, tau
+    ):
+        """
+        Calculate the coefficient of gradients.
+        """
+        assert order in [1, 2, 3, 4]
+        assert order == len(lambda_list), (
+            "the length of lambda list must be equal to the order"
+        )
+        coefficients = []
+        lagrange_coefficient = self.lagrange_polynomial_coefficient(
+            order - 1, lambda_list
+        )
+        for i in range(order):
+            coefficient = 0
+            for j in range(order):
+                if self.predict_x0:
+                    coefficient += lagrange_coefficient[i][
+                        j
+                    ] * self.get_coefficients_exponential_positive(
+                        order - 1 - j, interval_start, interval_end, tau
+                    )
+                else:
+                    coefficient += lagrange_coefficient[i][
+                        j
+                    ] * self.get_coefficients_exponential_negative(
+                        order - 1 - j, interval_start, interval_end
+                    )
+            coefficients.append(coefficient)
+        assert len(coefficients) == order, (
+            "the length of coefficients does not match the order"
+        )
+        return coefficients
+
+    def adams_bashforth_update(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Predictor, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def adams_moulton_update(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        t_list = t_prev_list + [t]
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def adams_bashforth_update_few_steps(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Predictor, with the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        if self.predict_x0:
+            if (
+                order == 2
+            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to unipc. Note: This is used only for few steps sampling.
+                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
+                # ODE case
+                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
+                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
+                gradient_coefficients[0] += (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h**2 / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2)
+                    )
+                    / (
+                        ns.marginal_lambda(t_prev_list[-1])
+                        - ns.marginal_lambda(t_prev_list[-2])
+                    )
+                )
+                gradient_coefficients[1] -= (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h**2 / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2)
+                    )
+                    / (
+                        ns.marginal_lambda(t_prev_list[-1])
+                        - ns.marginal_lambda(t_prev_list[-2])
+                    )
+                )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def adams_moulton_update_few_steps(
+        self, order, x, tau, model_prev_list, t_prev_list, noise, t
+    ):
+        """
+        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        """
+
+        assert order in [
+            1,
+            2,
+            3,
+            4,
+        ], (
+            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
+        )
+
+        # get noise schedule
+        ns = self.noise_schedule
+        alpha_t = ns.marginal_alpha(t)
+        sigma_t = ns.marginal_std(t)
+        lambda_t = ns.marginal_lambda(t)
+        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
+        sigma_prev = ns.marginal_std(t_prev_list[-1])
+        gradient_part = torch.zeros_like(x)
+        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
+        lambda_list = []
+        t_list = t_prev_list + [t]
+        for i in range(order):
+            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
+        gradient_coefficients = self.get_coefficients_fn(
+            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
+        )
+
+        if self.predict_x0:
+            if (
+                order == 2
+            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to UniPC. Note: This is used only for few steps sampling.
+                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
+                # ODE case
+                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
+                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
+                gradient_coefficients[0] += (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2 * h)
+                    )
+                )
+                gradient_coefficients[1] -= (
+                    1.0
+                    * torch.exp((1 + tau**2) * lambda_t)
+                    * (
+                        h / 2
+                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
+                        / ((1 + tau**2) ** 2 * h)
+                    )
+                )
+
+        for i in range(order):
+            if self.predict_x0:
+                gradient_part += (
+                    (1 + tau**2)
+                    * sigma_t
+                    * torch.exp(-(tau**2) * lambda_t)
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+            else:
+                gradient_part += (
+                    -(1 + tau**2)
+                    * alpha_t
+                    * gradient_coefficients[i]
+                    * model_prev_list[-(i + 1)]
+                )
+
+        if self.predict_x0:
+            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
+        else:
+            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
+
+        if self.predict_x0:
+            x_t = (
+                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
+                + gradient_part
+                + noise_part
+            )
+        else:
+            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
+
+        return x_t
+
+    def sample_few_steps(
+        self,
+        x,
+        tau,
+        steps=5,
+        t_start=None,
+        t_end=None,
+        skip_type="time",
+        skip_order=1,
+        predictor_order=3,
+        corrector_order=4,
+        pc_mode="PEC",
+        return_intermediate=False,
+    ):
+        """
+        For the PC-mode, please refer to the wiki page
+        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
+        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
+        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
+        """
+
+        skip_first_step = False
+        skip_final_step = True
+        lower_order_final = True
+        denoise_to_zero = False
+
+        assert pc_mode in [
+            "PEC",
+            "PECE",
+        ], "Predictor-corrector mode only supports PEC and PECE"
+        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+
+        device = x.device
+        intermediates = []
+        with torch.no_grad():
+            assert steps >= max(predictor_order, corrector_order - 1)
+            timesteps = self.get_time_steps(
+                skip_type=skip_type,
+                t_T=t_T,
+                t_0=t_0,
+                N=steps,
+                order=skip_order,
+                device=device,
+            )
+            assert timesteps.shape[0] - 1 == steps
+            # Init the initial values.
+            step = 0
+            t = timesteps[step]
+            noise = torch.randn_like(x)
+            t_prev_list = [t]
+            # do not evaluate if skip_first_step
+            if skip_first_step:
+                if self.predict_x0:
+                    alpha_t = self.noise_schedule.marginal_alpha(t)
+                    sigma_t = self.noise_schedule.marginal_std(t)
+                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
+                else:
+                    model_prev_list = [x]
+            else:
+                model_prev_list = [self.model_fn(x, t)]
+
+            if self.correcting_xt_fn is not None:
+                x = self.correcting_xt_fn(x, t, step)
+            if return_intermediate:
+                intermediates.append(x)
+
+            # determine the first several values
+            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
+                t = timesteps[step]
+                predictor_order_used = min(predictor_order, step)
+                corrector_order_used = min(corrector_order, step + 1)
+                noise = torch.randn_like(x)
+                # predictor step
+                x_p = self.adams_bashforth_update_few_steps(
+                    order=predictor_order_used,
+                    x=x,
+                    tau=tau(t),
+                    model_prev_list=model_prev_list,
+                    t_prev_list=t_prev_list,
+                    noise=noise,
+                    t=t,
+                )
+                # evaluation step
+                model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                model_prev_list.append(model_x)
+                # corrector step
+                if corrector_order > 0:
+                    x = self.adams_moulton_update_few_steps(
+                        order=corrector_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x = x_p
+
+                # evaluation step if correction and mode = pece
+                if corrector_order > 0:
+                    if pc_mode == "PECE":
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+
+            for step in tqdm(
+                range(max(predictor_order, corrector_order - 1), steps + 1)
+            ):
+                if lower_order_final:
+                    predictor_order_used = min(predictor_order, steps - step + 1)
+                    corrector_order_used = min(corrector_order, steps - step + 2)
+
+                else:
+                    predictor_order_used = predictor_order
+                    corrector_order_used = corrector_order
+                t = timesteps[step]
+                noise = torch.randn_like(x)
+
+                # predictor step
+                if skip_final_step and step == steps and not denoise_to_zero:
+                    x_p = self.adams_bashforth_update_few_steps(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=0,
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x_p = self.adams_bashforth_update_few_steps(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+
+                # evaluation step
+                # do not evaluate if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                # do not update if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_prev_list.append(model_x)
+
+                # corrector step
+                # do not correct if skip_final_step and step = steps
+                if corrector_order > 0:
+                    if not skip_final_step or step < steps:
+                        x = self.adams_moulton_update_few_steps(
+                            order=corrector_order_used,
+                            x=x,
+                            tau=tau(t),
+                            model_prev_list=model_prev_list,
+                            t_prev_list=t_prev_list,
+                            noise=noise,
+                            t=t,
+                        )
+                    else:
+                        x = x_p
+                else:
+                    x = x_p
+
+                # evaluation step if mode = pece and step != steps
+                if corrector_order > 0:
+                    if pc_mode == "PECE" and step < steps:
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+                del model_prev_list[0]
+
+            if denoise_to_zero:
+                t = torch.ones((1,)).to(device) * t_0
+                x = self.denoise_to_zero_fn(x, t)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step + 1)
+                if return_intermediate:
+                    intermediates.append(x)
+        if return_intermediate:
+            return x, intermediates
+        else:
+            return x
+
+    def sample_more_steps(
+        self,
+        x,
+        tau,
+        steps=20,
+        t_start=None,
+        t_end=None,
+        skip_type="time",
+        skip_order=1,
+        predictor_order=3,
+        corrector_order=4,
+        pc_mode="PEC",
+        return_intermediate=False,
+    ):
+        """
+        For the PC-mode, please refer to the wiki page
+        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
+        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
+        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
+        """
+
+        skip_first_step = False
+        skip_final_step = False
+        lower_order_final = True
+        denoise_to_zero = True
+
+        assert pc_mode in [
+            "PEC",
+            "PECE",
+        ], "Predictor-corrector mode only supports PEC and PECE"
+        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
+        t_T = self.noise_schedule.T if t_start is None else t_start
+        assert t_0 > 0 and t_T > 0, (
+            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
+        )
+
+        device = x.device
+        intermediates = []
+        with torch.no_grad():
+            assert steps >= max(predictor_order, corrector_order - 1)
+            timesteps = self.get_time_steps(
+                skip_type=skip_type,
+                t_T=t_T,
+                t_0=t_0,
+                N=steps,
+                order=skip_order,
+                device=device,
+            )
+            assert timesteps.shape[0] - 1 == steps
+            # Init the initial values.
+            step = 0
+            t = timesteps[step]
+            noise = torch.randn_like(x)
+            t_prev_list = [t]
+            # do not evaluate if skip_first_step
+            if skip_first_step:
+                if self.predict_x0:
+                    alpha_t = self.noise_schedule.marginal_alpha(t)
+                    sigma_t = self.noise_schedule.marginal_std(t)
+                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
+                else:
+                    model_prev_list = [x]
+            else:
+                model_prev_list = [self.model_fn(x, t)]
+
+            if self.correcting_xt_fn is not None:
+                x = self.correcting_xt_fn(x, t, step)
+            if return_intermediate:
+                intermediates.append(x)
+
+            # determine the first several values
+            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
+                t = timesteps[step]
+                predictor_order_used = min(predictor_order, step)
+                corrector_order_used = min(corrector_order, step + 1)
+                noise = torch.randn_like(x)
+                # predictor step
+                x_p = self.adams_bashforth_update(
+                    order=predictor_order_used,
+                    x=x,
+                    tau=tau(t),
+                    model_prev_list=model_prev_list,
+                    t_prev_list=t_prev_list,
+                    noise=noise,
+                    t=t,
+                )
+                # evaluation step
+                model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                model_prev_list.append(model_x)
+                # corrector step
+                if corrector_order > 0:
+                    x = self.adams_moulton_update(
+                        order=corrector_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x = x_p
+
+                # evaluation step if mode = pece
+                if corrector_order > 0:
+                    if pc_mode == "PECE":
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+
+            for step in tqdm(
+                range(max(predictor_order, corrector_order - 1), steps + 1)
+            ):
+                if lower_order_final:
+                    predictor_order_used = min(predictor_order, steps - step + 1)
+                    corrector_order_used = min(corrector_order, steps - step + 2)
+
+                else:
+                    predictor_order_used = predictor_order
+                    corrector_order_used = corrector_order
+                t = timesteps[step]
+                noise = torch.randn_like(x)
+
+                # predictor step
+                if skip_final_step and step == steps and not denoise_to_zero:
+                    x_p = self.adams_bashforth_update(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=0,
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+                else:
+                    x_p = self.adams_bashforth_update(
+                        order=predictor_order_used,
+                        x=x,
+                        tau=tau(t),
+                        model_prev_list=model_prev_list,
+                        t_prev_list=t_prev_list,
+                        noise=noise,
+                        t=t,
+                    )
+
+                # evaluation step
+                # do not evaluate if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_x = self.model_fn(x_p, t)
+
+                # update model_list
+                # do not update if skip_final_step and step = steps
+                if not skip_final_step or step < steps:
+                    model_prev_list.append(model_x)
+
+                # corrector step
+                # do not correct if skip_final_step and step = steps
+                if corrector_order > 0:
+                    if not skip_final_step or step < steps:
+                        x = self.adams_moulton_update(
+                            order=corrector_order_used,
+                            x=x,
+                            tau=tau(t),
+                            model_prev_list=model_prev_list,
+                            t_prev_list=t_prev_list,
+                            noise=noise,
+                            t=t,
+                        )
+                    else:
+                        x = x_p
+                else:
+                    x = x_p
+
+                # evaluation step if mode = pece and step != steps
+                if corrector_order > 0:
+                    if pc_mode == "PECE" and step < steps:
+                        model_x = self.model_fn(x, t)
+                        del model_prev_list[-1]
+                        model_prev_list.append(model_x)
+
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step)
+                if return_intermediate:
+                    intermediates.append(x)
+
+                t_prev_list.append(t)
+                del model_prev_list[0]
+
+            if denoise_to_zero:
+                t = torch.ones((1,)).to(device) * t_0
+                x = self.denoise_to_zero_fn(x, t)
+                if self.correcting_xt_fn is not None:
+                    x = self.correcting_xt_fn(x, t, step + 1)
+                if return_intermediate:
+                    intermediates.append(x)
+        if return_intermediate:
+            return x, intermediates
+        else:
+            return x
+
+    def sample(
+        self,
+        mode,
+        x,
+        tau,
+        steps,
+        t_start=None,
+        t_end=None,
+        skip_type="time",
+        skip_order=1,
+        predictor_order=3,
+        corrector_order=4,
+        pc_mode="PEC",
+        return_intermediate=False,
+    ):
+        """
+        For the PC-mode, please refer to the wiki page
+        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
+        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
+        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
+
+        'few_steps' mode is recommended. The differences between 'few_steps' and 'more_steps' are as below:
+        1) 'few_steps' do not correct at final step and do not denoise to zero, while 'more_steps' do these two.
+        Thus the NFEs for 'few_steps' = steps, NFEs for 'more_steps' = steps + 2
+        For most of the experiments and tasks, we find these two operations do not have much help to sample quality.
+        2) 'few_steps' use a rescaling trick as in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
+        We find it will slightly improve the sample quality especially in few steps.
+        """
+        assert mode in [
+            "few_steps",
+            "more_steps",
+        ], "mode must be either 'few_steps' or 'more_steps'"
+        if mode == "few_steps":
+            return self.sample_few_steps(
+                x=x,
+                tau=tau,
+                steps=steps,
+                t_start=t_start,
+                t_end=t_end,
+                skip_type=skip_type,
+                skip_order=skip_order,
+                predictor_order=predictor_order,
+                corrector_order=corrector_order,
+                pc_mode=pc_mode,
+                return_intermediate=return_intermediate,
+            )
+        else:
+            return self.sample_more_steps(
+                x=x,
+                tau=tau,
+                steps=steps,
+                t_start=t_start,
+                t_end=t_end,
+                skip_type=skip_type,
+                skip_order=skip_order,
+                predictor_order=predictor_order,
+                corrector_order=corrector_order,
+                pc_mode=pc_mode,
+                return_intermediate=return_intermediate,
+            )
+
+
+#############################################################
+# other utility functions
+#############################################################
+
+
+def interpolate_fn(x, xp, yp):
+    """
+    A piecewise linear function y = f(x), using xp and yp as keypoints.
+    We implement f(x) in a differentiable way (i.e. applicable for autograd).
+    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
+    Args:
+        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
+        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
+        yp: PyTorch tensor with shape [C, K].
+    Returns:
+        The function values f(x), with shape [N, C].
+    """
+    N, K = x.shape[0], xp.shape[1]
+    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
+    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
+    x_idx = torch.argmin(x_indices, dim=2)
+    cand_start_idx = x_idx - 1
+    start_idx = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(1, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    end_idx = torch.where(
+        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
+    )
+    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
+    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
+    start_idx2 = torch.where(
+        torch.eq(x_idx, 0),
+        torch.tensor(0, device=x.device),
+        torch.where(
+            torch.eq(x_idx, K),
+            torch.tensor(K - 2, device=x.device),
+            cand_start_idx,
+        ),
+    )
+    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
+    start_y = torch.gather(
+        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
+    ).squeeze(2)
+    end_y = torch.gather(
+        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
+    ).squeeze(2)
+    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
+    return cand
+
+
+def expand_dims(v, dims):
+    """
+    Expand the tensor `v` to the dim `dims`.
+    Args:
+        `v`: a PyTorch tensor with shape [N].
+        `dim`: a `int`.
+    Returns:
+        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
+    """
+    return v[(...,) + (None,) * (dims - 1)]
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/timestep_sampler.py b/image/sana/sana-600m/packages/Sana/diffusion/model/timestep_sampler.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/timestep_sampler.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/timestep_sampler.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/utils.py b/image/sana/sana-600m/packages/Sana/diffusion/model/utils.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/utils.py
rename to image/sana/sana-600m/packages/Sana/diffusion/model/utils.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/sa_sampler.py b/image/sana/sana-600m/packages/Sana/diffusion/sa_sampler.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/sa_sampler.py
rename to image/sana/sana-600m/packages/Sana/diffusion/sa_sampler.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/sa_solver_diffusers.py b/image/sana/sana-600m/packages/Sana/diffusion/sa_solver_diffusers.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/sa_solver_diffusers.py
rename to image/sana/sana-600m/packages/Sana/diffusion/sa_solver_diffusers.py
diff --git a/llama/llama-3_1-8b-instruct-sglang/model/__init__.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/__init__.py
old mode 100644
new mode 100755
similarity index 100%
rename from llama/llama-3_1-8b-instruct-sglang/model/__init__.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/__init__.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/checkpoint.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/checkpoint.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/checkpoint.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/checkpoint.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/config.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/config.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/config.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/config.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/data_sampler.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/data_sampler.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/data_sampler.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/data_sampler.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/dist_utils.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/dist_utils.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/dist_utils.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/dist_utils.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/import_utils.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/import_utils.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/import_utils.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/import_utils.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/logger.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/logger.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/logger.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/logger.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/lr_scheduler.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/lr_scheduler.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/lr_scheduler.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/lr_scheduler.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/misc.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/misc.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/misc.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/misc.py
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/optimizer.py b/image/sana/sana-600m/packages/Sana/diffusion/utils/optimizer.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/optimizer.py
rename to image/sana/sana-600m/packages/Sana/diffusion/utils/optimizer.py
diff --git a/sana/sana_600M/packages/Sana/environment_setup.sh b/image/sana/sana-600m/packages/Sana/environment_setup.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/environment_setup.sh
rename to image/sana/sana-600m/packages/Sana/environment_setup.sh
diff --git a/sana/sana_600M/packages/Sana/pyproject.toml b/image/sana/sana-600m/packages/Sana/pyproject.toml
similarity index 100%
rename from sana/sana_600M/packages/Sana/pyproject.toml
rename to image/sana/sana-600m/packages/Sana/pyproject.toml
diff --git a/sana/sana_600M/packages/Sana/sana/cli/run.py b/image/sana/sana-600m/packages/Sana/sana/cli/run.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/sana/cli/run.py
rename to image/sana/sana-600m/packages/Sana/sana/cli/run.py
diff --git a/sana/sana_600M/packages/Sana/sana/cli/upload2hf.py b/image/sana/sana-600m/packages/Sana/sana/cli/upload2hf.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/sana/cli/upload2hf.py
rename to image/sana/sana-600m/packages/Sana/sana/cli/upload2hf.py
diff --git a/sana/sana_600M/packages/Sana/sana/tools/__init__.py b/image/sana/sana-600m/packages/Sana/sana/tools/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/sana/tools/__init__.py
rename to image/sana/sana-600m/packages/Sana/sana/tools/__init__.py
diff --git a/sana/sana_600M/packages/Sana/sana/tools/download.py b/image/sana/sana-600m/packages/Sana/sana/tools/download.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/sana/tools/download.py
rename to image/sana/sana-600m/packages/Sana/sana/tools/download.py
diff --git a/sana/sana_600M/packages/Sana/sana/tools/hf_utils.py b/image/sana/sana-600m/packages/Sana/sana/tools/hf_utils.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/sana/tools/hf_utils.py
rename to image/sana/sana-600m/packages/Sana/sana/tools/hf_utils.py
diff --git a/sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric.sh b/image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric.sh
rename to image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric_dpg.sh b/image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric_dpg.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric_dpg.sh
rename to image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric_dpg.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric_geneval.sh b/image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric_geneval.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric_geneval.sh
rename to image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric_geneval.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh b/image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh
rename to image/sana/sana-600m/packages/Sana/scripts/bash_run_inference_metric_imagereward.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/infer_metric_run_inference_metric.sh b/image/sana/sana-600m/packages/Sana/scripts/infer_metric_run_inference_metric.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/infer_metric_run_inference_metric.sh
rename to image/sana/sana-600m/packages/Sana/scripts/infer_metric_run_inference_metric.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh b/image/sana/sana-600m/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh
rename to image/sana/sana-600m/packages/Sana/scripts/infer_metric_run_inference_metric_geneval.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/infer_run_inference.sh b/image/sana/sana-600m/packages/Sana/scripts/infer_run_inference.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/infer_run_inference.sh
rename to image/sana/sana-600m/packages/Sana/scripts/infer_run_inference.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/infer_run_inference_geneval.sh b/image/sana/sana-600m/packages/Sana/scripts/infer_run_inference_geneval.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/infer_run_inference_geneval.sh
rename to image/sana/sana-600m/packages/Sana/scripts/infer_run_inference_geneval.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh b/image/sana/sana-600m/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh
rename to image/sana/sana-600m/packages/Sana/scripts/infer_run_inference_geneval_diffusers.sh
diff --git a/sana/sana_600M/packages/Sana/scripts/inference.py b/image/sana/sana-600m/packages/Sana/scripts/inference.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/inference.py
rename to image/sana/sana-600m/packages/Sana/scripts/inference.py
diff --git a/sana/sana_600M/packages/Sana/scripts/inference_dpg.py b/image/sana/sana-600m/packages/Sana/scripts/inference_dpg.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/inference_dpg.py
rename to image/sana/sana-600m/packages/Sana/scripts/inference_dpg.py
diff --git a/sana/sana_600M/packages/Sana/scripts/inference_geneval.py b/image/sana/sana-600m/packages/Sana/scripts/inference_geneval.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/inference_geneval.py
rename to image/sana/sana-600m/packages/Sana/scripts/inference_geneval.py
diff --git a/sana/sana_600M/packages/Sana/scripts/inference_geneval_diffusers.py b/image/sana/sana-600m/packages/Sana/scripts/inference_geneval_diffusers.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/inference_geneval_diffusers.py
rename to image/sana/sana-600m/packages/Sana/scripts/inference_geneval_diffusers.py
diff --git a/sana/sana_600M/packages/Sana/scripts/inference_image_reward.py b/image/sana/sana-600m/packages/Sana/scripts/inference_image_reward.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/inference_image_reward.py
rename to image/sana/sana-600m/packages/Sana/scripts/inference_image_reward.py
diff --git a/sana/sana_600M/packages/Sana/scripts/interface.py b/image/sana/sana-600m/packages/Sana/scripts/interface.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/interface.py
rename to image/sana/sana-600m/packages/Sana/scripts/interface.py
diff --git a/sana/sana_600M/packages/Sana/scripts/style.css b/image/sana/sana-600m/packages/Sana/scripts/style.css
similarity index 100%
rename from sana/sana_600M/packages/Sana/scripts/style.css
rename to image/sana/sana-600m/packages/Sana/scripts/style.css
diff --git a/sana/sana_600M/packages/Sana/tests/bash/entry.sh b/image/sana/sana-600m/packages/Sana/tests/bash/entry.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tests/bash/entry.sh
rename to image/sana/sana-600m/packages/Sana/tests/bash/entry.sh
diff --git a/sana/sana_600M/packages/Sana/tests/bash/test_inference.sh b/image/sana/sana-600m/packages/Sana/tests/bash/test_inference.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tests/bash/test_inference.sh
rename to image/sana/sana-600m/packages/Sana/tests/bash/test_inference.sh
diff --git a/sana/sana_600M/packages/Sana/tests/bash/test_training_1epoch.sh b/image/sana/sana-600m/packages/Sana/tests/bash/test_training_1epoch.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tests/bash/test_training_1epoch.sh
rename to image/sana/sana-600m/packages/Sana/tests/bash/test_training_1epoch.sh
diff --git a/llama/llama-3_1-8b-instruct/model/__init__.py b/image/sana/sana-600m/packages/Sana/tools/__init__.py
similarity index 100%
rename from llama/llama-3_1-8b-instruct/model/__init__.py
rename to image/sana/sana-600m/packages/Sana/tools/__init__.py
diff --git a/sana/sana_600M/packages/Sana/tools/convert_py_to_yaml.py b/image/sana/sana-600m/packages/Sana/tools/convert_py_to_yaml.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/convert_py_to_yaml.py
rename to image/sana/sana-600m/packages/Sana/tools/convert_py_to_yaml.py
diff --git a/sana/sana_600M/packages/Sana/tools/convert_sana_pag_to_diffusers.py b/image/sana/sana-600m/packages/Sana/tools/convert_sana_pag_to_diffusers.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/convert_sana_pag_to_diffusers.py
rename to image/sana/sana-600m/packages/Sana/tools/convert_sana_pag_to_diffusers.py
diff --git a/sana/sana_600M/packages/Sana/tools/convert_sana_to_diffusers.py b/image/sana/sana-600m/packages/Sana/tools/convert_sana_to_diffusers.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/convert_sana_to_diffusers.py
rename to image/sana/sana-600m/packages/Sana/tools/convert_sana_to_diffusers.py
diff --git a/sana/sana_600M/packages/Sana/tools/download.py b/image/sana/sana-600m/packages/Sana/tools/download.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/download.py
rename to image/sana/sana-600m/packages/Sana/tools/download.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/.gitignore b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/.gitignore
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/.gitignore
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/.gitignore
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/LICENSE b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/LICENSE
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/LICENSE
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/LICENSE
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/README.md b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/README.md
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/README.md
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/README.md
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/clip_score.py b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/clip_score.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/clip_score.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/clip_score.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/setup.py b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/setup.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/setup.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/setup.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__init__.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/src/clip_score/__main__.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py b/image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/clip-score/src/clip_score/clip_score.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/compute_clipscore.sh b/image/sana/sana-600m/packages/Sana/tools/metrics/compute_clipscore.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/compute_clipscore.sh
rename to image/sana/sana-600m/packages/Sana/tools/metrics/compute_clipscore.sh
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/compute_dpg.sh b/image/sana/sana-600m/packages/Sana/tools/metrics/compute_dpg.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/compute_dpg.sh
rename to image/sana/sana-600m/packages/Sana/tools/metrics/compute_dpg.sh
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/compute_fid_embedding.sh b/image/sana/sana-600m/packages/Sana/tools/metrics/compute_fid_embedding.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/compute_fid_embedding.sh
rename to image/sana/sana-600m/packages/Sana/tools/metrics/compute_fid_embedding.sh
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/compute_geneval.sh b/image/sana/sana-600m/packages/Sana/tools/metrics/compute_geneval.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/compute_geneval.sh
rename to image/sana/sana-600m/packages/Sana/tools/metrics/compute_geneval.sh
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/compute_imagereward.sh b/image/sana/sana-600m/packages/Sana/tools/metrics/compute_imagereward.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/compute_imagereward.sh
rename to image/sana/sana-600m/packages/Sana/tools/metrics/compute_imagereward.sh
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py b/image/sana/sana-600m/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/dpg_bench/compute_dpg_bench.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv b/image/sana/sana-600m/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv
rename to image/sana/sana-600m/packages/Sana/tools/metrics/dpg_bench/dpg_bench.csv
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/dpg_bench/requirements.txt b/image/sana/sana-600m/packages/Sana/tools/metrics/dpg_bench/requirements.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/dpg_bench/requirements.txt
rename to image/sana/sana-600m/packages/Sana/tools/metrics/dpg_bench/requirements.txt
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/LICENSE b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/LICENSE
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/LICENSE
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/LICENSE
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/README.md b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/README.md
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/README.md
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/README.md
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/annotations_clip.csv
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/annotations_if-xl.csv
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/annotations_sdv2.csv
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/annotations/mturk_hit_template.html
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/environment.yml b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/environment.yml
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/environment.yml
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/environment.yml
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/download_models.sh
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/evaluate_images.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/object_names.txt
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/evaluation/summary_scores.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/generation/diffusers_generate.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/images/geneval_figure_1.png
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/create_prompts.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/evaluation_metadata.jsonl
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/generation_prompts.txt
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/object_names.txt b/image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/object_names.txt
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/geneval/prompts/object_names.txt
rename to image/sana/sana-600m/packages/Sana/tools/metrics/geneval/prompts/object_names.txt
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json b/image/sana/sana-600m/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json
rename to image/sana/sana-600m/packages/Sana/tools/metrics/image_reward/benchmark-prompts-dict.json
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/image_reward/compute_image_reward.py b/image/sana/sana-600m/packages/Sana/tools/metrics/image_reward/compute_image_reward.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/image_reward/compute_image_reward.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/image_reward/compute_image_reward.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/.gitignore b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/.gitignore
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/.gitignore
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/.gitignore
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/CHANGELOG.md
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/LICENSE b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/LICENSE
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/LICENSE
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/LICENSE
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/README.md b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/README.md
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/README.md
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/README.md
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/compute_fid.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/noxfile.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/noxfile.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/noxfile.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/noxfile.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/setup.cfg b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/setup.cfg
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/setup.cfg
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/setup.cfg
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/setup.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/setup.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/setup.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/setup.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__init__.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/__main__.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/fid_score.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/src/pytorch_fid/inception.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py b/image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/pytorch-fid/tests/test_fid_score.py
diff --git a/sana/sana_600M/packages/Sana/tools/metrics/utils.py b/image/sana/sana-600m/packages/Sana/tools/metrics/utils.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/metrics/utils.py
rename to image/sana/sana-600m/packages/Sana/tools/metrics/utils.py
diff --git a/sana/sana_600M/packages/Sana/train_scripts/train.py b/image/sana/sana-600m/packages/Sana/train_scripts/train.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/train_scripts/train.py
rename to image/sana/sana-600m/packages/Sana/train_scripts/train.py
diff --git a/sana/sana_600M/packages/Sana/train_scripts/train.sh b/image/sana/sana-600m/packages/Sana/train_scripts/train.sh
similarity index 100%
rename from sana/sana_600M/packages/Sana/train_scripts/train.sh
rename to image/sana/sana-600m/packages/Sana/train_scripts/train.sh
diff --git a/image/segment-anything/README.md b/image/segment-anything/README.md
new file mode 100644
index 000000000..1f26a1276
--- /dev/null
+++ b/image/segment-anything/README.md
@@ -0,0 +1,31 @@
+# Segment Anything
+
+Deploy Segment Anything for image segmentation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image segmentation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "image_url": "https://as2.ftcdn.net/v2/jpg/00/66/26/87/1000_F_66268784_jccdcfdpf2vmq5X8raYA8JQT0sziZ1H9.jpg"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `python3-opencv`
diff --git a/image/segment-anything/config.yaml b/image/segment-anything/config.yaml
new file mode 100644
index 000000000..afea7f154
--- /dev/null
+++ b/image/segment-anything/config.yaml
@@ -0,0 +1,26 @@
+description: "Segment Anything for image segmentation"
+environment_variables: {}
+external_data:
+- local_data_path: sam_vit_h_4b8939.pth
+  url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
+external_package_dirs: []
+model_metadata:
+  repo_id: "facebook/sam-vit-huge"
+  example_model_input:
+    image_url: https://as2.ftcdn.net/v2/jpg/00/66/26/87/1000_F_66268784_jccdcfdpf2vmq5X8raYA8JQT0sziZ1H9.jpg
+model_name: Segment Anything
+python_version: py310
+requirements:
+- git+https://github.com/facebookresearch/segment-anything.git@6fdee8f2727f4506cfbbe553e23b895e27956588
+- opencv-python==4.8.1.78
+- torch==2.1.0
+- torchvision==0.16.0
+- pycocotools==2.0.7
+resources:
+  accelerator: A10G
+  cpu: 1000m
+  memory: 10Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- python3-opencv
diff --git a/llama/llama-3_1_70b-instruct/model/__init__.py b/image/segment-anything/model/__init__.py
similarity index 100%
rename from llama/llama-3_1_70b-instruct/model/__init__.py
rename to image/segment-anything/model/__init__.py
diff --git a/segment-anything/model/model.py b/image/segment-anything/model/model.py
similarity index 100%
rename from segment-anything/model/model.py
rename to image/segment-anything/model/model.py
diff --git a/image/stable-diffusion/dreamshaper-lcm/README.md b/image/stable-diffusion/dreamshaper-lcm/README.md
new file mode 100644
index 000000000..3928394a5
--- /dev/null
+++ b/image/stable-diffusion/dreamshaper-lcm/README.md
@@ -0,0 +1,31 @@
+# Dreamshaper Latent Consistency Model
+
+Deploy Dreamshaper Latent Consistency Model for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/stable-diffusion/dreamshaper-lcm/config.yaml b/image/stable-diffusion/dreamshaper-lcm/config.yaml
new file mode 100644
index 000000000..1b5ffd430
--- /dev/null
+++ b/image/stable-diffusion/dreamshaper-lcm/config.yaml
@@ -0,0 +1,21 @@
+description: "Dreamshaper Latent Consistency Model for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "SimianLuo/LCM_Dreamshaper_v7"
+  example_model_input:
+    prompt: A photo of an astronaut riding a horse
+model_name: Dreamshaper Latent Consistency Model
+python_version: py311
+requirements:
+- diffusers==0.21.4
+- transformers==4.34.1
+- accelerate==0.23.0
+- torch==2.1.0
+resources:
+  accelerator: A10G
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/llama/llama-7b-exllama-streaming/model/__init__.py b/image/stable-diffusion/dreamshaper-lcm/model/__init__.py
similarity index 100%
rename from llama/llama-7b-exllama-streaming/model/__init__.py
rename to image/stable-diffusion/dreamshaper-lcm/model/__init__.py
diff --git a/stable-diffusion/dreamshaper-lcm/model/model.py b/image/stable-diffusion/dreamshaper-lcm/model/model.py
similarity index 100%
rename from stable-diffusion/dreamshaper-lcm/model/model.py
rename to image/stable-diffusion/dreamshaper-lcm/model/model.py
diff --git a/image/stable-diffusion/playground-v2-trt/README.md b/image/stable-diffusion/playground-v2-trt/README.md
new file mode 100644
index 000000000..99add8568
--- /dev/null
+++ b/image/stable-diffusion/playground-v2-trt/README.md
@@ -0,0 +1,35 @@
+# Playground v2 - TensorRT
+
+Generate original images from text prompts.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/playground-v2-trt-8.6.1.post1-engine-A100](https://huggingface.co/baseten/playground-v2-trt-8.6.1.post1-engine-A100) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+}'
+```
+
+## Configuration highlights
+
+- Base image: `nvcr.io/nvidia/pytorch:23.11-py3`
+- Predict concurrency: **1**
+- System packages: `python3.10-venv, ffmpeg, libsm6, libxext6`
+- Environment variables: `HF_HUB_ENABLE_HF_TRANSFER`
diff --git a/image/stable-diffusion/playground-v2-trt/config.yaml b/image/stable-diffusion/playground-v2-trt/config.yaml
new file mode 100644
index 000000000..d7197438e
--- /dev/null
+++ b/image/stable-diffusion/playground-v2-trt/config.yaml
@@ -0,0 +1,57 @@
+base_image:
+  image: nvcr.io/nvidia/pytorch:23.11-py3
+  python_executable_path: /usr/bin/python
+description: Generate original images from text prompts.
+environment_variables:
+  HF_HUB_ENABLE_HF_TRANSFER: 1
+external_package_dirs: []
+model_cache:
+- repo_id: baseten/playground-v2-trt-8.6.1.post1-engine-A100
+  use_volume: false
+- allow_patterns:
+  - config.json
+  - diffusion_pytorch_model.safetensors
+  repo_id: madebyollin/sdxl-vae-fp16-fix
+  use_volume: false
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - playground-v2.safetensors
+  repo_id: playgroundai/playground-v2-1024px-aesthetic
+  use_volume: false
+model_metadata:
+  example_model_input:
+    prompt: Astronaut in a jungle, cold color palette, muted colors, detailed, 8k
+  pretty_name: Playground v2 - TensorRT
+  tags:
+  - image-generation
+model_name: Playground v2 - TensorRT
+python_version: py39
+requirements:
+- accelerate==0.25.0
+- colored==2.2.4
+- cuda-python==12.3.0
+- ftfy==6.1.3
+- nvtx==0.2.10
+- opencv-python==4.8.0.74
+- scipy==1.11.4
+- transformers==4.31.0
+- safetensors==0.4.1
+- hf_transfer==0.1.4
+- diffusers==0.23.1
+- invisible-watermark>=0.2.0
+- --extra-index-url https://pypi.ngc.nvidia.com
+- polygraphy==0.49.9
+- --extra-index-url https://pypi.nvidia.com
+- tensorrt==8.6.1.post1
+resources:
+  accelerator: A100
+  use_gpu: true
+runtime:
+  predict_concurrency: 1
+secrets: {}
+system_packages:
+- python3.10-venv
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/llama/llama-7b-exllama/model/__init__.py b/image/stable-diffusion/playground-v2-trt/model/__init__.py
similarity index 100%
rename from llama/llama-7b-exllama/model/__init__.py
rename to image/stable-diffusion/playground-v2-trt/model/__init__.py
diff --git a/stable-diffusion/playground-v2-trt/model/model.py b/image/stable-diffusion/playground-v2-trt/model/model.py
similarity index 100%
rename from stable-diffusion/playground-v2-trt/model/model.py
rename to image/stable-diffusion/playground-v2-trt/model/model.py
diff --git a/stable-diffusion/playground-v2-trt/packages/diffusion/trtclip.py b/image/stable-diffusion/playground-v2-trt/packages/diffusion/trtclip.py
similarity index 100%
rename from stable-diffusion/playground-v2-trt/packages/diffusion/trtclip.py
rename to image/stable-diffusion/playground-v2-trt/packages/diffusion/trtclip.py
diff --git a/stable-diffusion/playground-v2-trt/packages/diffusion/trtunet.py b/image/stable-diffusion/playground-v2-trt/packages/diffusion/trtunet.py
similarity index 100%
rename from stable-diffusion/playground-v2-trt/packages/diffusion/trtunet.py
rename to image/stable-diffusion/playground-v2-trt/packages/diffusion/trtunet.py
diff --git a/stable-diffusion/playground-v2-trt/show.py b/image/stable-diffusion/playground-v2-trt/show.py
similarity index 100%
rename from stable-diffusion/playground-v2-trt/show.py
rename to image/stable-diffusion/playground-v2-trt/show.py
diff --git a/image/stable-diffusion/sd-textual-inversion/README.md b/image/stable-diffusion/sd-textual-inversion/README.md
new file mode 100644
index 000000000..5794c95b4
--- /dev/null
+++ b/image/stable-diffusion/sd-textual-inversion/README.md
@@ -0,0 +1,31 @@
+# SD_Textual_Inversion
+
+Deploy SD_Textual_Inversion for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/stable-diffusion/sd-textual-inversion/config.yaml b/image/stable-diffusion/sd-textual-inversion/config.yaml
new file mode 100644
index 000000000..1570d85c0
--- /dev/null
+++ b/image/stable-diffusion/sd-textual-inversion/config.yaml
@@ -0,0 +1,26 @@
+description: "Stable Diffusion with Textual Inversion for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "SG161222/Realistic_Vision_V4.0_noVAE"
+  example_model_input:
+    prompt: A photo of an astronaut riding a horse in the style of LulaCipher
+  pretty_name: Stable Diffusion - Textual Inversion
+  tags:
+  - image-generation
+model_name: SD_Textual_Inversion
+python_version: py311
+requirements:
+- diffusers==0.16.1
+- transformers==4.36.0
+- ftfy==6.1.3
+- accelerate==0.25.0
+- torch==2.1.0
+- pillow==10.1.0
+resources:
+  accelerator: T4
+  cpu: 500m
+  memory: 512Mi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/stable-diffusion/sd-textual-inversion/data/LulaCipher.bin b/image/stable-diffusion/sd-textual-inversion/data/LulaCipher.bin
similarity index 100%
rename from stable-diffusion/sd-textual-inversion/data/LulaCipher.bin
rename to image/stable-diffusion/sd-textual-inversion/data/LulaCipher.bin
diff --git a/llama/llama-7b-vllm/model/__init__.py b/image/stable-diffusion/sd-textual-inversion/model/__init__.py
similarity index 100%
rename from llama/llama-7b-vllm/model/__init__.py
rename to image/stable-diffusion/sd-textual-inversion/model/__init__.py
diff --git a/stable-diffusion/sd-textual-inversion/model/model.py b/image/stable-diffusion/sd-textual-inversion/model/model.py
similarity index 100%
rename from stable-diffusion/sd-textual-inversion/model/model.py
rename to image/stable-diffusion/sd-textual-inversion/model/model.py
diff --git a/image/stable-diffusion/sd-turbo/README.md b/image/stable-diffusion/sd-turbo/README.md
new file mode 100644
index 000000000..1372ab9ec
--- /dev/null
+++ b/image/stable-diffusion/sd-turbo/README.md
@@ -0,0 +1,32 @@
+# SD Turbo
+
+Deploy [stabilityai/sdxl-turbo](https://huggingface.co/stabilityai/sdxl-turbo) for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [stabilityai/sdxl-turbo](https://huggingface.co/stabilityai/sdxl-turbo) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "A tree in a field under the night sky"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/stable-diffusion/sd-turbo/config.yaml b/image/stable-diffusion/sd-turbo/config.yaml
new file mode 100644
index 000000000..e9ffc0152
--- /dev/null
+++ b/image/stable-diffusion/sd-turbo/config.yaml
@@ -0,0 +1,30 @@
+description: "stabilityai/sdxl-turbo for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_cache:
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - '*.txt'
+  repo_id: stabilityai/sdxl-turbo
+  use_volume: false
+model_metadata:
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: A tree in a field under the night sky
+  pretty_name: SD Turbo
+  tags:
+  - image-generation
+model_name: SD Turbo
+python_version: py311
+requirements:
+- torch==2.0.1
+- transformers==4.35.2
+- diffusers==0.23.1
+- accelerate==0.24.1
+resources:
+  accelerator: T4
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/llama/llama-7b/model/__init__.py b/image/stable-diffusion/sd-turbo/model/__init__.py
similarity index 100%
rename from llama/llama-7b/model/__init__.py
rename to image/stable-diffusion/sd-turbo/model/__init__.py
diff --git a/stable-diffusion/sd-turbo/model/model.py b/image/stable-diffusion/sd-turbo/model/model.py
similarity index 100%
rename from stable-diffusion/sd-turbo/model/model.py
rename to image/stable-diffusion/sd-turbo/model/model.py
diff --git a/image/stable-diffusion/sdxl-controlnet-canny/README.md b/image/stable-diffusion/sdxl-controlnet-canny/README.md
new file mode 100644
index 000000000..de0fd7cbe
--- /dev/null
+++ b/image/stable-diffusion/sdxl-controlnet-canny/README.md
@@ -0,0 +1,31 @@
+# SDXL ControlNet Canny
+
+Deploy SDXL ControlNet Canny for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G:2 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/stable-diffusion/sdxl-controlnet-canny/baseten-logo.gif b/image/stable-diffusion/sdxl-controlnet-canny/baseten-logo.gif
similarity index 100%
rename from stable-diffusion/sdxl-controlnet-canny/baseten-logo.gif
rename to image/stable-diffusion/sdxl-controlnet-canny/baseten-logo.gif
diff --git a/image/stable-diffusion/sdxl-controlnet-canny/config.yaml b/image/stable-diffusion/sdxl-controlnet-canny/config.yaml
new file mode 100644
index 000000000..cd246fbf6
--- /dev/null
+++ b/image/stable-diffusion/sdxl-controlnet-canny/config.yaml
@@ -0,0 +1,33 @@
+description: "SDXL ControlNet Canny for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-diffusion-xl-base-1.0"
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: aerial view, a futuristic research complex in a bright foggy jungle, hard
+      lighting
+    image: "<base64-encoded PNG image>"
+  model_metadata: null
+  pretty_name: Stable Diffusion ControlNet
+  tags:
+  - image-generation
+model_name: SDXL ControlNet Canny
+python_version: py39
+requirements:
+- accelerate==0.23.0
+- transformers==4.33.2
+- safetensors==0.3.3
+- opencv-python==4.8.0.76
+- diffusers==0.21.2
+resources:
+  accelerator: A10G:2
+  cpu: 3500m
+  memory: 20Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/llava/llava-1.6-sgl/model/__init__.py b/image/stable-diffusion/sdxl-controlnet-canny/model/__init__.py
similarity index 100%
rename from llava/llava-1.6-sgl/model/__init__.py
rename to image/stable-diffusion/sdxl-controlnet-canny/model/__init__.py
diff --git a/stable-diffusion/sdxl-controlnet-canny/model/model.py b/image/stable-diffusion/sdxl-controlnet-canny/model/model.py
similarity index 100%
rename from stable-diffusion/sdxl-controlnet-canny/model/model.py
rename to image/stable-diffusion/sdxl-controlnet-canny/model/model.py
diff --git a/image/stable-diffusion/sdxl-controlnet-depth/README.md b/image/stable-diffusion/sdxl-controlnet-depth/README.md
new file mode 100644
index 000000000..48a775323
--- /dev/null
+++ b/image/stable-diffusion/sdxl-controlnet-depth/README.md
@@ -0,0 +1,31 @@
+# SDXL ControlNet Depth
+
+Deploy SDXL ControlNet Depth for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G:2 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "large bed, abstract painting on the wall, fluffy rug on the floor, ambient lighting, extremely detailed"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/image/stable-diffusion/sdxl-controlnet-depth/config.yaml b/image/stable-diffusion/sdxl-controlnet-depth/config.yaml
new file mode 100644
index 000000000..92cd2829d
--- /dev/null
+++ b/image/stable-diffusion/sdxl-controlnet-depth/config.yaml
@@ -0,0 +1,32 @@
+description: "SDXL ControlNet Depth for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-diffusion-xl-base-1.0"
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: large bed, abstract painting on the wall, fluffy rug on the floor, ambient
+      lighting, extremely detailed
+    image: "<base64-encoded PNG image>"
+  pretty_name: Stable Diffusion ControlNet Depth
+  tags:
+  - image-generation
+model_name: SDXL ControlNet Depth
+python_version: py39
+requirements:
+- accelerate==0.23.0
+- transformers==4.33.2
+- safetensors==0.3.3
+- opencv-python==4.8.0.76
+- diffusers==0.21.2
+resources:
+  accelerator: A10G:2
+  cpu: 3500m
+  memory: 20Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/llava/llava-v1.5-7b/model/__init__.py b/image/stable-diffusion/sdxl-controlnet-depth/model/__init__.py
similarity index 100%
rename from llava/llava-v1.5-7b/model/__init__.py
rename to image/stable-diffusion/sdxl-controlnet-depth/model/__init__.py
diff --git a/stable-diffusion/sdxl-controlnet-depth/model/model.py b/image/stable-diffusion/sdxl-controlnet-depth/model/model.py
similarity index 100%
rename from stable-diffusion/sdxl-controlnet-depth/model/model.py
rename to image/stable-diffusion/sdxl-controlnet-depth/model/model.py
diff --git a/image/stable-diffusion/sdxl-controlnet/README.md b/image/stable-diffusion/sdxl-controlnet/README.md
new file mode 100644
index 000000000..c7847277a
--- /dev/null
+++ b/image/stable-diffusion/sdxl-controlnet/README.md
@@ -0,0 +1,31 @@
+# SDXL ControlNet
+
+Deploy SDXL ControlNet for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "aerial view, a futuristic research complex in a bright foggy jungle, hard lighting"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/stable-diffusion/sdxl-controlnet/baseten-logo.gif b/image/stable-diffusion/sdxl-controlnet/baseten-logo.gif
similarity index 100%
rename from stable-diffusion/sdxl-controlnet/baseten-logo.gif
rename to image/stable-diffusion/sdxl-controlnet/baseten-logo.gif
diff --git a/image/stable-diffusion/sdxl-controlnet/config.yaml b/image/stable-diffusion/sdxl-controlnet/config.yaml
new file mode 100644
index 000000000..11067692e
--- /dev/null
+++ b/image/stable-diffusion/sdxl-controlnet/config.yaml
@@ -0,0 +1,32 @@
+description: "SDXL ControlNet for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-diffusion-xl-base-1.0"
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: aerial view, a futuristic research complex in a bright foggy jungle, hard
+      lighting
+    image: "<base64-encoded PNG image>"
+  pretty_name: Stable Diffusion ControlNet
+  tags:
+  - image-generation
+model_name: SDXL ControlNet
+python_version: py39
+requirements:
+- accelerate==0.25.0
+- transformers==4.36.0
+- safetensors==0.4.1
+- opencv-python==4.8.1.78
+- diffusers==0.24.0
+resources:
+  accelerator: A10G
+  cpu: 3500m
+  memory: 20Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/__init__.py b/image/stable-diffusion/sdxl-controlnet/model/__init__.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/__init__.py
rename to image/stable-diffusion/sdxl-controlnet/model/__init__.py
diff --git a/stable-diffusion/sdxl-controlnet/model/model.py b/image/stable-diffusion/sdxl-controlnet/model/model.py
similarity index 100%
rename from stable-diffusion/sdxl-controlnet/model/model.py
rename to image/stable-diffusion/sdxl-controlnet/model/model.py
diff --git a/image/stable-diffusion/sdxl-lightning/README.md b/image/stable-diffusion/sdxl-lightning/README.md
new file mode 100644
index 000000000..44d42fc42
--- /dev/null
+++ b/image/stable-diffusion/sdxl-lightning/README.md
@@ -0,0 +1,31 @@
+# SDXL Lightning
+
+Deploy SDXL Lightning for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "A tree in a field under the night sky"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/stable-diffusion/sdxl-lightning/config.yaml b/image/stable-diffusion/sdxl-lightning/config.yaml
new file mode 100644
index 000000000..6154693a1
--- /dev/null
+++ b/image/stable-diffusion/sdxl-lightning/config.yaml
@@ -0,0 +1,26 @@
+description: "SDXL Lightning for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "ByteDance/SDXL-Lightning"
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: A tree in a field under the night sky
+  pretty_name: SDXL Lightning
+  tags:
+  - image-generation
+model_name: SDXL Lightning
+python_version: py310
+requirements:
+- torch==2.0.1
+- transformers==4.35.2
+- diffusers==0.23.1
+- hf_transfer==0.1.4
+- xformers==0.0.22
+- accelerate==0.24.1
+resources:
+  accelerator: A100
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/llava/llava-v1.6-34b/model/__init__.py b/image/stable-diffusion/sdxl-lightning/model/__init__.py
similarity index 100%
rename from llava/llava-v1.6-34b/model/__init__.py
rename to image/stable-diffusion/sdxl-lightning/model/__init__.py
diff --git a/stable-diffusion/sdxl-lightning/model/model.py b/image/stable-diffusion/sdxl-lightning/model/model.py
similarity index 100%
rename from stable-diffusion/sdxl-lightning/model/model.py
rename to image/stable-diffusion/sdxl-lightning/model/model.py
diff --git a/image/stable-diffusion/sdxl-lora-swapping/README.md b/image/stable-diffusion/sdxl-lora-swapping/README.md
new file mode 100644
index 000000000..59a134a77
--- /dev/null
+++ b/image/stable-diffusion/sdxl-lora-swapping/README.md
@@ -0,0 +1,35 @@
+# Stable Diffusion XL with LoRA Swapping
+
+Deploy Stable Diffusion XL with LoRA Swapping for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "lora": {
+    "repo_id": "nerijs/pixel-art-xl",
+    "weights": "pixel-art-xl.safetensors"
+  },
+  "prompt": "pixel art, an baby giraffe"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/image/stable-diffusion/sdxl-lora-swapping/config.yaml b/image/stable-diffusion/sdxl-lora-swapping/config.yaml
new file mode 100644
index 000000000..5aae62310
--- /dev/null
+++ b/image/stable-diffusion/sdxl-lora-swapping/config.yaml
@@ -0,0 +1,28 @@
+description: "Stable Diffusion XL with LoRA Swapping for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: stabilityai/stable-diffusion-xl-base-1.0
+  example_model_input:
+    lora:
+      repo_id: nerijs/pixel-art-xl
+      weights: pixel-art-xl.safetensors
+    prompt: pixel art, an baby giraffe
+model_name: Stable Diffusion XL with LoRA Swapping
+python_version: py311
+requirements:
+- accelerate==0.23.0
+- transformers==4.33.2
+- safetensors==0.3.3
+- opencv-python==4.8.0.76
+- diffusers==0.21.2
+resources:
+  accelerator: A100
+  cpu: 3500m
+  memory: 20Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/magic-animate/model/__init__.py b/image/stable-diffusion/sdxl-lora-swapping/model/__init__.py
similarity index 100%
rename from magic-animate/model/__init__.py
rename to image/stable-diffusion/sdxl-lora-swapping/model/__init__.py
diff --git a/stable-diffusion/sdxl-lora-swapping/model/model.py b/image/stable-diffusion/sdxl-lora-swapping/model/model.py
similarity index 100%
rename from stable-diffusion/sdxl-lora-swapping/model/model.py
rename to image/stable-diffusion/sdxl-lora-swapping/model/model.py
diff --git a/image/stable-diffusion/sdxl-lora/README.md b/image/stable-diffusion/sdxl-lora/README.md
new file mode 100644
index 000000000..d7f527625
--- /dev/null
+++ b/image/stable-diffusion/sdxl-lora/README.md
@@ -0,0 +1,31 @@
+# Stable Diffusion XL with LoRA
+
+Deploy Stable Diffusion XL with LoRA for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/image/stable-diffusion/sdxl-lora/config.yaml b/image/stable-diffusion/sdxl-lora/config.yaml
new file mode 100644
index 000000000..717160710
--- /dev/null
+++ b/image/stable-diffusion/sdxl-lora/config.yaml
@@ -0,0 +1,25 @@
+description: "Stable Diffusion XL with LoRA for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-diffusion-xl-base-1.0"
+  example_model_input:
+    prompt: A photo of an astronaut riding a horse
+model_name: Stable Diffusion XL with LoRA
+python_version: py311
+requirements:
+- accelerate==0.25.0
+- transformers==4.36.0
+- safetensors==0.4.1
+- opencv-python==4.8.1.78
+- diffusers==0.24.0
+resources:
+  accelerator: A10G
+  cpu: 3500m
+  memory: 20Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/metavoice-1b/model/__init__.py b/image/stable-diffusion/sdxl-lora/model/__init__.py
similarity index 100%
rename from metavoice-1b/model/__init__.py
rename to image/stable-diffusion/sdxl-lora/model/__init__.py
diff --git a/stable-diffusion/sdxl-lora/model/model.py b/image/stable-diffusion/sdxl-lora/model/model.py
similarity index 100%
rename from stable-diffusion/sdxl-lora/model/model.py
rename to image/stable-diffusion/sdxl-lora/model/model.py
diff --git a/image/stable-diffusion/sdxl-turbo/README.md b/image/stable-diffusion/sdxl-turbo/README.md
new file mode 100644
index 000000000..b8b3bc107
--- /dev/null
+++ b/image/stable-diffusion/sdxl-turbo/README.md
@@ -0,0 +1,32 @@
+# SDXL Turbo
+
+Deploy [stabilityai/sdxl-turbo](https://huggingface.co/stabilityai/sdxl-turbo) for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [stabilityai/sdxl-turbo](https://huggingface.co/stabilityai/sdxl-turbo) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "A tree in a field under the night sky"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/stable-diffusion/sdxl-turbo/config.yaml b/image/stable-diffusion/sdxl-turbo/config.yaml
new file mode 100644
index 000000000..8e2e52dc3
--- /dev/null
+++ b/image/stable-diffusion/sdxl-turbo/config.yaml
@@ -0,0 +1,34 @@
+description: "stabilityai/sdxl-turbo for image generation"
+environment_variables: {}
+external_package_dirs: []
+model_cache:
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - '*.txt'
+  repo_id: stabilityai/sdxl-turbo
+  use_volume: false
+model_metadata:
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: A tree in a field under the night sky
+  pretty_name: SDXL Turbo
+  tags:
+  - image-generation
+model_name: SDXL Turbo
+python_version: py310
+requirements:
+- torch==2.0.1
+- transformers==4.35.2
+- diffusers==0.23.1
+- hf_transfer==0.1.4
+- xformers==0.0.22
+- accelerate==0.24.1
+resources:
+  accelerator: T4
+  cpu: '3'
+  memory: 20Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/midnight/model/__init__.py b/image/stable-diffusion/sdxl-turbo/model/__init__.py
similarity index 100%
rename from midnight/model/__init__.py
rename to image/stable-diffusion/sdxl-turbo/model/__init__.py
diff --git a/stable-diffusion/sdxl-turbo/model/model.py b/image/stable-diffusion/sdxl-turbo/model/model.py
similarity index 100%
rename from stable-diffusion/sdxl-turbo/model/model.py
rename to image/stable-diffusion/sdxl-turbo/model/model.py
diff --git a/image/stable-diffusion/stable-diffusion-3-medium/README.md b/image/stable-diffusion/stable-diffusion-3-medium/README.md
new file mode 100644
index 000000000..cff014088
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-3-medium/README.md
@@ -0,0 +1,35 @@
+# Stable Diffusion 3 Medium
+
+Deploy [stabilityai/stable-diffusion-3-medium-diffusers](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [stabilityai/stable-diffusion-3-medium-diffusers](https://huggingface.co/stabilityai/stable-diffusion-3-medium-diffusers) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py310 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
+- Environment variables: `HF_HUB_OFFLINE`
diff --git a/image/stable-diffusion/stable-diffusion-3-medium/config.yaml b/image/stable-diffusion/stable-diffusion-3-medium/config.yaml
new file mode 100644
index 000000000..28d973a95
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-3-medium/config.yaml
@@ -0,0 +1,27 @@
+description: "stabilityai/stable-diffusion-3-medium-diffusers for image generation"
+environment_variables:
+  HF_HUB_OFFLINE: 1
+external_package_dirs: []
+model_metadata:
+  example_model_input:
+    prompt: A photo of an astronaut riding a horse
+model_cache:
+  - repo_id: stabilityai/stable-diffusion-3-medium-diffusers
+    use_volume: false
+model_name: Stable Diffusion 3 Medium
+python_version: py310
+requirements:
+  - diffusers==0.29.0
+  - transformers==4.36.0
+  - accelerate==0.25.0
+  - sentencepiece==0.1.99
+  - protobuf==4.25.1
+resources:
+  accelerator: A100
+  use_gpu: true
+secrets:
+  hf_access_token: ""
+system_packages:
+  - ffmpeg
+  - libsm6
+  - libxext6
diff --git a/mistral/mistral-7b-chat/model/__init__.py b/image/stable-diffusion/stable-diffusion-3-medium/model/__init__.py
similarity index 100%
rename from mistral/mistral-7b-chat/model/__init__.py
rename to image/stable-diffusion/stable-diffusion-3-medium/model/__init__.py
diff --git a/stable-diffusion/stable-diffusion-3-medium/model/model.py b/image/stable-diffusion/stable-diffusion-3-medium/model/model.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-3-medium/model/model.py
rename to image/stable-diffusion/stable-diffusion-3-medium/model/model.py
diff --git a/image/stable-diffusion/stable-diffusion-inpainting-trt/README.md b/image/stable-diffusion/stable-diffusion-inpainting-trt/README.md
new file mode 100644
index 000000000..dfec6f6c2
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-inpainting-trt/README.md
@@ -0,0 +1,31 @@
+# Stable Diffusion Inpainting TRT
+
+Deploy Stable Diffusion Inpainting TRT for image generation on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/stable-diffusion/stable-diffusion-inpainting-trt/config.yaml b/image/stable-diffusion/stable-diffusion-inpainting-trt/config.yaml
new file mode 100644
index 000000000..05ef54a83
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-inpainting-trt/config.yaml
@@ -0,0 +1,18 @@
+description: "Stable Diffusion Inpainting with TensorRT for image inpainting"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "runwayml/stable-diffusion-inpainting"
+  example_model_input:
+    prompt: A beautiful vase of flowers on a table
+    image: "<base64-encoded image>"
+    mask: "<base64-encoded mask image>"
+model_name: Stable Diffusion Inpainting TRT
+python_version: py310
+requirements: []
+requirements_file: requirements.txt
+resources:
+  accelerator: A10G
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/mistral/mistral-7b-instruct-vllm/model/__init__.py b/image/stable-diffusion/stable-diffusion-inpainting-trt/model/__init__.py
similarity index 100%
rename from mistral/mistral-7b-instruct-vllm/model/__init__.py
rename to image/stable-diffusion/stable-diffusion-inpainting-trt/model/__init__.py
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/model/model.py b/image/stable-diffusion/stable-diffusion-inpainting-trt/model/model.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-inpainting-trt/model/model.py
rename to image/stable-diffusion/stable-diffusion-inpainting-trt/model/model.py
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/inpaint_pipeline.py b/image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/inpaint_pipeline.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/inpaint_pipeline.py
rename to image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/inpaint_pipeline.py
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/models.py b/image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/models.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/models.py
rename to image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/models.py
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/stable_diffusion_pipeline.py b/image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/stable_diffusion_pipeline.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/stable_diffusion_pipeline.py
rename to image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/stable_diffusion_pipeline.py
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/utilities.py b/image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/utilities.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/utilities.py
rename to image/stable-diffusion/stable-diffusion-inpainting-trt/packages/helpers/utilities.py
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/requirements.txt b/image/stable-diffusion/stable-diffusion-inpainting-trt/requirements.txt
similarity index 100%
rename from stable-diffusion/stable-diffusion-inpainting-trt/requirements.txt
rename to image/stable-diffusion/stable-diffusion-inpainting-trt/requirements.txt
diff --git a/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/README.md b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/README.md
new file mode 100644
index 000000000..184be9ec3
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/README.md
@@ -0,0 +1,35 @@
+# Stable Diffusion XL - TensorRT
+
+Generate original images from text prompts.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/sdxl-1.0-trt-8.6.1.post1-engine-H100](https://huggingface.co/baseten/sdxl-1.0-trt-8.6.1.post1-engine-H100) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | H100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+}'
+```
+
+## Configuration highlights
+
+- Base image: `nvcr.io/nvidia/pytorch:23.11-py3`
+- Predict concurrency: **1**
+- System packages: `python3.10-venv, ffmpeg, libsm6, libxext6`
+- Environment variables: `HF_HUB_ENABLE_HF_TRANSFER`
diff --git a/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/config.yaml b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/config.yaml
new file mode 100644
index 000000000..a89232fa8
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/config.yaml
@@ -0,0 +1,65 @@
+base_image:
+  image: nvcr.io/nvidia/pytorch:23.11-py3
+  python_executable_path: /usr/bin/python
+description: Generate original images from text prompts.
+environment_variables:
+  HF_HUB_ENABLE_HF_TRANSFER: 1
+external_package_dirs: []
+model_cache:
+- repo_id: baseten/sdxl-1.0-trt-8.6.1.post1-engine-H100
+  use_volume: false
+- allow_patterns:
+  - config.json
+  - diffusion_pytorch_model.safetensors
+  repo_id: madebyollin/sdxl-vae-fp16-fix
+  use_volume: false
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - sd_xl_base_1.0.safetensors
+  repo_id: stabilityai/stable-diffusion-xl-base-1.0
+  use_volume: false
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - sd_xl_refiner_1.0.safetensors
+  repo_id: stabilityai/stable-diffusion-xl-refiner-1.0
+  use_volume: false
+model_metadata:
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: Astronaut in a jungle, cold color palette, muted colors, detailed, 8k
+  pretty_name: Stable Diffusion XL - TensorRT
+  tags:
+  - image-generation
+model_name: Stable Diffusion XL - TensorRT
+python_version: py39
+requirements:
+- accelerate==0.25.0
+- colored==2.2.4
+- cuda-python==12.3.0
+- ftfy==6.1.3
+- nvtx==0.2.10
+- opencv-python==4.8.0.74
+- scipy==1.11.4
+- transformers==4.31.0
+- safetensors==0.4.1
+- hf_transfer==0.1.4
+- diffusers==0.23.1
+- invisible-watermark>=0.2.0
+- --extra-index-url https://pypi.ngc.nvidia.com
+- polygraphy==0.49.9
+- --extra-index-url https://pypi.nvidia.com
+- tensorrt==8.6.1.post1
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency: 1
+secrets: {}
+system_packages:
+- python3.10-venv
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/mistral/mistral-7b-instruct/model/__init__.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/__init__.py
similarity index 100%
rename from mistral/mistral-7b-instruct/model/__init__.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/__init__.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/model.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/model.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/model.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/model.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtclip.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtclip.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtclip.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtclip.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtunet.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtunet.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtunet.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/packages/diffusion/trtunet.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/show.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/show.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt-h100/show.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/show.py
diff --git a/image/stable-diffusion/stable-diffusion-xl-1.0-trt/README.md b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/README.md
new file mode 100644
index 000000000..9a7b6dd87
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/README.md
@@ -0,0 +1,35 @@
+# Stable Diffusion XL - TensorRT
+
+Generate original images from text prompts.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/sdxl-1.0-trt-8.6.1.post1-engine](https://huggingface.co/baseten/sdxl-1.0-trt-8.6.1.post1-engine) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
+}'
+```
+
+## Configuration highlights
+
+- Base image: `nvcr.io/nvidia/pytorch:23.11-py3`
+- Predict concurrency: **1**
+- System packages: `python3.10-venv, ffmpeg, libsm6, libxext6`
+- Environment variables: `HF_HUB_ENABLE_HF_TRANSFER`
diff --git a/image/stable-diffusion/stable-diffusion-xl-1.0-trt/config.yaml b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/config.yaml
new file mode 100644
index 000000000..ba99ec3bf
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/config.yaml
@@ -0,0 +1,65 @@
+base_image:
+  image: nvcr.io/nvidia/pytorch:23.11-py3
+  python_executable_path: /usr/bin/python
+description: Generate original images from text prompts.
+environment_variables:
+  HF_HUB_ENABLE_HF_TRANSFER: 1
+external_package_dirs: []
+model_cache:
+- repo_id: baseten/sdxl-1.0-trt-8.6.1.post1-engine
+  use_volume: false
+- allow_patterns:
+  - config.json
+  - diffusion_pytorch_model.safetensors
+  repo_id: madebyollin/sdxl-vae-fp16-fix
+  use_volume: false
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - sd_xl_base_1.0.safetensors
+  repo_id: stabilityai/stable-diffusion-xl-base-1.0
+  use_volume: false
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - sd_xl_refiner_1.0.safetensors
+  repo_id: stabilityai/stable-diffusion-xl-refiner-1.0
+  use_volume: false
+model_metadata:
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: Astronaut in a jungle, cold color palette, muted colors, detailed, 8k
+  pretty_name: Stable Diffusion XL - TensorRT
+  tags:
+  - image-generation
+model_name: Stable Diffusion XL - TensorRT
+python_version: py39
+requirements:
+- accelerate==0.25.0
+- colored==2.2.4
+- cuda-python==12.3.0
+- ftfy==6.1.3
+- nvtx==0.2.10
+- opencv-python==4.8.0.74
+- scipy==1.11.4
+- transformers==4.31.0
+- safetensors==0.4.1
+- hf_transfer==0.1.4
+- diffusers==0.23.1
+- invisible-watermark>=0.2.0
+- --extra-index-url https://pypi.ngc.nvidia.com
+- polygraphy==0.49.9
+- --extra-index-url https://pypi.nvidia.com
+- tensorrt==8.6.1.post1
+resources:
+  accelerator: A100
+  use_gpu: true
+runtime:
+  predict_concurrency: 1
+secrets: {}
+system_packages:
+- python3.10-venv
+- ffmpeg
+- libsm6
+- libxext6
diff --git a/mistral/mistral-7b/model/__init__.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/model/__init__.py
similarity index 100%
rename from mistral/mistral-7b/model/__init__.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt/model/__init__.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt/model/model.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/model/model.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt/model/model.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt/model/model.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtclip.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtclip.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtclip.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtclip.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtunet.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtunet.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtunet.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt/packages/diffusion/trtunet.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt/show.py b/image/stable-diffusion/stable-diffusion-xl-1.0-trt/show.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt/show.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0-trt/show.py
diff --git a/image/stable-diffusion/stable-diffusion-xl-1.0/README.md b/image/stable-diffusion/stable-diffusion-xl-1.0/README.md
new file mode 100644
index 000000000..eecc45bcd
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion-xl-1.0/README.md
@@ -0,0 +1,33 @@
+# Stable Diffusion XL
+
+Generate original images from text prompts.
+
+| Property | Value |
+|----------|-------|
+| Model | [madebyollin/sdxl-vae-fp16-fix](https://huggingface.co/madebyollin/sdxl-vae-fp16-fix) |
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "A tree in a field under the night sky",
+  "use_refiner": true
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/stable-diffusion/stable-diffusion-xl-1.0/config.yaml b/image/stable-diffusion/stable-diffusion-xl-1.0/config.yaml
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0/config.yaml
rename to image/stable-diffusion/stable-diffusion-xl-1.0/config.yaml
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/model/__init__.py b/image/stable-diffusion/stable-diffusion-xl-1.0/model/__init__.py
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/model/__init__.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0/model/__init__.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0/model/model.py b/image/stable-diffusion/stable-diffusion-xl-1.0/model/model.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0/model/model.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0/model/model.py
diff --git a/stable-diffusion/stable-diffusion-xl-1.0/show.py b/image/stable-diffusion/stable-diffusion-xl-1.0/show.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0/show.py
rename to image/stable-diffusion/stable-diffusion-xl-1.0/show.py
diff --git a/image/stable-diffusion/stable-diffusion/README.md b/image/stable-diffusion/stable-diffusion/README.md
new file mode 100644
index 000000000..b2de1c4bd
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion/README.md
@@ -0,0 +1,31 @@
+# Stable Diffusion
+
+Generate original images from text prompts.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/image/stable-diffusion/stable-diffusion/config.yaml b/image/stable-diffusion/stable-diffusion/config.yaml
new file mode 100644
index 000000000..862e75187
--- /dev/null
+++ b/image/stable-diffusion/stable-diffusion/config.yaml
@@ -0,0 +1,36 @@
+description: Generate original images from text prompts.
+environment_variables: {}
+external_data:
+- local_data_path: unet/diffusion_pytorch_model.bin
+  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/stable-diffusion-truss/unet/diffusion_pytorch_model.bin
+- local_data_path: text_encoder/pytorch_model.bin
+  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/stable-diffusion-truss/text_encoder/pytorch_model.bin
+- local_data_path: vae/diffusion_pytorch_model.bin
+  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/stable-diffusion-truss/vae/diffusion_pytorch_model.bin
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-diffusion-2-1-base"
+  avatar_url: https://cdn.baseten.co/production/static/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/sd.png
+  example_model_input:
+    prompt: A photo of an astronaut riding a horse
+  pretty_name: Stable Diffusion
+  tags:
+  - image-generation
+model_name: Stable Diffusion
+python_version: py39
+requirements:
+- diffusers==0.24.0
+- transformers==4.36.0
+- torch==2.1.0
+- scipy==1.11.4
+- accelerate==0.25.0
+- pillow==10.1.0
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 14Gi
+  use_gpu: true
+secrets: {}
+spec_version: "2.0"
+system_packages: []
diff --git a/stable-diffusion/stable-diffusion/data/model_index.json b/image/stable-diffusion/stable-diffusion/data/model_index.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/model_index.json
rename to image/stable-diffusion/stable-diffusion/data/model_index.json
diff --git a/stable-diffusion/stable-diffusion/data/scheduler/scheduler_config.json b/image/stable-diffusion/stable-diffusion/data/scheduler/scheduler_config.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/scheduler/scheduler_config.json
rename to image/stable-diffusion/stable-diffusion/data/scheduler/scheduler_config.json
diff --git a/stable-diffusion/stable-diffusion/data/text_encoder/config.json b/image/stable-diffusion/stable-diffusion/data/text_encoder/config.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/text_encoder/config.json
rename to image/stable-diffusion/stable-diffusion/data/text_encoder/config.json
diff --git a/stable-diffusion/stable-diffusion/data/tokenizer/merges.txt b/image/stable-diffusion/stable-diffusion/data/tokenizer/merges.txt
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/tokenizer/merges.txt
rename to image/stable-diffusion/stable-diffusion/data/tokenizer/merges.txt
diff --git a/stable-diffusion/stable-diffusion/data/tokenizer/special_tokens_map.json b/image/stable-diffusion/stable-diffusion/data/tokenizer/special_tokens_map.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/tokenizer/special_tokens_map.json
rename to image/stable-diffusion/stable-diffusion/data/tokenizer/special_tokens_map.json
diff --git a/stable-diffusion/stable-diffusion/data/tokenizer/tokenizer_config.json b/image/stable-diffusion/stable-diffusion/data/tokenizer/tokenizer_config.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/tokenizer/tokenizer_config.json
rename to image/stable-diffusion/stable-diffusion/data/tokenizer/tokenizer_config.json
diff --git a/stable-diffusion/stable-diffusion/data/tokenizer/vocab.json b/image/stable-diffusion/stable-diffusion/data/tokenizer/vocab.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/tokenizer/vocab.json
rename to image/stable-diffusion/stable-diffusion/data/tokenizer/vocab.json
diff --git a/stable-diffusion/stable-diffusion/data/unet/config.json b/image/stable-diffusion/stable-diffusion/data/unet/config.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/unet/config.json
rename to image/stable-diffusion/stable-diffusion/data/unet/config.json
diff --git a/stable-diffusion/stable-diffusion/data/vae/config.json b/image/stable-diffusion/stable-diffusion/data/vae/config.json
similarity index 100%
rename from stable-diffusion/stable-diffusion/data/vae/config.json
rename to image/stable-diffusion/stable-diffusion/data/vae/config.json
diff --git a/mistral/mixtral-8x22b/model/__init__.py b/image/stable-diffusion/stable-diffusion/model/__init__.py
similarity index 100%
rename from mistral/mixtral-8x22b/model/__init__.py
rename to image/stable-diffusion/stable-diffusion/model/__init__.py
diff --git a/stable-diffusion/stable-diffusion/model/model.py b/image/stable-diffusion/stable-diffusion/model/model.py
similarity index 100%
rename from stable-diffusion/stable-diffusion/model/model.py
rename to image/stable-diffusion/stable-diffusion/model/model.py
diff --git a/stable-diffusion/stable-diffusion/show.py b/image/stable-diffusion/stable-diffusion/show.py
similarity index 100%
rename from stable-diffusion/stable-diffusion/show.py
rename to image/stable-diffusion/stable-diffusion/show.py
diff --git a/image/stable-diffusion/stable-video-diffusion/README.md b/image/stable-diffusion/stable-video-diffusion/README.md
new file mode 100644
index 000000000..6eea00f77
--- /dev/null
+++ b/image/stable-diffusion/stable-video-diffusion/README.md
@@ -0,0 +1,31 @@
+# Stable Video Diffusion
+
+Stable Video Diffusion can turn any image into a short video.
+
+| Property | Value |
+|----------|-------|
+| Task | Image generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "A photo of a cat in a field of sunflowers"}'
+```
+
+> The response may contain base64-encoded image data.
+
+## Configuration highlights
+
+- System packages: `libgl1-mesa-glx, ffmpeg`
diff --git a/image/stable-diffusion/stable-video-diffusion/config.yaml b/image/stable-diffusion/stable-video-diffusion/config.yaml
new file mode 100644
index 000000000..c8c150352
--- /dev/null
+++ b/image/stable-diffusion/stable-video-diffusion/config.yaml
@@ -0,0 +1,42 @@
+description: Stable Video Diffusion can turn any image into a short video.
+environment_variables: {}
+external_data:
+- local_data_path: weights/svd.safetensors
+  url: https://huggingface.co/stabilityai/stable-video-diffusion-img2vid/resolve/main/svd.safetensors
+- local_data_path: weights/ViT-L-14.pt
+  url: https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt
+external_package_dirs: []
+model_metadata:
+  repo_id: "stabilityai/stable-video-diffusion-img2vid"
+  avatar_url: https://cdn.baseten.co/production/static/explore/stability.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/stable-video-diffusion.png
+  example_model_input:
+    image: "<base64-encoded image>"
+  tags:
+  - image-to-video
+model_name: Stable Video Diffusion
+python_version: py310
+requirements:
+- einops==0.7.0
+- fire==0.5.0
+- omegaconf==2.3.0
+- git+https://github.com/openai/CLIP.git@2dbac9065bb0b4ffc28ecf0e94758261d1ddfdb0
+- lightning==2.1.2
+- kornia==0.7.0
+- open-clip-torch==2.23.0
+- invisible-watermark==0.2.0
+- xformers==0.0.22
+- opencv-python==4.8.0.76
+- scipy==1.11.3
+- transformers==4.35.2
+- hf_transfer==0.1.4
+- git+https://github.com/Stability-AI/generative-models.git@059d8e9cd9c55aea1ef2ece39abf605efb8b7cc9
+resources:
+  accelerator: A100
+  cpu: '4'
+  memory: 16Gi
+  use_gpu: true
+secrets: {}
+system_packages:
+- libgl1-mesa-glx
+- ffmpeg
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/model/__init__.py b/image/stable-diffusion/stable-video-diffusion/model/__init__.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/model/__init__.py
rename to image/stable-diffusion/stable-video-diffusion/model/__init__.py
diff --git a/stable-diffusion/stable-video-diffusion/model/helper.py b/image/stable-diffusion/stable-video-diffusion/model/helper.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/helper.py
rename to image/stable-diffusion/stable-video-diffusion/model/helper.py
diff --git a/stable-diffusion/stable-video-diffusion/model/model.py b/image/stable-diffusion/stable-video-diffusion/model/model.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/model.py
rename to image/stable-diffusion/stable-video-diffusion/model/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/__init__.py b/image/stable-diffusion/stable-video-diffusion/model/scripts/__init__.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/__init__.py
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/__init__.py
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd.yaml b/image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd.yaml
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd.yaml
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd.yaml
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_image_decoder.yaml b/image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_image_decoder.yaml
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_image_decoder.yaml
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_image_decoder.yaml
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt.yaml b/image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt.yaml
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt.yaml
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt.yaml
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt_image_decoder.yaml b/image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt_image_decoder.yaml
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt_image_decoder.yaml
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/configs/svd_xt_image_decoder.yaml
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/sampling/simple_video_sample.py b/image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/simple_video_sample.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/sampling/simple_video_sample.py
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/sampling/simple_video_sample.py
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/tests/attention.py b/image/stable-diffusion/stable-video-diffusion/model/scripts/tests/attention.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/tests/attention.py
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/tests/attention.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/__init__.py b/image/stable-diffusion/stable-video-diffusion/model/scripts/util/__init__.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/__init__.py
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/util/__init__.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/model/__init__.py b/image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/__init__.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/model/__init__.py
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/__init__.py
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/nsfw_and_watermark_dectection.py b/image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/nsfw_and_watermark_dectection.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/util/detection/nsfw_and_watermark_dectection.py
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/nsfw_and_watermark_dectection.py
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/p_head_v1.npz b/image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/p_head_v1.npz
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/util/detection/p_head_v1.npz
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/p_head_v1.npz
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/w_head_v1.npz b/image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/w_head_v1.npz
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/util/detection/w_head_v1.npz
rename to image/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/w_head_v1.npz
diff --git a/stable-diffusion/stable-video-diffusion/sample_images/cheetah.jpeg b/image/stable-diffusion/stable-video-diffusion/sample_images/cheetah.jpeg
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/sample_images/cheetah.jpeg
rename to image/stable-diffusion/stable-video-diffusion/sample_images/cheetah.jpeg
diff --git a/stable-diffusion/stable-video-diffusion/sample_images/racecar.jpeg b/image/stable-diffusion/stable-video-diffusion/sample_images/racecar.jpeg
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/sample_images/racecar.jpeg
rename to image/stable-diffusion/stable-video-diffusion/sample_images/racecar.jpeg
diff --git a/infrastructure/README.md b/infrastructure/README.md
new file mode 100644
index 000000000..2022564dd
--- /dev/null
+++ b/infrastructure/README.md
@@ -0,0 +1,28 @@
+# Infrastructure Examples
+
+Examples demonstrating advanced Truss features, custom serving infrastructure, and integrations. These cover topics like custom servers, gRPC, chaining models, caching, and specialized tooling.
+
+| Directory | Description |
+|-----------|-------------|
+| [custom-server](custom-server/) | Custom inference servers using SGLang, LMDeploy, and other engines with Dockerfile-based Truss configs |
+| [grpc](grpc/) | Serve a model over gRPC instead of HTTP |
+| [custom-engine-builder-control](custom-engine-builder-control/) | Custom engine builder with fine-grained control over the build process |
+| [chains-examples](chains-examples/) | Truss Chains examples for multi-model pipelines |
+| [multiprocessing](multiprocessing/) | Use Python multiprocessing within a Truss model |
+| [model-cache](model-cache/) | Cache model weights across deployments for faster cold starts |
+| [metrics](metrics/) | Export custom metrics from a Truss model |
+| [jsonformatter](jsonformatter/) | Custom JSON formatting for model outputs |
+| [ngram-speculator](ngram-speculator/) | N-gram speculative decoding for faster LLM inference |
+| [llama-cpp-server](llama-cpp-server/) | Serve models using llama.cpp as the backend |
+| [binocular](binocular/) | Binocular LLM-generated text detection |
+| [layoutlm-document-qa](layoutlm-document-qa/) | LayoutLM document question answering |
+| [autodesk-wala](autodesk-wala/) | Autodesk WALA model integration |
+| [paddlepaddle](paddlepaddle/) | PaddlePaddle framework model serving |
+
+## Deploying
+
+Each example can be deployed to Baseten with:
+
+```bash
+truss push
+```
diff --git a/infrastructure/autodesk-wala/README.md b/infrastructure/autodesk-wala/README.md
new file mode 100644
index 000000000..ab2d06db7
--- /dev/null
+++ b/infrastructure/autodesk-wala/README.md
@@ -0,0 +1,31 @@
+# ADSKAILab/WaLa-SV-1B
+
+Deploy ADSKAILab/WaLa-SV-1B using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Infrastructure / Custom server |
+| Engine | Custom (Truss) |
+| GPU | H100_40GB |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"image_b64": "<base64_encoded_image>"}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/infrastructure/autodesk-wala/config.yaml b/infrastructure/autodesk-wala/config.yaml
new file mode 100644
index 000000000..d5e1ea751
--- /dev/null
+++ b/infrastructure/autodesk-wala/config.yaml
@@ -0,0 +1,12 @@
+description: "Autodesk WaLa SV 1B via custom server"
+model_metadata:
+  repo_id: "ADSKAILab/WaLa-SV-1B"
+  example_model_input: {"image_b64": "<base64_encoded_image>", "output_format": "obj", "seed": 42}
+model_name: ADSKAILab/WaLa-SV-1B
+python_version: py311
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+requirements_file: ./requirements.txt # should cover all the dependencies for WaLa, taken from repo
+secrets:
+  hf_access_token: null
diff --git a/autodesk-wala/examples/README.md b/infrastructure/autodesk-wala/examples/README.md
similarity index 100%
rename from autodesk-wala/examples/README.md
rename to infrastructure/autodesk-wala/examples/README.md
diff --git a/autodesk-wala/examples/depth_maps_4/10.png b/infrastructure/autodesk-wala/examples/depth_maps_4/10.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_4/10.png
rename to infrastructure/autodesk-wala/examples/depth_maps_4/10.png
diff --git a/autodesk-wala/examples/depth_maps_4/26.png b/infrastructure/autodesk-wala/examples/depth_maps_4/26.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_4/26.png
rename to infrastructure/autodesk-wala/examples/depth_maps_4/26.png
diff --git a/autodesk-wala/examples/depth_maps_4/3.png b/infrastructure/autodesk-wala/examples/depth_maps_4/3.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_4/3.png
rename to infrastructure/autodesk-wala/examples/depth_maps_4/3.png
diff --git a/autodesk-wala/examples/depth_maps_4/6.png b/infrastructure/autodesk-wala/examples/depth_maps_4/6.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_4/6.png
rename to infrastructure/autodesk-wala/examples/depth_maps_4/6.png
diff --git a/autodesk-wala/examples/depth_maps_6/10.png b/infrastructure/autodesk-wala/examples/depth_maps_6/10.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_6/10.png
rename to infrastructure/autodesk-wala/examples/depth_maps_6/10.png
diff --git a/autodesk-wala/examples/depth_maps_6/26.png b/infrastructure/autodesk-wala/examples/depth_maps_6/26.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_6/26.png
rename to infrastructure/autodesk-wala/examples/depth_maps_6/26.png
diff --git a/autodesk-wala/examples/depth_maps_6/3.png b/infrastructure/autodesk-wala/examples/depth_maps_6/3.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_6/3.png
rename to infrastructure/autodesk-wala/examples/depth_maps_6/3.png
diff --git a/autodesk-wala/examples/depth_maps_6/49.png b/infrastructure/autodesk-wala/examples/depth_maps_6/49.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_6/49.png
rename to infrastructure/autodesk-wala/examples/depth_maps_6/49.png
diff --git a/autodesk-wala/examples/depth_maps_6/50.png b/infrastructure/autodesk-wala/examples/depth_maps_6/50.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_6/50.png
rename to infrastructure/autodesk-wala/examples/depth_maps_6/50.png
diff --git a/autodesk-wala/examples/depth_maps_6/6.png b/infrastructure/autodesk-wala/examples/depth_maps_6/6.png
similarity index 100%
rename from autodesk-wala/examples/depth_maps_6/6.png
rename to infrastructure/autodesk-wala/examples/depth_maps_6/6.png
diff --git a/autodesk-wala/examples/multi_view/003.png b/infrastructure/autodesk-wala/examples/multi_view/003.png
similarity index 100%
rename from autodesk-wala/examples/multi_view/003.png
rename to infrastructure/autodesk-wala/examples/multi_view/003.png
diff --git a/autodesk-wala/examples/multi_view/006.png b/infrastructure/autodesk-wala/examples/multi_view/006.png
similarity index 100%
rename from autodesk-wala/examples/multi_view/006.png
rename to infrastructure/autodesk-wala/examples/multi_view/006.png
diff --git a/autodesk-wala/examples/multi_view/010.png b/infrastructure/autodesk-wala/examples/multi_view/010.png
similarity index 100%
rename from autodesk-wala/examples/multi_view/010.png
rename to infrastructure/autodesk-wala/examples/multi_view/010.png
diff --git a/autodesk-wala/examples/multi_view/026.png b/infrastructure/autodesk-wala/examples/multi_view/026.png
similarity index 100%
rename from autodesk-wala/examples/multi_view/026.png
rename to infrastructure/autodesk-wala/examples/multi_view/026.png
diff --git a/autodesk-wala/examples/pointcloud/ring.h5df b/infrastructure/autodesk-wala/examples/pointcloud/ring.h5df
similarity index 100%
rename from autodesk-wala/examples/pointcloud/ring.h5df
rename to infrastructure/autodesk-wala/examples/pointcloud/ring.h5df
diff --git a/autodesk-wala/examples/single_depth_map/49.png b/infrastructure/autodesk-wala/examples/single_depth_map/49.png
similarity index 100%
rename from autodesk-wala/examples/single_depth_map/49.png
rename to infrastructure/autodesk-wala/examples/single_depth_map/49.png
diff --git a/autodesk-wala/examples/single_view/table.png b/infrastructure/autodesk-wala/examples/single_view/table.png
similarity index 100%
rename from autodesk-wala/examples/single_view/table.png
rename to infrastructure/autodesk-wala/examples/single_view/table.png
diff --git a/autodesk-wala/examples/sketch/tree.png b/infrastructure/autodesk-wala/examples/sketch/tree.png
similarity index 100%
rename from autodesk-wala/examples/sketch/tree.png
rename to infrastructure/autodesk-wala/examples/sketch/tree.png
diff --git a/autodesk-wala/examples/voxel/horse_16.json b/infrastructure/autodesk-wala/examples/voxel/horse_16.json
similarity index 100%
rename from autodesk-wala/examples/voxel/horse_16.json
rename to infrastructure/autodesk-wala/examples/voxel/horse_16.json
diff --git a/autodesk-wala/examples/voxel/horse_32.json b/infrastructure/autodesk-wala/examples/voxel/horse_32.json
similarity index 100%
rename from autodesk-wala/examples/voxel/horse_32.json
rename to infrastructure/autodesk-wala/examples/voxel/horse_32.json
diff --git a/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/__init__.py b/infrastructure/autodesk-wala/model/__init__.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/__init__.py
rename to infrastructure/autodesk-wala/model/__init__.py
diff --git a/autodesk-wala/model/model.py b/infrastructure/autodesk-wala/model/model.py
similarity index 100%
rename from autodesk-wala/model/model.py
rename to infrastructure/autodesk-wala/model/model.py
diff --git a/autodesk-wala/output.obj b/infrastructure/autodesk-wala/output.obj
similarity index 100%
rename from autodesk-wala/output.obj
rename to infrastructure/autodesk-wala/output.obj
diff --git a/mistral/mixtral-8x7b-instruct-vllm/model/__init__.py b/infrastructure/autodesk-wala/packages/src/__init__.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-vllm/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/__init__.py
diff --git a/autodesk-wala/packages/src/bpe_simple_vocab_16e6.txt.gz b/infrastructure/autodesk-wala/packages/src/bpe_simple_vocab_16e6.txt.gz
similarity index 100%
rename from autodesk-wala/packages/src/bpe_simple_vocab_16e6.txt.gz
rename to infrastructure/autodesk-wala/packages/src/bpe_simple_vocab_16e6.txt.gz
diff --git a/autodesk-wala/packages/src/clip_mod.py b/infrastructure/autodesk-wala/packages/src/clip_mod.py
similarity index 100%
rename from autodesk-wala/packages/src/clip_mod.py
rename to infrastructure/autodesk-wala/packages/src/clip_mod.py
diff --git a/autodesk-wala/packages/src/dataset_utils.py b/infrastructure/autodesk-wala/packages/src/dataset_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/dataset_utils.py
rename to infrastructure/autodesk-wala/packages/src/dataset_utils.py
diff --git a/mistral/pixtral-12b/model/__init__.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/__init__.py
similarity index 100%
rename from mistral/pixtral-12b/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/__init__.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/dwt.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/dwt.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/dwt.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/dwt.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/dwt_utils.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/dwt_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/dwt_utils.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/dwt_utils.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/fp16_util.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/fp16_util.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/fp16_util.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/fp16_util.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/gaussian_diffusion.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/gaussian_diffusion.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/gaussian_diffusion.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/gaussian_diffusion.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/latent_points.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/latent_points.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/latent_points.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/latent_points.py
diff --git a/multiprocessing/model/__init__.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/__init__ copy.py
similarity index 100%
rename from multiprocessing/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/__init__ copy.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/__init__.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/__init__.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/__init__.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/ball_query.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/ball_query.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/ball_query.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/ball_query.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/frustum.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/frustum.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/frustum.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/frustum.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/__init__.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/__init__.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/__init__.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/backend.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/backend.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/backend.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/backend.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/ball_query.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/ball_query.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/ball_query.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/ball_query.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/devoxelization.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/devoxelization.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/devoxelization.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/devoxelization.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/grouping.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/grouping.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/grouping.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/grouping.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/interpolatation.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/interpolatation.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/interpolatation.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/interpolatation.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/loss.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/loss.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/loss.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/loss.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/sampling.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/sampling.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/sampling.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/sampling.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cu b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cu
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cu
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cu
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cuh b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cuh
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cuh
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.cuh
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.hpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.hpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.hpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/ball_query/ball_query.hpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/bindings.cpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/bindings.cpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/bindings.cpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/bindings.cpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/cuda_utils.cuh b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/cuda_utils.cuh
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/cuda_utils.cuh
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/cuda_utils.cuh
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cu b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cu
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cu
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cu
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cuh b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cuh
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cuh
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.cuh
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.hpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.hpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.hpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/grouping/grouping.hpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cu b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cu
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cu
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cu
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cuh b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cuh
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cuh
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.cuh
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.hpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.hpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.hpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/neighbor_interpolate.hpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cu b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cu
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cu
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cu
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cuh b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cuh
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cuh
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.cuh
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.hpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.hpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.hpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/interpolate/trilinear_devox.hpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cu b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cu
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cu
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cu
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cuh b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cuh
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cuh
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.cuh
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.hpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.hpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.hpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/sampling/sampling.hpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/utils.hpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/utils.hpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/utils.hpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/utils.hpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cu b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cu
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cu
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cu
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cuh b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cuh
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cuh
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.cuh
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.hpp b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.hpp
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.hpp
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/src/voxelization/vox.hpp
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/functional/voxelization.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/voxelization.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/functional/voxelization.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/functional/voxelization.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/loss.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/loss.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/loss.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/loss.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/pointnet.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/pointnet.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/pointnet.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/pointnet.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/pvconv.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/pvconv.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/pvconv.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/pvconv.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/se.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/se.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/se.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/se.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/shared_mlp.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/shared_mlp.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/shared_mlp.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/shared_mlp.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/modules/voxelization.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/voxelization.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/modules/voxelization.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/modules/voxelization.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/network_ae.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/network_ae.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/network_ae.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/network_ae.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/nn.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/nn.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/nn.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/nn.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/point_voxels.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/point_voxels.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/point_voxels.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/point_voxels.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/resample.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/resample.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/resample.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/resample.py
diff --git a/autodesk-wala/packages/src/diffusion_modules/sparse_network.py b/infrastructure/autodesk-wala/packages/src/diffusion_modules/sparse_network.py
similarity index 100%
rename from autodesk-wala/packages/src/diffusion_modules/sparse_network.py
rename to infrastructure/autodesk-wala/packages/src/diffusion_modules/sparse_network.py
diff --git a/musicgen-large/model/__init__.py b/infrastructure/autodesk-wala/packages/src/experiments/__init__.py
similarity index 100%
rename from musicgen-large/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/experiments/__init__.py
diff --git a/musicgen-melody/model/__init__.py b/infrastructure/autodesk-wala/packages/src/experiments/utils/__init__.py
similarity index 100%
rename from musicgen-melody/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/experiments/utils/__init__.py
diff --git a/autodesk-wala/packages/src/experiments/utils/wavelet_utils.py b/infrastructure/autodesk-wala/packages/src/experiments/utils/wavelet_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/experiments/utils/wavelet_utils.py
rename to infrastructure/autodesk-wala/packages/src/experiments/utils/wavelet_utils.py
diff --git a/ngram-speculator/truss/model/__init__.py b/infrastructure/autodesk-wala/packages/src/latent_model/__init__.py
similarity index 100%
rename from ngram-speculator/truss/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/__init__.py
diff --git a/autodesk-wala/packages/src/latent_model/abstract_volume_nn.py b/infrastructure/autodesk-wala/packages/src/latent_model/abstract_volume_nn.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/abstract_volume_nn.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/abstract_volume_nn.py
diff --git a/autodesk-wala/packages/src/latent_model/continous_diffusion_interface.py b/infrastructure/autodesk-wala/packages/src/latent_model/continous_diffusion_interface.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/continous_diffusion_interface.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/continous_diffusion_interface.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/dwt.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/dwt.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/dwt.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/dwt.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/dwt_utils.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/dwt_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/dwt_utils.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/dwt_utils.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/fp16_util.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/fp16_util.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/fp16_util.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/fp16_util.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion_wavelet.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion_wavelet.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion_wavelet.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/gaussian_diffusion_wavelet.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/nn.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/nn.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/nn.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/nn.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/resample.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/resample.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/resample.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/resample.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/sparse_network.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/sparse_network.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/sparse_network.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/sparse_network.py
diff --git a/autodesk-wala/packages/src/latent_model/diffusion_modules/utils.py b/infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/utils.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/diffusion_modules/utils.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/diffusion_modules/utils.py
diff --git a/autodesk-wala/packages/src/latent_model/dit_utils.py b/infrastructure/autodesk-wala/packages/src/latent_model/dit_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/dit_utils.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/dit_utils.py
diff --git a/autodesk-wala/packages/src/latent_model/gaussian_diffusion.py b/infrastructure/autodesk-wala/packages/src/latent_model/gaussian_diffusion.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/gaussian_diffusion.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/gaussian_diffusion.py
diff --git a/autodesk-wala/packages/src/latent_model/latent_dit_utils.py b/infrastructure/autodesk-wala/packages/src/latent_model/latent_dit_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/latent_dit_utils.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/latent_dit_utils.py
diff --git a/autodesk-wala/packages/src/latent_model/latent_uvit_utils.py b/infrastructure/autodesk-wala/packages/src/latent_model/latent_uvit_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/latent_uvit_utils.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/latent_uvit_utils.py
diff --git a/autodesk-wala/packages/src/latent_model/points_network.py b/infrastructure/autodesk-wala/packages/src/latent_model/points_network.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/points_network.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/points_network.py
diff --git a/autodesk-wala/packages/src/latent_model/quantize.py b/infrastructure/autodesk-wala/packages/src/latent_model/quantize.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/quantize.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/quantize.py
diff --git a/autodesk-wala/packages/src/latent_model/utils.py b/infrastructure/autodesk-wala/packages/src/latent_model/utils.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/utils.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/utils.py
diff --git a/autodesk-wala/packages/src/latent_model/voxels_network.py b/infrastructure/autodesk-wala/packages/src/latent_model/voxels_network.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/voxels_network.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/voxels_network.py
diff --git a/autodesk-wala/packages/src/latent_model/wavelet_utils.py b/infrastructure/autodesk-wala/packages/src/latent_model/wavelet_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/wavelet_utils.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/wavelet_utils.py
diff --git a/autodesk-wala/packages/src/latent_model/wavelet_vq_model.py b/infrastructure/autodesk-wala/packages/src/latent_model/wavelet_vq_model.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_model/wavelet_vq_model.py
rename to infrastructure/autodesk-wala/packages/src/latent_model/wavelet_vq_model.py
diff --git a/autodesk-wala/packages/src/latent_module.py b/infrastructure/autodesk-wala/packages/src/latent_module.py
similarity index 100%
rename from autodesk-wala/packages/src/latent_module.py
rename to infrastructure/autodesk-wala/packages/src/latent_module.py
diff --git a/autodesk-wala/packages/src/model_utils.py b/infrastructure/autodesk-wala/packages/src/model_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/model_utils.py
rename to infrastructure/autodesk-wala/packages/src/model_utils.py
diff --git a/autodesk-wala/packages/src/mvdream/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/__init__.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/__init__.py
diff --git a/autodesk-wala/packages/src/mvdream/callbacks.py b/infrastructure/autodesk-wala/packages/src/mvdream/callbacks.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/callbacks.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/callbacks.py
diff --git a/autodesk-wala/packages/src/mvdream/camera_utils.py b/infrastructure/autodesk-wala/packages/src/mvdream/camera_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/camera_utils.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/camera_utils.py
diff --git a/autodesk-wala/packages/src/mvdream/coco_prompts.py b/infrastructure/autodesk-wala/packages/src/mvdream/coco_prompts.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/coco_prompts.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/coco_prompts.py
diff --git a/autodesk-wala/packages/src/mvdream/configs/sd-v1.yaml b/infrastructure/autodesk-wala/packages/src/mvdream/configs/sd-v1.yaml
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/configs/sd-v1.yaml
rename to infrastructure/autodesk-wala/packages/src/mvdream/configs/sd-v1.yaml
diff --git a/autodesk-wala/packages/src/mvdream/configs/sd-v2-base.yaml b/infrastructure/autodesk-wala/packages/src/mvdream/configs/sd-v2-base.yaml
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/configs/sd-v2-base.yaml
rename to infrastructure/autodesk-wala/packages/src/mvdream/configs/sd-v2-base.yaml
diff --git a/autodesk-wala/packages/src/mvdream/constants.py b/infrastructure/autodesk-wala/packages/src/mvdream/constants.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/constants.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/constants.py
diff --git a/autodesk-wala/packages/src/mvdream/helper.py b/infrastructure/autodesk-wala/packages/src/mvdream/helper.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/helper.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/helper.py
diff --git a/nous-capybara/nous-capybara-34b-openai/model/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/__init__.py
similarity index 100%
rename from nous-capybara/nous-capybara-34b-openai/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/__init__.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/interface.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/interface.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/interface.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/interface.py
diff --git a/nous-capybara/nous-capybara-34b/model/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/__init__.py
similarity index 100%
rename from nous-capybara/nous-capybara-34b/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/__init__.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/models/autoencoder.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/autoencoder.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/models/autoencoder.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/autoencoder.py
diff --git a/nsql/model/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/__init__.py
similarity index 100%
rename from nsql/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/__init__.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/ddim.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/ddim.py
similarity index 99%
rename from autodesk-wala/packages/src/mvdream/ldm/models/diffusion/ddim.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/ddim.py
index 922655006..06ac62a88 100644
--- a/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/ddim.py
+++ b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/models/diffusion/ddim.py
@@ -38,8 +38,8 @@ def make_schedule(
         assert alphas_cumprod.shape[0] == self.ddpm_num_timesteps, (
             "alphas have to be defined for each timestep"
         )
-        to_torch = (
-            lambda x: x.clone().detach().to(torch.float32).to(self.model.betas.device)
+        to_torch = lambda x: (
+            x.clone().detach().to(torch.float32).to(self.model.betas.device)
         )
 
         self.register_buffer("betas", to_torch(self.model.betas))
diff --git a/nvidia/parakeet-tdt-0_6b-v2/model/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/__init__.py
similarity index 100%
rename from nvidia/parakeet-tdt-0_6b-v2/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/__init__.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/attention.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/attention.py
similarity index 98%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/attention.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/attention.py
index a5d68e8ea..d957f8acc 100644
--- a/autodesk-wala/packages/src/mvdream/ldm/modules/attention.py
+++ b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/attention.py
@@ -220,11 +220,13 @@ def forward(self, x, context=None, mask=None):
 
         b, _, _ = q.shape
         q, k, v = map(
-            lambda t: t.unsqueeze(3)
-            .reshape(b, t.shape[1], self.heads, self.dim_head)
-            .permute(0, 2, 1, 3)
-            .reshape(b * self.heads, t.shape[1], self.dim_head)
-            .contiguous(),
+            lambda t: (
+                t.unsqueeze(3)
+                .reshape(b, t.shape[1], self.heads, self.dim_head)
+                .permute(0, 2, 1, 3)
+                .reshape(b * self.heads, t.shape[1], self.dim_head)
+                .contiguous()
+            ),
             (q, k, v),
         )
 
diff --git a/phi/phi-3-mini-128k-instruct/model/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/__init__.py
similarity index 100%
rename from phi/phi-3-mini-128k-instruct/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/__init__.py
diff --git a/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/model.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/model.py
new file mode 100644
index 000000000..79c4885ea
--- /dev/null
+++ b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/model.py
@@ -0,0 +1,1022 @@
+# pytorch_diffusion + derived encoder decoder
+import math
+import torch
+import torch.nn as nn
+import numpy as np
+from einops import rearrange
+from typing import Optional, Any
+
+from ..attention import MemoryEfficientCrossAttention
+
+try:
+    import xformers
+    import xformers.ops
+
+    XFORMERS_IS_AVAILBLE = True
+except:
+    XFORMERS_IS_AVAILBLE = False
+    print("No module 'xformers'. Proceeding without it.")
+
+
+def get_timestep_embedding(timesteps, embedding_dim):
+    """
+    This matches the implementation in Denoising Diffusion Probabilistic Models:
+    From Fairseq.
+    Build sinusoidal embeddings.
+    This matches the implementation in tensor2tensor, but differs slightly
+    from the description in Section 3.5 of "Attention Is All You Need".
+    """
+    assert len(timesteps.shape) == 1
+
+    half_dim = embedding_dim // 2
+    emb = math.log(10000) / (half_dim - 1)
+    # emb = torch.exp(torch.arange(half_dim, dtype=torch.float32) * -emb)
+    emb = torch.exp(torch.arange(half_dim, dtype=torch.bfloat16) * -emb)
+    emb = emb.to(device=timesteps.device)
+    # emb = timesteps.float()[:, None] * emb[None, :]
+    emb = timesteps[:, None] * emb[None, :]
+    emb = torch.cat([torch.sin(emb), torch.cos(emb)], dim=1)
+    if embedding_dim % 2 == 1:  # zero pad
+        emb = torch.nn.functional.pad(emb, (0, 1, 0, 0))
+    return emb
+
+
+def nonlinearity(x):
+    # swish
+    return x * torch.sigmoid(x)
+
+
+def Normalize(in_channels, num_groups=32):
+    return torch.nn.GroupNorm(
+        num_groups=num_groups, num_channels=in_channels, eps=1e-6, affine=True
+    )
+
+
+class Upsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=3, stride=1, padding=1
+            )
+
+    def forward(self, x):
+        x = torch.nn.functional.interpolate(x, scale_factor=2.0, mode="nearest")
+        if self.with_conv:
+            x = self.conv(x)
+        return x
+
+
+class Downsample(nn.Module):
+    def __init__(self, in_channels, with_conv):
+        super().__init__()
+        self.with_conv = with_conv
+        if self.with_conv:
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=3, stride=2, padding=0
+            )
+
+    def forward(self, x):
+        if self.with_conv:
+            pad = (0, 1, 0, 1)
+            x = torch.nn.functional.pad(x, pad, mode="constant", value=0)
+            x = self.conv(x)
+        else:
+            x = torch.nn.functional.avg_pool2d(x, kernel_size=2, stride=2)
+        return x
+
+
+class ResnetBlock(nn.Module):
+    def __init__(
+        self,
+        *,
+        in_channels,
+        out_channels=None,
+        conv_shortcut=False,
+        dropout,
+        temb_channels=512,
+    ):
+        super().__init__()
+        self.in_channels = in_channels
+        out_channels = in_channels if out_channels is None else out_channels
+        self.out_channels = out_channels
+        self.use_conv_shortcut = conv_shortcut
+
+        self.norm1 = Normalize(in_channels)
+        self.conv1 = torch.nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if temb_channels > 0:
+            self.temb_proj = torch.nn.Linear(temb_channels, out_channels)
+        self.norm2 = Normalize(out_channels)
+        self.dropout = torch.nn.Dropout(dropout)
+        self.conv2 = torch.nn.Conv2d(
+            out_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                self.conv_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=3, stride=1, padding=1
+                )
+            else:
+                self.nin_shortcut = torch.nn.Conv2d(
+                    in_channels, out_channels, kernel_size=1, stride=1, padding=0
+                )
+
+    def forward(self, x, temb):
+        h = x
+        h = self.norm1(h)
+        h = nonlinearity(h)
+        h = self.conv1(h)
+
+        if temb is not None:
+            h = h + self.temb_proj(nonlinearity(temb))[:, :, None, None]
+
+        h = self.norm2(h)
+        h = nonlinearity(h)
+        h = self.dropout(h)
+        h = self.conv2(h)
+
+        if self.in_channels != self.out_channels:
+            if self.use_conv_shortcut:
+                x = self.conv_shortcut(x)
+            else:
+                x = self.nin_shortcut(x)
+
+        return x + h
+
+
+class AttnBlock(nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        b, c, h, w = q.shape
+        q = q.reshape(b, c, h * w)
+        q = q.permute(0, 2, 1)  # b,hw,c
+        k = k.reshape(b, c, h * w)  # b,c,hw
+        w_ = torch.bmm(q, k)  # b,hw,hw    w[b,i,j]=sum_c q[b,i,c]k[b,c,j]
+        w_ = w_ * (int(c) ** (-0.5))
+        w_ = torch.nn.functional.softmax(w_, dim=2)
+
+        # attend to values
+        v = v.reshape(b, c, h * w)
+        w_ = w_.permute(0, 2, 1)  # b,hw,hw (first hw of k, second of q)
+        h_ = torch.bmm(v, w_)  # b, c,hw (hw of q) h_[b,c,j] = sum_i v[b,c,i] w_[b,i,j]
+        h_ = h_.reshape(b, c, h, w)
+
+        h_ = self.proj_out(h_)
+
+        return x + h_
+
+
+class MemoryEfficientAttnBlock(nn.Module):
+    """
+    Uses xformers efficient implementation,
+    see https://github.com/MatthieuTPHR/diffusers/blob/d80b531ff8060ec1ea982b65a1b8df70f73aa67c/src/diffusers/models/attention.py#L223
+    Note: this is a single-head self-attention operation
+    """
+
+    #
+    def __init__(self, in_channels):
+        super().__init__()
+        self.in_channels = in_channels
+
+        self.norm = Normalize(in_channels)
+        self.q = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.k = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.v = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.proj_out = torch.nn.Conv2d(
+            in_channels, in_channels, kernel_size=1, stride=1, padding=0
+        )
+        self.attention_op: Optional[Any] = None
+
+    def forward(self, x):
+        h_ = x
+        h_ = self.norm(h_)
+        q = self.q(h_)
+        k = self.k(h_)
+        v = self.v(h_)
+
+        # compute attention
+        B, C, H, W = q.shape
+        q, k, v = map(lambda x: rearrange(x, "b c h w -> b (h w) c"), (q, k, v))
+
+        q, k, v = map(
+            lambda t: (
+                t.unsqueeze(3)
+                .reshape(B, t.shape[1], 1, C)
+                .permute(0, 2, 1, 3)
+                .reshape(B * 1, t.shape[1], C)
+                .contiguous()
+            ),
+            (q, k, v),
+        )
+        out = xformers.ops.memory_efficient_attention(
+            q, k, v, attn_bias=None, op=self.attention_op
+        )
+
+        out = (
+            out.unsqueeze(0)
+            .reshape(B, 1, out.shape[1], C)
+            .permute(0, 2, 1, 3)
+            .reshape(B, out.shape[1], C)
+        )
+        out = rearrange(out, "b (h w) c -> b c h w", b=B, h=H, w=W, c=C)
+        out = self.proj_out(out)
+        return x + out
+
+
+class MemoryEfficientCrossAttentionWrapper(MemoryEfficientCrossAttention):
+    def forward(self, x, context=None, mask=None):
+        b, c, h, w = x.shape
+        x = rearrange(x, "b c h w -> b (h w) c")
+        out = super().forward(x, context=context, mask=mask)
+        out = rearrange(out, "b (h w) c -> b c h w", h=h, w=w, c=c)
+        return x + out
+
+
+def make_attn(in_channels, attn_type="vanilla", attn_kwargs=None):
+    assert attn_type in [
+        "vanilla",
+        "vanilla-xformers",
+        "memory-efficient-cross-attn",
+        "linear",
+        "none",
+    ], f"attn_type {attn_type} unknown"
+    if XFORMERS_IS_AVAILBLE and attn_type == "vanilla":
+        attn_type = "vanilla-xformers"
+    print(f"making attention of type '{attn_type}' with {in_channels} in_channels")
+    if attn_type == "vanilla":
+        assert attn_kwargs is None
+        return AttnBlock(in_channels)
+    elif attn_type == "vanilla-xformers":
+        print(f"building MemoryEfficientAttnBlock with {in_channels} in_channels...")
+        return MemoryEfficientAttnBlock(in_channels)
+    elif type == "memory-efficient-cross-attn":
+        attn_kwargs["query_dim"] = in_channels
+        return MemoryEfficientCrossAttentionWrapper(**attn_kwargs)
+    elif attn_type == "none":
+        return nn.Identity(in_channels)
+    else:
+        raise NotImplementedError()
+
+
+class Model(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        use_timestep=True,
+        use_linear_attn=False,
+        attn_type="vanilla",
+    ):
+        super().__init__()
+        if use_linear_attn:
+            attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = self.ch * 4
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        self.use_timestep = use_timestep
+        if self.use_timestep:
+            # timestep embedding
+            self.temb = nn.Module()
+            self.temb.dense = nn.ModuleList(
+                [
+                    torch.nn.Linear(self.ch, self.temb_ch),
+                    torch.nn.Linear(self.temb_ch, self.temb_ch),
+                ]
+            )
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(
+            in_channels, self.ch, kernel_size=3, stride=1, padding=1
+        )
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            skip_in = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                if i_block == self.num_res_blocks:
+                    skip_in = ch * in_ch_mult[i_level]
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in + skip_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, out_ch, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x, t=None, context=None):
+        # assert x.shape[2] == x.shape[3] == self.resolution
+        if context is not None:
+            # assume aligned context, cat along channel axis
+            x = torch.cat((x, context), dim=1)
+        if self.use_timestep:
+            # timestep embedding
+            assert t is not None
+            temb = get_timestep_embedding(t, self.ch)
+            temb = self.temb.dense[0](temb)
+            temb = nonlinearity(temb)
+            temb = self.temb.dense[1](temb)
+        else:
+            temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](
+                    torch.cat([h, hs.pop()], dim=1), temb
+                )
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+    def get_last_layer(self):
+        return self.conv_out.weight
+
+
+class Encoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        double_z=True,
+        use_linear_attn=False,
+        attn_type="vanilla",
+        **ignore_kwargs,
+    ):
+        super().__init__()
+        if use_linear_attn:
+            attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+
+        # downsampling
+        self.conv_in = torch.nn.Conv2d(
+            in_channels, self.ch, kernel_size=3, stride=1, padding=1
+        )
+
+        curr_res = resolution
+        in_ch_mult = (1,) + tuple(ch_mult)
+        self.in_ch_mult = in_ch_mult
+        self.down = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_in = ch * in_ch_mult[i_level]
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            down = nn.Module()
+            down.block = block
+            down.attn = attn
+            if i_level != self.num_resolutions - 1:
+                down.downsample = Downsample(block_in, resamp_with_conv)
+                curr_res = curr_res // 2
+            self.down.append(down)
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in,
+            2 * z_channels if double_z else z_channels,
+            kernel_size=3,
+            stride=1,
+            padding=1,
+        )
+
+    def forward(self, x):
+        # timestep embedding
+        temb = None
+
+        # downsampling
+        hs = [self.conv_in(x)]
+        for i_level in range(self.num_resolutions):
+            for i_block in range(self.num_res_blocks):
+                h = self.down[i_level].block[i_block](hs[-1], temb)
+                if len(self.down[i_level].attn) > 0:
+                    h = self.down[i_level].attn[i_block](h)
+                hs.append(h)
+            if i_level != self.num_resolutions - 1:
+                hs.append(self.down[i_level].downsample(hs[-1]))
+
+        # middle
+        h = hs[-1]
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # end
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class Decoder(nn.Module):
+    def __init__(
+        self,
+        *,
+        ch,
+        out_ch,
+        ch_mult=(1, 2, 4, 8),
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        in_channels,
+        resolution,
+        z_channels,
+        give_pre_end=False,
+        tanh_out=False,
+        use_linear_attn=False,
+        attn_type="vanilla",
+        **ignorekwargs,
+    ):
+        super().__init__()
+        if use_linear_attn:
+            attn_type = "linear"
+        self.ch = ch
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        self.resolution = resolution
+        self.in_channels = in_channels
+        self.give_pre_end = give_pre_end
+        self.tanh_out = tanh_out
+
+        # compute in_ch_mult, block_in and curr_res at lowest res
+        in_ch_mult = (1,) + tuple(ch_mult)
+        block_in = ch * ch_mult[self.num_resolutions - 1]
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.z_shape = (1, z_channels, curr_res, curr_res)
+        print(
+            "Working with z of shape {} = {} dimensions.".format(
+                self.z_shape, np.prod(self.z_shape)
+            )
+        )
+
+        # z to block_in
+        self.conv_in = torch.nn.Conv2d(
+            z_channels, block_in, kernel_size=3, stride=1, padding=1
+        )
+
+        # middle
+        self.mid = nn.Module()
+        self.mid.block_1 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+        self.mid.attn_1 = make_attn(block_in, attn_type=attn_type)
+        self.mid.block_2 = ResnetBlock(
+            in_channels=block_in,
+            out_channels=block_in,
+            temb_channels=self.temb_ch,
+            dropout=dropout,
+        )
+
+        # upsampling
+        self.up = nn.ModuleList()
+        for i_level in reversed(range(self.num_resolutions)):
+            block = nn.ModuleList()
+            attn = nn.ModuleList()
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+                if curr_res in attn_resolutions:
+                    attn.append(make_attn(block_in, attn_type=attn_type))
+            up = nn.Module()
+            up.block = block
+            up.attn = attn
+            if i_level != 0:
+                up.upsample = Upsample(block_in, resamp_with_conv)
+                curr_res = curr_res * 2
+            self.up.insert(0, up)  # prepend to get consistent order
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, out_ch, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, z):
+        # assert z.shape[1:] == self.z_shape[1:]
+        self.last_z_shape = z.shape
+
+        # timestep embedding
+        temb = None
+
+        # z to block_in
+        h = self.conv_in(z)
+
+        # middle
+        h = self.mid.block_1(h, temb)
+        h = self.mid.attn_1(h)
+        h = self.mid.block_2(h, temb)
+
+        # upsampling
+        for i_level in reversed(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.up[i_level].block[i_block](h, temb)
+                if len(self.up[i_level].attn) > 0:
+                    h = self.up[i_level].attn[i_block](h)
+            if i_level != 0:
+                h = self.up[i_level].upsample(h)
+
+        # end
+        if self.give_pre_end:
+            return h
+
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        if self.tanh_out:
+            h = torch.tanh(h)
+        return h
+
+
+class SimpleDecoder(nn.Module):
+    def __init__(self, in_channels, out_channels, *args, **kwargs):
+        super().__init__()
+        self.model = nn.ModuleList(
+            [
+                nn.Conv2d(in_channels, in_channels, 1),
+                ResnetBlock(
+                    in_channels=in_channels,
+                    out_channels=2 * in_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                ),
+                ResnetBlock(
+                    in_channels=2 * in_channels,
+                    out_channels=4 * in_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                ),
+                ResnetBlock(
+                    in_channels=4 * in_channels,
+                    out_channels=2 * in_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                ),
+                nn.Conv2d(2 * in_channels, in_channels, 1),
+                Upsample(in_channels, with_conv=True),
+            ]
+        )
+        # end
+        self.norm_out = Normalize(in_channels)
+        self.conv_out = torch.nn.Conv2d(
+            in_channels, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x):
+        for i, layer in enumerate(self.model):
+            if i in [1, 2, 3]:
+                x = layer(x, None)
+            else:
+                x = layer(x)
+
+        h = self.norm_out(x)
+        h = nonlinearity(h)
+        x = self.conv_out(h)
+        return x
+
+
+class UpsampleDecoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        out_channels,
+        ch,
+        num_res_blocks,
+        resolution,
+        ch_mult=(2, 2),
+        dropout=0.0,
+    ):
+        super().__init__()
+        # upsampling
+        self.temb_ch = 0
+        self.num_resolutions = len(ch_mult)
+        self.num_res_blocks = num_res_blocks
+        block_in = in_channels
+        curr_res = resolution // 2 ** (self.num_resolutions - 1)
+        self.res_blocks = nn.ModuleList()
+        self.upsample_blocks = nn.ModuleList()
+        for i_level in range(self.num_resolutions):
+            res_block = []
+            block_out = ch * ch_mult[i_level]
+            for i_block in range(self.num_res_blocks + 1):
+                res_block.append(
+                    ResnetBlock(
+                        in_channels=block_in,
+                        out_channels=block_out,
+                        temb_channels=self.temb_ch,
+                        dropout=dropout,
+                    )
+                )
+                block_in = block_out
+            self.res_blocks.append(nn.ModuleList(res_block))
+            if i_level != self.num_resolutions - 1:
+                self.upsample_blocks.append(Upsample(block_in, True))
+                curr_res = curr_res * 2
+
+        # end
+        self.norm_out = Normalize(block_in)
+        self.conv_out = torch.nn.Conv2d(
+            block_in, out_channels, kernel_size=3, stride=1, padding=1
+        )
+
+    def forward(self, x):
+        # upsampling
+        h = x
+        for k, i_level in enumerate(range(self.num_resolutions)):
+            for i_block in range(self.num_res_blocks + 1):
+                h = self.res_blocks[i_level][i_block](h, None)
+            if i_level != self.num_resolutions - 1:
+                h = self.upsample_blocks[k](h)
+        h = self.norm_out(h)
+        h = nonlinearity(h)
+        h = self.conv_out(h)
+        return h
+
+
+class LatentRescaler(nn.Module):
+    def __init__(self, factor, in_channels, mid_channels, out_channels, depth=2):
+        super().__init__()
+        # residual block, interpolate, residual block
+        self.factor = factor
+        self.conv_in = nn.Conv2d(
+            in_channels, mid_channels, kernel_size=3, stride=1, padding=1
+        )
+        self.res_block1 = nn.ModuleList(
+            [
+                ResnetBlock(
+                    in_channels=mid_channels,
+                    out_channels=mid_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                )
+                for _ in range(depth)
+            ]
+        )
+        self.attn = AttnBlock(mid_channels)
+        self.res_block2 = nn.ModuleList(
+            [
+                ResnetBlock(
+                    in_channels=mid_channels,
+                    out_channels=mid_channels,
+                    temb_channels=0,
+                    dropout=0.0,
+                )
+                for _ in range(depth)
+            ]
+        )
+
+        self.conv_out = nn.Conv2d(
+            mid_channels,
+            out_channels,
+            kernel_size=1,
+        )
+
+    def forward(self, x):
+        x = self.conv_in(x)
+        for block in self.res_block1:
+            x = block(x, None)
+        x = torch.nn.functional.interpolate(
+            x,
+            size=(
+                int(round(x.shape[2] * self.factor)),
+                int(round(x.shape[3] * self.factor)),
+            ),
+        )
+        x = self.attn(x)
+        for block in self.res_block2:
+            x = block(x, None)
+        x = self.conv_out(x)
+        return x
+
+
+class MergedRescaleEncoder(nn.Module):
+    def __init__(
+        self,
+        in_channels,
+        ch,
+        resolution,
+        out_ch,
+        num_res_blocks,
+        attn_resolutions,
+        dropout=0.0,
+        resamp_with_conv=True,
+        ch_mult=(1, 2, 4, 8),
+        rescale_factor=1.0,
+        rescale_module_depth=1,
+    ):
+        super().__init__()
+        intermediate_chn = ch * ch_mult[-1]
+        self.encoder = Encoder(
+            in_channels=in_channels,
+            num_res_blocks=num_res_blocks,
+            ch=ch,
+            ch_mult=ch_mult,
+            z_channels=intermediate_chn,
+            double_z=False,
+            resolution=resolution,
+            attn_resolutions=attn_resolutions,
+            dropout=dropout,
+            resamp_with_conv=resamp_with_conv,
+            out_ch=None,
+        )
+        self.rescaler = LatentRescaler(
+            factor=rescale_factor,
+            in_channels=intermediate_chn,
+            mid_channels=intermediate_chn,
+            out_channels=out_ch,
+            depth=rescale_module_depth,
+        )
+
+    def forward(self, x):
+        x = self.encoder(x)
+        x = self.rescaler(x)
+        return x
+
+
+class MergedRescaleDecoder(nn.Module):
+    def __init__(
+        self,
+        z_channels,
+        out_ch,
+        resolution,
+        num_res_blocks,
+        attn_resolutions,
+        ch,
+        ch_mult=(1, 2, 4, 8),
+        dropout=0.0,
+        resamp_with_conv=True,
+        rescale_factor=1.0,
+        rescale_module_depth=1,
+    ):
+        super().__init__()
+        tmp_chn = z_channels * ch_mult[-1]
+        self.decoder = Decoder(
+            out_ch=out_ch,
+            z_channels=tmp_chn,
+            attn_resolutions=attn_resolutions,
+            dropout=dropout,
+            resamp_with_conv=resamp_with_conv,
+            in_channels=None,
+            num_res_blocks=num_res_blocks,
+            ch_mult=ch_mult,
+            resolution=resolution,
+            ch=ch,
+        )
+        self.rescaler = LatentRescaler(
+            factor=rescale_factor,
+            in_channels=z_channels,
+            mid_channels=tmp_chn,
+            out_channels=tmp_chn,
+            depth=rescale_module_depth,
+        )
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Upsampler(nn.Module):
+    def __init__(self, in_size, out_size, in_channels, out_channels, ch_mult=2):
+        super().__init__()
+        assert out_size >= in_size
+        num_blocks = int(np.log2(out_size // in_size)) + 1
+        factor_up = 1.0 + (out_size % in_size)
+        print(
+            f"Building {self.__class__.__name__} with in_size: {in_size} --> out_size {out_size} and factor {factor_up}"
+        )
+        self.rescaler = LatentRescaler(
+            factor=factor_up,
+            in_channels=in_channels,
+            mid_channels=2 * in_channels,
+            out_channels=in_channels,
+        )
+        self.decoder = Decoder(
+            out_ch=out_channels,
+            resolution=out_size,
+            z_channels=in_channels,
+            num_res_blocks=2,
+            attn_resolutions=[],
+            in_channels=None,
+            ch=in_channels,
+            ch_mult=[ch_mult for _ in range(num_blocks)],
+        )
+
+    def forward(self, x):
+        x = self.rescaler(x)
+        x = self.decoder(x)
+        return x
+
+
+class Resize(nn.Module):
+    def __init__(self, in_channels=None, learned=False, mode="bilinear"):
+        super().__init__()
+        self.with_conv = learned
+        self.mode = mode
+        if self.with_conv:
+            print(
+                f"Note: {self.__class__.__name} uses learned downsampling and will ignore the fixed {mode} mode"
+            )
+            raise NotImplementedError()
+            assert in_channels is not None
+            # no asymmetric padding in torch conv, must do it ourselves
+            self.conv = torch.nn.Conv2d(
+                in_channels, in_channels, kernel_size=4, stride=2, padding=1
+            )
+
+    def forward(self, x, scale_factor=1.0):
+        if scale_factor == 1.0:
+            return x
+        else:
+            x = torch.nn.functional.interpolate(
+                x, mode=self.mode, align_corners=False, scale_factor=scale_factor
+            )
+        return x
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/openaimodel.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/openaimodel.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/openaimodel.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/openaimodel.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/util.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/util.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/util.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/diffusionmodules/util.py
diff --git a/phi/phi-3-mini-4k-instruct/model/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/distributions/__init__.py
similarity index 100%
rename from phi/phi-3-mini-4k-instruct/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/distributions/__init__.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/distributions/distributions.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/distributions/distributions.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/distributions/distributions.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/distributions/distributions.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/ema.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/ema.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/ema.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/ema.py
diff --git a/phi/phi-3.5-mini/model/__init__.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/encoders/__init__.py
similarity index 100%
rename from phi/phi-3.5-mini/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/encoders/__init__.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/modules/encoders/modules.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/encoders/modules.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/modules/encoders/modules.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/modules/encoders/modules.py
diff --git a/autodesk-wala/packages/src/mvdream/ldm/util.py b/infrastructure/autodesk-wala/packages/src/mvdream/ldm/util.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/ldm/util.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/ldm/util.py
diff --git a/autodesk-wala/packages/src/mvdream/model_zoo.py b/infrastructure/autodesk-wala/packages/src/mvdream/model_zoo.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/model_zoo.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/model_zoo.py
diff --git a/autodesk-wala/packages/src/mvdream/trainer_old.py b/infrastructure/autodesk-wala/packages/src/mvdream/trainer_old.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/trainer_old.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/trainer_old.py
diff --git a/autodesk-wala/packages/src/mvdream/trainermodule.py b/infrastructure/autodesk-wala/packages/src/mvdream/trainermodule.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream/trainermodule.py
rename to infrastructure/autodesk-wala/packages/src/mvdream/trainermodule.py
diff --git a/autodesk-wala/packages/src/mvdream_module.py b/infrastructure/autodesk-wala/packages/src/mvdream_module.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream_module.py
rename to infrastructure/autodesk-wala/packages/src/mvdream_module.py
diff --git a/autodesk-wala/packages/src/mvdream_utils.py b/infrastructure/autodesk-wala/packages/src/mvdream_utils.py
similarity index 100%
rename from autodesk-wala/packages/src/mvdream_utils.py
rename to infrastructure/autodesk-wala/packages/src/mvdream_utils.py
diff --git a/autodesk-wala/packages/src/networks/callbacks.py b/infrastructure/autodesk-wala/packages/src/networks/callbacks.py
similarity index 100%
rename from autodesk-wala/packages/src/networks/callbacks.py
rename to infrastructure/autodesk-wala/packages/src/networks/callbacks.py
diff --git a/piper-tts/model/__init__.py b/infrastructure/autodesk-wala/packages/src/utils/__init__.py
similarity index 100%
rename from piper-tts/model/__init__.py
rename to infrastructure/autodesk-wala/packages/src/utils/__init__.py
diff --git a/autodesk-wala/packages/src/utils/visualization.py b/infrastructure/autodesk-wala/packages/src/utils/visualization.py
similarity index 100%
rename from autodesk-wala/packages/src/utils/visualization.py
rename to infrastructure/autodesk-wala/packages/src/utils/visualization.py
diff --git a/autodesk-wala/requirements.txt b/infrastructure/autodesk-wala/requirements.txt
similarity index 100%
rename from autodesk-wala/requirements.txt
rename to infrastructure/autodesk-wala/requirements.txt
diff --git a/autodesk-wala/test.py b/infrastructure/autodesk-wala/test.py
similarity index 100%
rename from autodesk-wala/test.py
rename to infrastructure/autodesk-wala/test.py
diff --git a/autodesk-wala/vendor_wala.sh b/infrastructure/autodesk-wala/vendor_wala.sh
similarity index 100%
rename from autodesk-wala/vendor_wala.sh
rename to infrastructure/autodesk-wala/vendor_wala.sh
diff --git a/infrastructure/binocular/README.md b/infrastructure/binocular/README.md
new file mode 100644
index 000000000..598d8d94d
--- /dev/null
+++ b/infrastructure/binocular/README.md
@@ -0,0 +1,30 @@
+# Binoculars
+
+Deploy [tiiuae/falcon-7b](https://huggingface.co/tiiuae/falcon-7b) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [tiiuae/falcon-7b](https://huggingface.co/tiiuae/falcon-7b) |
+| Task | Infrastructure / Custom server |
+| Engine | Custom (Truss) |
+| GPU | A10G:2 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"text": "Your text to check for AI-generated content goes here. The input should be at least 64 tokens long for reliable detection."}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/infrastructure/binocular/config.yaml b/infrastructure/binocular/config.yaml
new file mode 100644
index 000000000..2d57b3548
--- /dev/null
+++ b/infrastructure/binocular/config.yaml
@@ -0,0 +1,26 @@
+description: "Binoculars AI-generated text detector using Falcon 7B"
+environment_variables: {}
+external_package_dirs: []
+model_cache:
+- allow_patterns:
+  - '*.bin'
+  ignore_patterns:
+  - coreml/*
+  repo_id: tiiuae/falcon-7b
+  use_volume: false
+- allow_patterns:
+  - '*.bin'
+  ignore_patterns:
+  - coreml/*
+  repo_id: tiiuae/falcon-7b-instruct
+  use_volume: false
+model_metadata:
+  example_model_input: {"text": "The quick brown fox jumps over the lazy dog. This is a sample text to analyze for AI-generated content detection."}
+model_name: Binoculars
+python_version: py311
+requirements:
+- git+https://github.com/ahans30/Binoculars.git
+resources:
+  accelerator: A10G:2
+secrets: {}
+system_packages: []
diff --git a/playground-v2-aesthetic/model/__init__.py b/infrastructure/binocular/model/__init__.py
similarity index 100%
rename from playground-v2-aesthetic/model/__init__.py
rename to infrastructure/binocular/model/__init__.py
diff --git a/binocular/model/model.py b/infrastructure/binocular/model/model.py
similarity index 100%
rename from binocular/model/model.py
rename to infrastructure/binocular/model/model.py
diff --git a/binocular/packages/config.py b/infrastructure/binocular/packages/config.py
similarity index 100%
rename from binocular/packages/config.py
rename to infrastructure/binocular/packages/config.py
diff --git a/chains-examples/docs/audio-transcription/README.md b/infrastructure/chains-examples/docs/audio-transcription/README.md
similarity index 100%
rename from chains-examples/docs/audio-transcription/README.md
rename to infrastructure/chains-examples/docs/audio-transcription/README.md
diff --git a/chains-examples/docs/audio-transcription/data_types.py b/infrastructure/chains-examples/docs/audio-transcription/data_types.py
similarity index 100%
rename from chains-examples/docs/audio-transcription/data_types.py
rename to infrastructure/chains-examples/docs/audio-transcription/data_types.py
diff --git a/chains-examples/docs/audio-transcription/helpers.py b/infrastructure/chains-examples/docs/audio-transcription/helpers.py
similarity index 100%
rename from chains-examples/docs/audio-transcription/helpers.py
rename to infrastructure/chains-examples/docs/audio-transcription/helpers.py
diff --git a/chains-examples/docs/audio-transcription/transcribe.py b/infrastructure/chains-examples/docs/audio-transcription/transcribe.py
similarity index 100%
rename from chains-examples/docs/audio-transcription/transcribe.py
rename to infrastructure/chains-examples/docs/audio-transcription/transcribe.py
diff --git a/chains-examples/docs/audio-transcription/whisper_chainlet.py b/infrastructure/chains-examples/docs/audio-transcription/whisper_chainlet.py
similarity index 100%
rename from chains-examples/docs/audio-transcription/whisper_chainlet.py
rename to infrastructure/chains-examples/docs/audio-transcription/whisper_chainlet.py
diff --git a/chains-examples/docs/poems/poems.py b/infrastructure/chains-examples/docs/poems/poems.py
similarity index 100%
rename from chains-examples/docs/poems/poems.py
rename to infrastructure/chains-examples/docs/poems/poems.py
diff --git a/infrastructure/custom-engine-builder-control/README.md b/infrastructure/custom-engine-builder-control/README.md
new file mode 100644
index 000000000..e10dcddf7
--- /dev/null
+++ b/infrastructure/custom-engine-builder-control/README.md
@@ -0,0 +1,62 @@
+# Briton-suffix-fanout-qwen3-8B
+
+Deploy [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) |
+| Task | Infrastructure / Custom server |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "max_tokens": 100,
+  "messages": [
+    {
+      "content": "You are a helpful assistant. To each math question, e.g. <Whats 1+1>, respond with all parts the question, '=', and answer.",
+      "role": "system"
+    }
+  ],
+  "stream": false,
+  "suffix_messages": [
+    [
+      {
+        "role": "system",
+        "content": "Whats 1+1"
+      }
+    ],
+    [
+      {
+        "role": "system",
+        "content": "Whats 2+2"
+      }
+    ]
+  ],
+  "temperature": 0.5,
+  "chat_template_kwargs": {
+    "enable_thinking": false
+  }
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **32,768**
+- Plugin: **use_fp8_context_fmha**
diff --git a/infrastructure/custom-engine-builder-control/config.yaml b/infrastructure/custom-engine-builder-control/config.yaml
new file mode 100644
index 000000000..458946fec
--- /dev/null
+++ b/infrastructure/custom-engine-builder-control/config.yaml
@@ -0,0 +1,49 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Custom engine builder control example with Qwen3 8B"
+model_metadata:
+  example_model_input:
+    max_tokens: 100
+    messages:
+      - content: You are a helpful assistant. To each math question, e.g. <Whats 1+1>, respond with all parts the question, '=', and answer.
+        role: system
+    stream: false
+    suffix_messages: [
+        # k=1
+        [{ "role": "system", "content": "Whats 1+1" }],
+        # k=2
+        [{ "role": "system", "content": "Whats 2+2" }],
+      ]
+    temperature: 0.5
+    chat_template_kwargs: { "enable_thinking": false }
+  tags:
+    - openai-compatible
+model_name: Briton-suffix-fanout-qwen3-8B
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-8B
+      revision: main
+      source: HF
+    max_batch_size: 128
+    max_num_tokens: 32768
+    max_seq_len: 32768
+    num_builder_gpus: 1
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 8
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: false
diff --git a/custom-engine-builder-control/model/model.py b/infrastructure/custom-engine-builder-control/model/model.py
similarity index 100%
rename from custom-engine-builder-control/model/model.py
rename to infrastructure/custom-engine-builder-control/model/model.py
diff --git a/infrastructure/custom-server/README.md b/infrastructure/custom-server/README.md
new file mode 100644
index 000000000..1346207df
--- /dev/null
+++ b/infrastructure/custom-server/README.md
@@ -0,0 +1,37 @@
+# Custom Server
+
+Examples of deploying models using `docker_server.start_command` instead of a traditional `model.py`. This is useful when deploying models that are already wrapped in an API (e.g. vLLM, SGLang, LMDeploy) or when using a custom Docker image that handles HTTP requests directly.
+
+| Example | Engine | Description |
+|---------|--------|-------------|
+| [llama3-8b-instruct-sglang](llama3-8b-instruct-sglang/) | SGLang | Llama 3 8B Instruct via SGLang server |
+| [llama3-70b-instruct-sglang](llama3-70b-instruct-sglang/) | SGLang | Llama 3 70B Instruct via SGLang server |
+| [llama3-70b-instruct-lmdeploy](llama3-70b-instruct-lmdeploy/) | LMDeploy | Llama 3 70B Instruct via LMDeploy server |
+| [llama3-8b-instruct-lmdeploy](llama3-8b-instruct-lmdeploy/) | LMDeploy | Llama 3 8B Instruct via LMDeploy server |
+| [deepseek-v2-5-instruct-sglang](deepseek-v2-5-instruct-sglang/) | SGLang | DeepSeek v2.5 Instruct via SGLang server |
+| [pixtral-12b](pixtral-12b/) | vLLM | Pixtral 12B multimodal model |
+| [infinity-embedding-server](infinity-embedding-server/) | Infinity | Embedding server using Infinity engine |
+| [ultravox-0.4](ultravox-0.4/) | vLLM | Ultravox 0.4 multimodal audio model |
+| [ultravox-0.5-8b](ultravox-0.5-8b/) | vLLM | Ultravox 0.5 8B multimodal audio model |
+| [ultravox-0.6-70b](ultravox-0.6-70b/) | vLLM | Ultravox 0.6 70B multimodal audio model |
+| [voxtral-mini-3b-2507](voxtral-mini-3b-2507/) | vLLM | Voxtral Mini 3B speech model |
+| [voxtral-small-24b-2507](voxtral-small-24b-2507/) | vLLM | Voxtral Small 24B speech model |
+
+## Deploy
+
+```sh
+truss push infrastructure/custom-server/llama3-8b-instruct-sglang
+```
+
+## How it works
+
+Instead of writing a `model.py` with `load()` and `predict()` methods, these examples use `docker_server.start_command` in `config.yaml` to launch an existing inference server:
+
+```yaml
+docker_server:
+  start_command: "python -m sglang.launch_server --model meta-llama/Meta-Llama-3-8B-Instruct ..."
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+```
+
+This avoids unnecessary overhead when the model already provides its own HTTP endpoint.
diff --git a/infrastructure/custom-server/deepseek-v2-5-instruct-sglang/README.md b/infrastructure/custom-server/deepseek-v2-5-instruct-sglang/README.md
new file mode 100644
index 000000000..bdc97795e
--- /dev/null
+++ b/infrastructure/custom-server/deepseek-v2-5-instruct-sglang/README.md
@@ -0,0 +1,32 @@
+# DeepSeek V2.5 1210 SGLang
+
+Deploy [deepseek-ai/DeepSeek-V2.5-1210](https://huggingface.co/deepseek-ai/DeepSeek-V2.5-1210) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-V2.5-1210](https://huggingface.co/deepseek-ai/DeepSeek-V2.5-1210) |
+| Task | Infrastructure / Custom server |
+| Engine | SGLang |
+| GPU | H100:8 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-V2.5-1210", "prompt": "What is the capital of France?", "max_tokens": 256}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.0.post1-cu124`
+- Predict concurrency: **32**
diff --git a/infrastructure/custom-server/deepseek-v2-5-instruct-sglang/config.yaml b/infrastructure/custom-server/deepseek-v2-5-instruct-sglang/config.yaml
new file mode 100644
index 000000000..954417dea
--- /dev/null
+++ b/infrastructure/custom-server/deepseek-v2-5-instruct-sglang/config.yaml
@@ -0,0 +1,20 @@
+description: "DeepSeek V2.5 via custom SGLang server"
+base_image:
+  image: lmsysorg/sglang:v0.4.0.post1-cu124
+model_metadata:
+  example_model_input: {"model": "deepseek-ai/DeepSeek-V2.5-1210", "prompt": "What is machine learning?", "max_tokens": 512}
+  repo_id: deepseek-ai/DeepSeek-V2.5-1210
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path deepseek-ai/DeepSeek-V2.5-1210 --port 8000 --tp 8 --trust-remote-code"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/completions
+  server_port: 8000
+resources:
+  accelerator: H100:8
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: DeepSeek V2.5 1210 SGLang
+environment_variables:
+  hf_access_token: null
diff --git a/infrastructure/custom-server/infinity-embedding-server/README.md b/infrastructure/custom-server/infinity-embedding-server/README.md
new file mode 100644
index 000000000..9df5e84ea
--- /dev/null
+++ b/infrastructure/custom-server/infinity-embedding-server/README.md
@@ -0,0 +1,33 @@
+# infinity-embedding-server
+
+Deploy [BAAI/bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [BAAI/bge-small-en-v1.5](https://huggingface.co/BAAI/bge-small-en-v1.5) |
+| Task | Infrastructure / Custom server |
+| Engine | Docker Server |
+| GPU | L4 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/embeddings \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "BAAI/bge-small-en-v1.5", "input": ["What is the meaning of life?"]}'
+```
+
+## Configuration highlights
+
+- Base image: `python:3.11-slim`
+- Predict concurrency: **40**
+- Environment variables: `INFINITY_MAX_CLIENT_BATCH_SIZE`, `INFINITY_QUEUE_SIZE`, `DO_NOT_TRACK`
diff --git a/infrastructure/custom-server/infinity-embedding-server/config.yaml b/infrastructure/custom-server/infinity-embedding-server/config.yaml
new file mode 100644
index 000000000..38760ae19
--- /dev/null
+++ b/infrastructure/custom-server/infinity-embedding-server/config.yaml
@@ -0,0 +1,29 @@
+description: "BGE Small EN v1.5 via Infinity embedding server"
+base_image:
+  image: python:3.11-slim
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) infinity_emb v2 --batch-size 64 --model-id BAAI/bge-small-en-v1.5 --revision main"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /embeddings
+  server_port: 7997
+build_commands: # optional step to download the weights of the model into the image
+- sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) infinity_emb v2 --preload-only --no-model-warmup --model-id BAAI/bge-small-en-v1.5 --revision main"
+resources:
+  accelerator: L4
+  use_gpu: true
+model_metadata:
+  repo_id: "BAAI/bge-small-en-v1.5"
+  example_model_input: {"input": ["What is deep learning?"], "model": "BAAI/bge-small-en-v1.5"}
+model_name: infinity-embedding-server
+requirements:
+- infinity-emb[all]==0.0.72
+runtime:
+  predict_concurrency : 40
+environment_variables:
+  hf_access_token: null
+  # constrain api to at most 256 sentences per request, for better load-balancing
+  INFINITY_MAX_CLIENT_BATCH_SIZE: 256
+  # constrain model to a max backpressure of INFINITY_MAX_CLIENT_BATCH_SIZE * predict_concurrency = 10241 requests
+  INFINITY_QUEUE_SIZE: 10241
+  DO_NOT_TRACK: 1
diff --git a/infrastructure/custom-server/llama3-70b-instruct-lmdeploy/README.md b/infrastructure/custom-server/llama3-70b-instruct-lmdeploy/README.md
new file mode 100644
index 000000000..5aff4c03a
--- /dev/null
+++ b/infrastructure/custom-server/llama3-70b-instruct-lmdeploy/README.md
@@ -0,0 +1,32 @@
+# Llama 3.1 70B Instruct LMDeploy
+
+Deploy [meta-llama/Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct) |
+| Task | Infrastructure / Custom server |
+| Engine | Docker Server |
+| GPU | H100:4 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-70B-Instruct", "prompt": "What is the capital of France?", "max_tokens": 256}'
+```
+
+## Configuration highlights
+
+- Base image: `openmmlab/lmdeploy:v0.6.4-cu12`
+- Predict concurrency: **32**
diff --git a/infrastructure/custom-server/llama3-70b-instruct-lmdeploy/config.yaml b/infrastructure/custom-server/llama3-70b-instruct-lmdeploy/config.yaml
new file mode 100644
index 000000000..925ae964d
--- /dev/null
+++ b/infrastructure/custom-server/llama3-70b-instruct-lmdeploy/config.yaml
@@ -0,0 +1,20 @@
+description: "Llama 3.1 70B Instruct via custom LMDeploy server"
+base_image:
+  image: openmmlab/lmdeploy:v0.6.4-cu12
+model_metadata:
+  example_model_input: {"model": "meta-llama/Llama-3.1-70B-Instruct", "prompt": "What is machine learning?", "max_tokens": 512}
+  repo_id: meta-llama/Llama-3.1-70B-Instruct
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m lmdeploy serve api_server meta-llama/Llama-3.1-70B-Instruct --server-port 8000 --tp 4"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/completions
+  server_port: 8000
+resources:
+  accelerator: H100:4
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: Llama 3.1 70B Instruct LMDeploy
+environment_variables:
+  hf_access_token: null
diff --git a/infrastructure/custom-server/llama3-70b-instruct-sglang/README.md b/infrastructure/custom-server/llama3-70b-instruct-sglang/README.md
new file mode 100644
index 000000000..6d7b41a07
--- /dev/null
+++ b/infrastructure/custom-server/llama3-70b-instruct-sglang/README.md
@@ -0,0 +1,32 @@
+# Llama 3.1 70B Instruct SGLang
+
+Deploy [meta-llama/Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct) |
+| Task | Infrastructure / Custom server |
+| Engine | SGLang |
+| GPU | H100:4 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-70B-Instruct", "prompt": "What is the capital of France?", "max_tokens": 256}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.0.post1-cu124`
+- Predict concurrency: **32**
diff --git a/infrastructure/custom-server/llama3-70b-instruct-sglang/config.yaml b/infrastructure/custom-server/llama3-70b-instruct-sglang/config.yaml
new file mode 100644
index 000000000..f46bb7b96
--- /dev/null
+++ b/infrastructure/custom-server/llama3-70b-instruct-sglang/config.yaml
@@ -0,0 +1,20 @@
+description: "Llama 3.1 70B Instruct via custom SGLang server"
+base_image:
+  image: lmsysorg/sglang:v0.4.0.post1-cu124
+model_metadata:
+  example_model_input: {"model": "meta-llama/Llama-3.1-70B-Instruct", "prompt": "What is machine learning?", "max_tokens": 512}
+  repo_id: meta-llama/Llama-3.1-70B-Instruct
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path meta-llama/Llama-3.1-70B-Instruct --port 8000 --tp 4"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/completions
+  server_port: 8000
+resources:
+  accelerator: H100:4
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: Llama 3.1 70B Instruct SGLang
+environment_variables:
+  hf_access_token: null
diff --git a/infrastructure/custom-server/llama3-8b-instruct-lmdeploy/README.md b/infrastructure/custom-server/llama3-8b-instruct-lmdeploy/README.md
new file mode 100644
index 000000000..a40d89e4f
--- /dev/null
+++ b/infrastructure/custom-server/llama3-8b-instruct-lmdeploy/README.md
@@ -0,0 +1,32 @@
+# Llama 3.1 8B Instruct LMDeploy
+
+Deploy [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Task | Infrastructure / Custom server |
+| Engine | Docker Server |
+| GPU | H100 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-8B-Instruct", "prompt": "What is the capital of France?", "max_tokens": 256}'
+```
+
+## Configuration highlights
+
+- Base image: `openmmlab/lmdeploy:v0.6.4-cu12`
+- Predict concurrency: **32**
diff --git a/infrastructure/custom-server/llama3-8b-instruct-lmdeploy/config.yaml b/infrastructure/custom-server/llama3-8b-instruct-lmdeploy/config.yaml
new file mode 100644
index 000000000..057f88a1c
--- /dev/null
+++ b/infrastructure/custom-server/llama3-8b-instruct-lmdeploy/config.yaml
@@ -0,0 +1,20 @@
+description: "Llama 3.1 8B Instruct via custom LMDeploy server"
+base_image:
+  image: openmmlab/lmdeploy:v0.6.4-cu12
+model_metadata:
+  example_model_input: {"model": "meta-llama/Llama-3.1-8B-Instruct", "prompt": "What is machine learning?", "max_tokens": 512}
+  repo_id: meta-llama/Llama-3.1-8B-Instruct
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m lmdeploy serve api_server meta-llama/Llama-3.1-8B-Instruct --server-port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/completions
+  server_port: 8000
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: Llama 3.1 8B Instruct LMDeploy
+environment_variables:
+  hf_access_token: null
diff --git a/infrastructure/custom-server/llama3-8b-instruct-sglang/README.md b/infrastructure/custom-server/llama3-8b-instruct-sglang/README.md
new file mode 100644
index 000000000..9f9511e7a
--- /dev/null
+++ b/infrastructure/custom-server/llama3-8b-instruct-sglang/README.md
@@ -0,0 +1,32 @@
+# Llama 3.1 8B Instruct SGLang
+
+Deploy [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Task | Infrastructure / Custom server |
+| Engine | SGLang |
+| GPU | H100 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-8B-Instruct", "prompt": "What is the capital of France?", "max_tokens": 256}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.0.post1-cu124`
+- Predict concurrency: **32**
diff --git a/infrastructure/custom-server/llama3-8b-instruct-sglang/config.yaml b/infrastructure/custom-server/llama3-8b-instruct-sglang/config.yaml
new file mode 100644
index 000000000..6926c2dc4
--- /dev/null
+++ b/infrastructure/custom-server/llama3-8b-instruct-sglang/config.yaml
@@ -0,0 +1,20 @@
+description: "Llama 3.1 8B Instruct via custom SGLang server"
+base_image:
+  image: lmsysorg/sglang:v0.4.0.post1-cu124
+model_metadata:
+  example_model_input: {"model": "meta-llama/Llama-3.1-8B-Instruct", "prompt": "What is machine learning?", "max_tokens": 512}
+  repo_id: meta-llama/Llama-3.1-8B-Instruct
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path meta-llama/Llama-3.1-8B-Instruct --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/completions
+  server_port: 8000
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: Llama 3.1 8B Instruct SGLang
+environment_variables:
+  hf_access_token: null
diff --git a/custom-server/llama3_eval.py b/infrastructure/custom-server/llama3_eval.py
similarity index 100%
rename from custom-server/llama3_eval.py
rename to infrastructure/custom-server/llama3_eval.py
diff --git a/infrastructure/custom-server/pixtral-12b/README.md b/infrastructure/custom-server/pixtral-12b/README.md
new file mode 100644
index 000000000..ee26c5a69
--- /dev/null
+++ b/infrastructure/custom-server/pixtral-12b/README.md
@@ -0,0 +1,54 @@
+# Pixtral 12B
+
+Deploy [mistralai/Pixtral-12B-2409](https://huggingface.co/mistralai/Pixtral-12B-2409) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Pixtral-12B-2409](https://huggingface.co/mistralai/Pixtral-12B-2409) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "pixtral",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "Describe this image in one sentence."
+        },
+        {
+          "type": "image_url",
+          "image_url": {
+            "url": "https://picsum.photos/id/237/200/300"
+          }
+        }
+      ]
+    }
+  ],
+  "stream": false,
+  "max_tokens": 512,
+  "temperature": 0.5
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.7.3`
+- Predict concurrency: **16**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/infrastructure/custom-server/pixtral-12b/config.yaml b/infrastructure/custom-server/pixtral-12b/config.yaml
new file mode 100644
index 000000000..4a76d2501
--- /dev/null
+++ b/infrastructure/custom-server/pixtral-12b/config.yaml
@@ -0,0 +1,47 @@
+description: "Pixtral 12B via custom vLLM server"
+base_image:
+  image: vllm/vllm-openai:v0.7.3
+model_metadata:
+  repo_id: mistralai/Pixtral-12B-2409
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  example_model_input: {
+    "model": "pixtral",
+    "messages": [
+      {
+      "role": "user",
+      "content": [
+        {
+        "type": "text",
+        "text": "Describe this image in one sentence."
+        },
+        {
+        "type": "image_url",
+        "image_url": {
+          "url": "https://picsum.photos/id/237/200/300"
+        }
+        }
+      ]
+      }
+    ],
+    "stream": false,
+    "max_tokens": 512,
+    "temperature": 0.5
+  }
+  tags:
+    - openai-compatible
+    - multimodal
+    - text-generation
+docker_server:
+  start_command: sh -c "vllm serve mistral-community/pixtral-12b --served-model-name pixtral --max-model-len 65536 --chat-template /app/data/pixtral12b.jinja --chat-template-content-format string --limit_mm_per_prompt 'image=4' --gpu-memory-utilization 0.95"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+runtime:
+  predict_concurrency : 16
+resources:
+  accelerator: H100
+  use_gpu: true
+model_name: Pixtral 12B
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
diff --git a/custom-server/pixtral-12b/data/pixtral12b.jinja b/infrastructure/custom-server/pixtral-12b/data/pixtral12b.jinja
similarity index 100%
rename from custom-server/pixtral-12b/data/pixtral12b.jinja
rename to infrastructure/custom-server/pixtral-12b/data/pixtral12b.jinja
diff --git a/infrastructure/custom-server/ultravox-0.4/README.md b/infrastructure/custom-server/ultravox-0.4/README.md
new file mode 100644
index 000000000..c6948d86f
--- /dev/null
+++ b/infrastructure/custom-server/ultravox-0.4/README.md
@@ -0,0 +1,50 @@
+# Ultravox v0.4
+
+Take in audio and text as input, generating text as usual
+
+| Property | Value |
+|----------|-------|
+| Model | [fixie-ai/ultravox-v0.4](https://huggingface.co/fixie-ai/ultravox-v0.4) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100_40GB |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "ultravox",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What is Lydia like?"
+        },
+        {
+          "type": "audio_url",
+          "audio_url": {
+            "url": "https://baseten-public.s3.us-west-2.amazonaws.com/fred-audio-tests/real.mp3"
+          }
+        }
+      ]
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.9.2`
+- Predict concurrency: **16**
diff --git a/custom-server/ultravox-0.4/config.yaml b/infrastructure/custom-server/ultravox-0.4/config.yaml
similarity index 100%
rename from custom-server/ultravox-0.4/config.yaml
rename to infrastructure/custom-server/ultravox-0.4/config.yaml
diff --git a/infrastructure/custom-server/ultravox-0.5-8b/README.md b/infrastructure/custom-server/ultravox-0.5-8b/README.md
new file mode 100644
index 000000000..e14f042fd
--- /dev/null
+++ b/infrastructure/custom-server/ultravox-0.5-8b/README.md
@@ -0,0 +1,52 @@
+# Ultravox v0.5 8B
+
+Take in audio and text as input, generating text as usual
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "ultravox",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What is Lydia like?"
+        },
+        {
+          "type": "audio_url",
+          "audio_url": {
+            "url": "https://baseten-public.s3.us-west-2.amazonaws.com/fred-audio-tests/real.mp3"
+          }
+        }
+      ]
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.9.2`
+- Predict concurrency: **16**
diff --git a/custom-server/ultravox-0.5-8b/config.yaml b/infrastructure/custom-server/ultravox-0.5-8b/config.yaml
similarity index 100%
rename from custom-server/ultravox-0.5-8b/config.yaml
rename to infrastructure/custom-server/ultravox-0.5-8b/config.yaml
diff --git a/infrastructure/custom-server/ultravox-0.6-70b/README.md b/infrastructure/custom-server/ultravox-0.6-70b/README.md
new file mode 100644
index 000000000..4a0bbba14
--- /dev/null
+++ b/infrastructure/custom-server/ultravox-0.6-70b/README.md
@@ -0,0 +1,52 @@
+# Ultravox v0.6 70B
+
+Take in audio and text as input, generating text as usual
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100:4 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "ultravox",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What is Lydia like?"
+        },
+        {
+          "type": "audio_url",
+          "audio_url": {
+            "url": "https://baseten-public.s3.us-west-2.amazonaws.com/fred-audio-tests/real.mp3"
+          }
+        }
+      ]
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.9.2`
+- Predict concurrency: **16**
diff --git a/custom-server/ultravox-0.6-70b/config.yaml b/infrastructure/custom-server/ultravox-0.6-70b/config.yaml
similarity index 100%
rename from custom-server/ultravox-0.6-70b/config.yaml
rename to infrastructure/custom-server/ultravox-0.6-70b/config.yaml
diff --git a/infrastructure/custom-server/voxtral-mini-3b-2507/README.md b/infrastructure/custom-server/voxtral-mini-3b-2507/README.md
new file mode 100644
index 000000000..9da13aa35
--- /dev/null
+++ b/infrastructure/custom-server/voxtral-mini-3b-2507/README.md
@@ -0,0 +1,46 @@
+# Voxtral Mini 3B 2507
+
+Take in audio and text as input, generating text as usual
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Voxtral-Mini-3B-2507](https://huggingface.co/mistralai/Voxtral-Mini-3B-2507) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100_40GB |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "voxtral-mini",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What is the name of the famous bicycle race in France?"
+        }
+      ]
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.10.0`
+- Predict concurrency: **16**
diff --git a/infrastructure/custom-server/voxtral-mini-3b-2507/config.yaml b/infrastructure/custom-server/voxtral-mini-3b-2507/config.yaml
new file mode 100644
index 000000000..2fc4d0a7c
--- /dev/null
+++ b/infrastructure/custom-server/voxtral-mini-3b-2507/config.yaml
@@ -0,0 +1,41 @@
+description: Take in audio and text as input, generating text as usual
+base_image:
+  image: vllm/vllm-openai:v0.10.0
+model_metadata:
+  repo_id: mistralai/Voxtral-Mini-3B-2507
+  avatar_url: https://cdn-avatars.huggingface.co/v1/production/uploads/634c17653d11eaedd88b314d/9OgyfKstSZtbmsmuG8MbU.png
+  example_model_input:
+    {
+      "model": "voxtral-mini",
+      "messages":
+        [
+          {
+            "role": "user",
+            "content":
+              [
+                {
+                  "type": "text",
+                  "text": "What is the name of the famous bicycle race in France?",
+                },
+              ],
+          },
+        ],
+    }
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Voxtral-Mini-3B-2507 --tokenizer_mode mistral --config_format mistral --load_format mistral --port 8000 --served-model-name voxtral-mini"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+runtime:
+  predict_concurrency: 16
+model_name: Voxtral Mini 3B 2507
+secrets:
+  hf_access_token: null
+requirements:
+  - vllm[audio]==0.10.0
diff --git a/infrastructure/custom-server/voxtral-small-24b-2507/README.md b/infrastructure/custom-server/voxtral-small-24b-2507/README.md
new file mode 100644
index 000000000..27cf0c677
--- /dev/null
+++ b/infrastructure/custom-server/voxtral-small-24b-2507/README.md
@@ -0,0 +1,46 @@
+# Voxtral Small 24B 2507
+
+Take in audio and text as input, generating text as usual
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Voxtral-Small-24B-2507](https://huggingface.co/mistralai/Voxtral-Small-24B-2507) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "voxtral-small",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What is the name of the famous bicycle race in France?"
+        }
+      ]
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.10.0`
+- Predict concurrency: **16**
diff --git a/infrastructure/custom-server/voxtral-small-24b-2507/config.yaml b/infrastructure/custom-server/voxtral-small-24b-2507/config.yaml
new file mode 100644
index 000000000..521681e74
--- /dev/null
+++ b/infrastructure/custom-server/voxtral-small-24b-2507/config.yaml
@@ -0,0 +1,41 @@
+description: Take in audio and text as input, generating text as usual
+base_image:
+  image: vllm/vllm-openai:v0.10.0
+model_metadata:
+  repo_id: mistralai/Voxtral-Small-24B-2507
+  avatar_url: https://cdn-avatars.huggingface.co/v1/production/uploads/634c17653d11eaedd88b314d/9OgyfKstSZtbmsmuG8MbU.png
+  example_model_input:
+    {
+      "model": "voxtral-small",
+      "messages":
+        [
+          {
+            "role": "user",
+            "content":
+              [
+                {
+                  "type": "text",
+                  "text": "What is the name of the famous bicycle race in France?",
+                },
+              ],
+          },
+        ],
+    }
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Voxtral-Small-24B-2507 --tokenizer_mode mistral --config_format mistral --load_format mistral --port 8000 --served-model-name voxtral-small"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency: 16
+model_name: Voxtral Small 24B 2507
+secrets:
+  hf_access_token: null
+requirements:
+  - vllm[audio]==0.10.0
diff --git a/grpc/Dockerfile b/infrastructure/grpc/Dockerfile
similarity index 100%
rename from grpc/Dockerfile
rename to infrastructure/grpc/Dockerfile
diff --git a/infrastructure/grpc/README.md b/infrastructure/grpc/README.md
new file mode 100644
index 000000000..e46928ade
--- /dev/null
+++ b/infrastructure/grpc/README.md
@@ -0,0 +1,49 @@
+# gRPC Model Example
+
+Deploy a model using gRPC transport on Baseten. This example demonstrates how to serve a gRPC service using a custom Docker server with Truss.
+
+| Property | Value |
+|----------|-------|
+| Task | Infrastructure / gRPC transport |
+| Engine | Docker Server |
+| GPU | A10G |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model uses **gRPC transport**, not HTTP/REST. Use the included `client.py` as a reference:
+
+```python
+import grpc
+import example_pb2
+import example_pb2_grpc
+
+channel = grpc.insecure_channel(
+    "model-<MODEL_ID>.grpc.api.baseten.co:80",
+)
+
+stub = example_pb2_grpc.GreeterStub(channel)
+
+request = example_pb2.HelloRequest(name="World")
+
+metadata = [
+    ("baseten-authorization", "Api-Key YOUR_BASETEN_API_KEY"),
+    ("baseten-model-id", "<MODEL_ID>"),
+]
+
+response = stub.SayHello(request, metadata=metadata)
+print(response.message)
+```
+
+The proto definition (`example.proto`) defines a simple `Greeter` service with a `SayHello` RPC method.
+
+## Configuration highlights
+
+- Transport: **gRPC** (not HTTP/REST -- curl will not work)
+- Base image: `your/repository:tag`
+- Docker server with custom `model.py` entrypoint
diff --git a/grpc/client.py b/infrastructure/grpc/client.py
similarity index 100%
rename from grpc/client.py
rename to infrastructure/grpc/client.py
diff --git a/infrastructure/grpc/config.yaml b/infrastructure/grpc/config.yaml
new file mode 100644
index 000000000..101368971
--- /dev/null
+++ b/infrastructure/grpc/config.yaml
@@ -0,0 +1,18 @@
+description: "Example model using gRPC transport"
+model_metadata:
+  example_model_input: {"name": "World"}
+model_name: "gRPC Model Example"
+base_image:
+  image: your/repository:tag
+docker_server:
+  start_command: python model.py
+  server_port: 8080
+  predict_endpoint: /
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+resources:
+  accelerator: A10G  # or your preferred GPU
+  use_gpu: true
+runtime:
+  transport:
+    kind: "grpc"
diff --git a/grpc/example.proto b/infrastructure/grpc/example.proto
similarity index 100%
rename from grpc/example.proto
rename to infrastructure/grpc/example.proto
diff --git a/grpc/example_pb2.py b/infrastructure/grpc/example_pb2.py
similarity index 100%
rename from grpc/example_pb2.py
rename to infrastructure/grpc/example_pb2.py
diff --git a/grpc/example_pb2_grpc.py b/infrastructure/grpc/example_pb2_grpc.py
similarity index 100%
rename from grpc/example_pb2_grpc.py
rename to infrastructure/grpc/example_pb2_grpc.py
diff --git a/grpc/model.py b/infrastructure/grpc/model.py
similarity index 100%
rename from grpc/model.py
rename to infrastructure/grpc/model.py
diff --git a/grpc/requirements.txt b/infrastructure/grpc/requirements.txt
similarity index 100%
rename from grpc/requirements.txt
rename to infrastructure/grpc/requirements.txt
diff --git a/infrastructure/jsonformatter/README.md b/infrastructure/jsonformatter/README.md
new file mode 100644
index 000000000..09d8b51b9
--- /dev/null
+++ b/infrastructure/jsonformatter/README.md
@@ -0,0 +1,29 @@
+# JsonFormatter
+
+Deploy JsonFormatter using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Infrastructure / Custom server |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "Generate a person'\''s name and age"}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/infrastructure/jsonformatter/config.yaml b/infrastructure/jsonformatter/config.yaml
new file mode 100644
index 000000000..63216e83f
--- /dev/null
+++ b/infrastructure/jsonformatter/config.yaml
@@ -0,0 +1,17 @@
+description: "JSON formatting utility model example"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "databricks/dolly-v2-3b"
+  example_model_input: {"prompt": "Generate a person's information"}
+  llm_model: databricks/dolly-v2-3b
+model_name: JsonFormatter
+python_version: py311
+requirements:
+- jsonformer==0.12.0
+- transformers==4.36.0
+- accelerate==0.25.0
+resources:
+  accelerator: A10G
+secrets: {}
+system_packages: []
diff --git a/jsonformatter/data/schema.json b/infrastructure/jsonformatter/data/schema.json
similarity index 100%
rename from jsonformatter/data/schema.json
rename to infrastructure/jsonformatter/data/schema.json
diff --git a/qwen/qwen-7b-chat/model/__init__.py b/infrastructure/jsonformatter/model/__init__.py
similarity index 100%
rename from qwen/qwen-7b-chat/model/__init__.py
rename to infrastructure/jsonformatter/model/__init__.py
diff --git a/jsonformatter/model/model.py b/infrastructure/jsonformatter/model/model.py
similarity index 100%
rename from jsonformatter/model/model.py
rename to infrastructure/jsonformatter/model/model.py
diff --git a/infrastructure/layoutlm-document-qa/README.md b/infrastructure/layoutlm-document-qa/README.md
new file mode 100644
index 000000000..92e913164
--- /dev/null
+++ b/infrastructure/layoutlm-document-qa/README.md
@@ -0,0 +1,32 @@
+# LayoutLM Document QA
+
+Extract information from images of invoices
+
+| Property | Value |
+|----------|-------|
+| Task | Infrastructure / Custom server |
+| Engine | Custom (Truss) |
+| GPU | CPU |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "What is the invoice number?",
+  "url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `tesseract-ocr`
diff --git a/infrastructure/layoutlm-document-qa/config.yaml b/infrastructure/layoutlm-document-qa/config.yaml
new file mode 100644
index 000000000..c058010ce
--- /dev/null
+++ b/infrastructure/layoutlm-document-qa/config.yaml
@@ -0,0 +1,27 @@
+description: Extract information from images of invoices
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "impira/layoutlm-document-qa"
+  avatar_url: https://cdn.baseten.co/production/static/explore/impira-logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/document-qa.png
+  example_model_input:
+    prompt: What is the invoice number?
+    url: https://templates.invoicehome.com/invoice-template-us-neat-750px.png
+  tags:
+  - text-generation
+model_name: LayoutLM Document QA
+python_version: py39
+requirements:
+- Pillow==10.0.0
+- pytesseract==0.3.10
+- torch==2.0.1
+- transformers==4.30.2
+resources:
+  accelerator: null
+  cpu: '4'
+  memory: 16Gi
+  use_gpu: false
+secrets: {}
+system_packages:
+- tesseract-ocr
diff --git a/qwen/qwen-vl/model/__init__.py b/infrastructure/layoutlm-document-qa/model/__init__.py
similarity index 100%
rename from qwen/qwen-vl/model/__init__.py
rename to infrastructure/layoutlm-document-qa/model/__init__.py
diff --git a/layoutlm-document-qa/model/model.py b/infrastructure/layoutlm-document-qa/model/model.py
similarity index 100%
rename from layoutlm-document-qa/model/model.py
rename to infrastructure/layoutlm-document-qa/model/model.py
diff --git a/infrastructure/llama-cpp-server/README.md b/infrastructure/llama-cpp-server/README.md
new file mode 100644
index 000000000..990c81ca8
--- /dev/null
+++ b/infrastructure/llama-cpp-server/README.md
@@ -0,0 +1,56 @@
+# llama cpp gemma 3 27b it qat q4_0
+
+Deploy [google/gemma-3-27b-it-qat-q4_0-gguf](https://huggingface.co/google/gemma-3-27b-it-qat-q4_0-gguf) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [google/gemma-3-27b-it-qat-q4_0-gguf](https://huggingface.co/google/gemma-3-27b-it-qat-q4_0-gguf) |
+| Task | Infrastructure / Custom server |
+| Engine | Docker Server |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "gemma",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "Describe this image in one sentence."
+        },
+        {
+          "type": "image_url",
+          "image_url": {
+            "url": "https://picsum.photos/id/237/200/300"
+          }
+        }
+      ]
+    }
+  ],
+  "stream": true,
+  "max_tokens": 512,
+  "temperature": 0.5
+}'
+```
+
+## Configuration highlights
+
+- Base image: `alphatozeta/llama-cpp-server:0.4`
+- Predict concurrency: **8**
+- Streaming: **enabled**
diff --git a/infrastructure/llama-cpp-server/config.yaml b/infrastructure/llama-cpp-server/config.yaml
new file mode 100644
index 000000000..9f10e096f
--- /dev/null
+++ b/infrastructure/llama-cpp-server/config.yaml
@@ -0,0 +1,47 @@
+description: "Gemma 3 27B via llama.cpp server"
+base_image:
+  image: alphatozeta/llama-cpp-server:0.4
+build_commands:
+  - pip install git+https://github.com/huggingface/transformers.git hf-xet
+model_metadata:
+  repo_id: google/gemma-3-27b-it-qat-q4_0-gguf
+  example_model_input: {
+    "model": "gemma",
+    "messages": [
+      {
+      "role": "user",
+      "content": [
+        {
+        "type": "text",
+        "text": "Describe this image in one sentence."
+        },
+        {
+        "type": "image_url",
+        "image_url": {
+          "url": "https://picsum.photos/id/237/200/300"
+        }
+        }
+      ]
+      }
+    ],
+    "stream": true,
+    "max_tokens": 512,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) huggingface-cli download google/gemma-3-27b-it-qat-q4_0-gguf --local-dir /app/gemma-3-27b-it-qat-q4_0-gguf && HF_TOKEN=$(cat /secrets/hf_access_token) huggingface-cli download google/gemma-3-1b-it-qat-q4_0-gguf --local-dir /app/gemma-3-1b-it-qat-q4_0-gguf && /app/llama-server -m /app/gemma-3-27b-it-qat-q4_0-gguf/gemma-3-27b-it-q4_0.gguf -md /app/gemma-3-1b-it-qat-q4_0-gguf/gemma-3-1b-it-q4_0.gguf --port 8000 -c 32768 -cd 32768 -ngl 999 -ngld 999 --draft-max 16 --draft-min 0 --prio 3 -fa --no-webui"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+requirements: []
+secrets:
+  hf_access_token: null
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency : 8
+model_name: llama cpp gemma 3 27b it qat q4_0
diff --git a/llama-cpp-server/cuda.Dockerfile b/infrastructure/llama-cpp-server/cuda.Dockerfile
similarity index 100%
rename from llama-cpp-server/cuda.Dockerfile
rename to infrastructure/llama-cpp-server/cuda.Dockerfile
diff --git a/llama-cpp-server/llama_server_help b/infrastructure/llama-cpp-server/llama_server_help
similarity index 100%
rename from llama-cpp-server/llama_server_help
rename to infrastructure/llama-cpp-server/llama_server_help
diff --git a/metrics/datadog/Dockerfile b/infrastructure/metrics/datadog/Dockerfile
similarity index 100%
rename from metrics/datadog/Dockerfile
rename to infrastructure/metrics/datadog/Dockerfile
diff --git a/infrastructure/metrics/datadog/README.md b/infrastructure/metrics/datadog/README.md
new file mode 100644
index 000000000..65d27137e
--- /dev/null
+++ b/infrastructure/metrics/datadog/README.md
@@ -0,0 +1,49 @@
+# truss_fastapi_datadog
+
+Deploy [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "messages": [
+    {
+      "role": "system",
+      "content": "You are a helpful assistant."
+    },
+    {
+      "role": "user",
+      "content": "What does Tongyi Qianwen mean?"
+    }
+  ],
+  "stream": false,
+  "model": "qwen30b",
+  "max_tokens": 512,
+  "temperature": 0.7
+}'
+```
+
+## Configuration highlights
+
+- Base image: `chriswirick/truss_fastapi_datadog_vllm:v0.11.0h`
+- Predict concurrency: **32**
+- Environment variables: `DD_SITE`, `DD_HOSTNAME`, `DD_SERVICE`, `DD_ENV`, `DD_RUN_PATH`, `DD_AUTH_TOKEN_FILE_PATH`, `DD_INVENTORIES_CHECKS_ENABLED`, `DD_OTLP_CONFIG_RECEIVER_PROTOCOLS_GRPC_ENDPOINT`, `DD_CLOUD_PROVIDER_METADATA`, `VLLM_LOGGING_LEVEL`
diff --git a/infrastructure/metrics/datadog/config.yaml b/infrastructure/metrics/datadog/config.yaml
new file mode 100644
index 000000000..5b449b343
--- /dev/null
+++ b/infrastructure/metrics/datadog/config.yaml
@@ -0,0 +1,43 @@
+description: "Datadog metrics integration example with Qwen3 30B"
+base_image:
+  image: chriswirick/truss_fastapi_datadog_vllm:v0.11.0h
+docker_server:
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  readiness_endpoint: /health
+  server_port: 8000
+  start_command: sh -c "export DD_API_KEY=$(cat /secrets/dd_api_key | tr -d '\n\r' | xargs) && mkdir -p /tmp/datadog-agent /var/log/datadog && /opt/datadog-agent/bin/agent/agent run 2>&1 & sleep 3 && HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve Qwen/Qwen3-30B-A3B --reasoning-parser deepseek_r1 --served-model-name qwen30b --port 8000"
+environment_variables:
+  DD_SITE: "us5.datadoghq.com"
+  DD_HOSTNAME: "truss-vllm-server"
+  DD_SERVICE: "truss-vllm"
+  DD_ENV: "production"
+  DD_RUN_PATH: "/tmp/datadog-agent"
+  DD_AUTH_TOKEN_FILE_PATH: "/tmp/datadog-agent/auth_token"
+  DD_INVENTORIES_CHECKS_ENABLED: "false"
+  DD_OTLP_CONFIG_RECEIVER_PROTOCOLS_GRPC_ENDPOINT: ""
+  DD_CLOUD_PROVIDER_METADATA: "[]"
+  VLLM_LOGGING_LEVEL: WARNING
+model_metadata:
+  repo_id: Qwen/Qwen3-30B-A3B
+  example_model_input:
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What does Tongyi Qianwen mean?"
+    stream: false
+    model: "qwen30b"
+    max_tokens: 512
+    temperature: 0.7
+  tags:
+    - openai-compatible
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
+model_name: truss_fastapi_datadog
+secrets:
+  dd_api_key: null
+  hf_access_token: null
diff --git a/metrics/datadog/datadog-vllm-guide.md b/infrastructure/metrics/datadog/datadog-vllm-guide.md
similarity index 100%
rename from metrics/datadog/datadog-vllm-guide.md
rename to infrastructure/metrics/datadog/datadog-vllm-guide.md
diff --git a/metrics/datadog/vllm_conf.yaml b/infrastructure/metrics/datadog/vllm_conf.yaml
similarity index 100%
rename from metrics/datadog/vllm_conf.yaml
rename to infrastructure/metrics/datadog/vllm_conf.yaml
diff --git a/infrastructure/model-cache/README.md b/infrastructure/model-cache/README.md
new file mode 100644
index 000000000..79fb3d7c8
--- /dev/null
+++ b/infrastructure/model-cache/README.md
@@ -0,0 +1,36 @@
+# Model Cache
+
+Demonstrates how to use Truss `model_cache` to cache model weights across deployments for faster cold starts. This example caches Stable Diffusion XL weights using volume-mounted HuggingFace repos.
+
+| Property | Value |
+|----------|-------|
+| Task | Infrastructure / Model caching |
+| Engine | Custom (Truss) |
+| GPU | CPU |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"text": "hello"}'
+```
+
+The model returns the input data along with the size of the cached model weights file.
+
+## Configuration highlights
+
+- Model cache: **volume-mounted** for fast cold starts
+- Two cached HuggingFace repos: `madebyollin/sdxl-vae-fp16-fix` and `stabilityai/stable-diffusion-xl-base-1.0`
+- Uses `revision` pinning and `allow_patterns` for selective downloads
+- `runtime_secret_name` for authenticated HF access at runtime
diff --git a/infrastructure/model-cache/config.yaml b/infrastructure/model-cache/config.yaml
new file mode 100644
index 000000000..e8af0f419
--- /dev/null
+++ b/infrastructure/model-cache/config.yaml
@@ -0,0 +1,32 @@
+description: "Example of model weight caching for faster cold starts"
+model_metadata:
+  example_model_input: {"prompt": "Hello, world!"}
+model_name: Hello Model Cache Qwen
+python_version: py311
+requirements: ["torch==2.5.1"]
+resources:
+  accelerator: null
+  cpu: "1"
+  memory: 8Gi
+  use_gpu: false
+secrets: { hf_access_token: null } # null is encouraged, as this will automatically use the one provided by baseten.co
+model_cache:
+  - repo_id: madebyollin/sdxl-vae-fp16-fix
+    revision: 207b116dae70ace3637169f1ddd2434b91b3a8cd
+    use_volume: true
+    volume_folder: sdxl-vae-fp16
+    allow_patterns:
+      - config.json
+      - diffusion_pytorch_model.safetensors
+    runtime_secret_name: hf_access_token
+    kind: "hf"
+  - repo_id: stabilityai/stable-diffusion-xl-base-1.0
+    revision: 462165984030d82259a11f4367a4eed129e94a7b
+    use_volume: true
+    volume_folder: stable-diffusion-xl-base
+    allow_patterns:
+      - "*.json"
+      - "*.fp16.safetensors"
+      - sd_xl_base_1.0.safetensors
+    runtime_secret_name: hf_access_token
+    kind: "hf"
diff --git a/model_cache/model/model.py b/infrastructure/model-cache/model/model.py
similarity index 100%
rename from model_cache/model/model.py
rename to infrastructure/model-cache/model/model.py
diff --git a/infrastructure/multiprocessing/README.md b/infrastructure/multiprocessing/README.md
new file mode 100644
index 000000000..03972f245
--- /dev/null
+++ b/infrastructure/multiprocessing/README.md
@@ -0,0 +1,29 @@
+# Model with multiprocessing pre/post-process
+
+Deploy Model with multiprocessing pre/post-process using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Infrastructure / Custom server |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"n": 100}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/infrastructure/multiprocessing/config.yaml b/infrastructure/multiprocessing/config.yaml
new file mode 100644
index 000000000..1d9d12bcb
--- /dev/null
+++ b/infrastructure/multiprocessing/config.yaml
@@ -0,0 +1,16 @@
+description: "Example of multiprocessing for pre/post-processing"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  example_model_input: {"n": 100}
+model_name: Model with multiprocessing pre/post-process
+python_version: py310
+requirements:
+- torch==2.1.0
+resources:
+  accelerator: A10G
+  cpu: '8'
+  memory: 8Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/sana/sana_1600M/model/__init__.py b/infrastructure/multiprocessing/model/__init__.py
similarity index 100%
rename from sana/sana_1600M/model/__init__.py
rename to infrastructure/multiprocessing/model/__init__.py
diff --git a/multiprocessing/model/model.py b/infrastructure/multiprocessing/model/model.py
similarity index 100%
rename from multiprocessing/model/model.py
rename to infrastructure/multiprocessing/model/model.py
diff --git a/multiprocessing/model/test.py b/infrastructure/multiprocessing/model/test.py
similarity index 100%
rename from multiprocessing/model/test.py
rename to infrastructure/multiprocessing/model/test.py
diff --git a/infrastructure/ngram-speculator/truss/README.md b/infrastructure/ngram-speculator/truss/README.md
new file mode 100644
index 000000000..07e174c11
--- /dev/null
+++ b/infrastructure/ngram-speculator/truss/README.md
@@ -0,0 +1,29 @@
+# ngram-speculator
+
+Deploy ngram-speculator using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Infrastructure / Custom server |
+| Engine | Custom (Truss) |
+| GPU | H100 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"messages": [{"role": "user", "content": "What is the capital of France?"}], "max_tokens": 256}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/infrastructure/ngram-speculator/truss/config.yaml b/infrastructure/ngram-speculator/truss/config.yaml
new file mode 100644
index 000000000..bcb3bf8cb
--- /dev/null
+++ b/infrastructure/ngram-speculator/truss/config.yaml
@@ -0,0 +1,17 @@
+description: "N-gram speculative decoding example via Truss"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "NousResearch/Meta-Llama-3.1-8B-Instruct"
+  example_model_input: {"messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}
+model_name: ngram-speculator
+python_version: py310
+requirements:
+- vllm==0.6.5
+- transformers==4.47.1
+resources:
+  accelerator: H100
+  use_gpu: True
+secrets: {}
+system_packages: []
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/__init__.py b/infrastructure/ngram-speculator/truss/model/__init__.py
old mode 100755
new mode 100644
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/__init__.py
rename to infrastructure/ngram-speculator/truss/model/__init__.py
diff --git a/ngram-speculator/truss/model/model.py b/infrastructure/ngram-speculator/truss/model/model.py
similarity index 100%
rename from ngram-speculator/truss/model/model.py
rename to infrastructure/ngram-speculator/truss/model/model.py
diff --git a/infrastructure/ngram-speculator/trussless/README.md b/infrastructure/ngram-speculator/trussless/README.md
new file mode 100644
index 000000000..1b48220b8
--- /dev/null
+++ b/infrastructure/ngram-speculator/trussless/README.md
@@ -0,0 +1,47 @@
+# ngram-speculator
+
+Deploy [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "llama",
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "What do llamas dream of?"
+        }
+      ]
+    }
+  ],
+  "stream": false,
+  "max_tokens": 512
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:nightly`
+- Predict concurrency: **16**
diff --git a/infrastructure/ngram-speculator/trussless/config.yaml b/infrastructure/ngram-speculator/trussless/config.yaml
new file mode 100644
index 000000000..693ab4386
--- /dev/null
+++ b/infrastructure/ngram-speculator/trussless/config.yaml
@@ -0,0 +1,35 @@
+description: "N-gram speculative decoding example via custom server"
+base_image:
+  image: vllm/vllm-openai:nightly
+model_metadata:
+  repo_id: Qwen/Qwen3-0.6B
+  example_model_input: {
+    "model": "llama",
+    "messages": [
+      {
+      "role": "user",
+      "content": [
+        {
+        "type": "text",
+        "text": "What do llamas dream of?"
+        }
+      ]
+      }
+    ],
+    "stream": false,
+    "max_tokens": 512,
+  }
+docker_server:
+  start_command: sh -c "vllm serve distilbert/distilgpt2 --served-model-name llama --tensor-parallel-size 1 --max-model-len 64 --max-num-seqs 1 --max-num-batched-tokens 64 --gpu-memory-utilization 0.7"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+runtime:
+  predict_concurrency : 16
+resources:
+  accelerator: H100
+  use_gpu: true
+model_name: ngram-speculator
+environment_variables:
+  hf_access_token: null
diff --git a/infrastructure/paddlepaddle/paddleocr-vl/README.md b/infrastructure/paddlepaddle/paddleocr-vl/README.md
new file mode 100644
index 000000000..d6cea6a1d
--- /dev/null
+++ b/infrastructure/paddlepaddle/paddleocr-vl/README.md
@@ -0,0 +1,52 @@
+# PaddleOCR-VL
+
+Deploy [PaddlePaddle/PaddleOCR-VL](https://huggingface.co/PaddlePaddle/PaddleOCR-VL) using a custom server configuration on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [PaddlePaddle/PaddleOCR-VL](https://huggingface.co/PaddlePaddle/PaddleOCR-VL) |
+| Task | Infrastructure / Custom server |
+| Engine | vLLM |
+| GPU | H100_40GB |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "image_url",
+          "image_url": {
+            "url": "https://ofasys-multimodal-wlcb-3-toshanghai.oss-accelerate.aliyuncs.com/wpf272043/keepme/image/receipt.png"
+          }
+        },
+        {
+          "type": "text",
+          "text": "OCR:"
+        }
+      ]
+    }
+  ],
+  "model": "PaddlePaddle/PaddleOCR-VL",
+  "max_tokens": 4096,
+  "temperature": 0.0
+}'
+```
+
+## Configuration highlights
+
+- Base image: `public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:0bf29fadf5f8b28817fbccb037fb70adaef3f7f1`
+- Predict concurrency: **128**
diff --git a/infrastructure/paddlepaddle/paddleocr-vl/config.yaml b/infrastructure/paddlepaddle/paddleocr-vl/config.yaml
new file mode 100644
index 000000000..e252cbcac
--- /dev/null
+++ b/infrastructure/paddlepaddle/paddleocr-vl/config.yaml
@@ -0,0 +1,34 @@
+description: "PaddleOCR VL for optical character recognition"
+base_image:
+  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:0bf29fadf5f8b28817fbccb037fb70adaef3f7f1
+# build_commands:
+#   - pip uninstall -y vllm
+#   - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git
+model_metadata:
+  repo_id: PaddlePaddle/PaddleOCR-VL
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: user
+        content:
+          - type: image_url
+            image_url:
+              url: "https://ofasys-multimodal-wlcb-3-toshanghai.oss-accelerate.aliyuncs.com/wpf272043/keepme/image/receipt.png"
+          - type: text
+            text: "OCR:"
+    model: "PaddlePaddle/PaddleOCR-VL"
+    max_tokens: 4096
+    temperature: 0.0
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: vllm serve PaddlePaddle/PaddleOCR-VL --trust-remote-code --max-num-batched-tokens 16384 --no-enable-prefix-caching --mm-processor-cache-gb 0 --tensor-parallel-size 1 --served-model-name PaddlePaddle/PaddleOCR-VL --host 0.0.0.0 --port 8000
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+model_name: PaddleOCR-VL
diff --git a/ip-adapter/README.md b/ip-adapter/README.md
deleted file mode 100644
index 38373cfda..000000000
--- a/ip-adapter/README.md
+++ /dev/null
@@ -1,46 +0,0 @@
-# IP Adapter
-
-This is a [Truss](https://truss.baseten.co/) for [IP Adapter](https://github.com/tencent-ailab/IP-Adapter). IP Adapter can create variations of a given input image based on a prompt, while retaining the aesthetic of the origina image.
-
-
-## Deploying IP Adapter
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd ip_adapter
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `ip_adapter` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking IP Adapter
-
-IP Adapter takes in two inputs, an optional `prompt` and a base64 encoded `image`. The output is a JSON blob with a single key, `result` with another base64 encoded image.
-
-
-```sh
-truss predict -d '{"image": "data:image/png;base64,iVBORw0KGgoA..."}'
-```
-
-You can also invoke your model via a REST API
-```
-curl -X POST https://app.baseten.co/model_versions/<YOUR_MODEL_VERSION_ID>/predict \
-     -H "Content-Type: application/json" \
-     -d '{
-           "image": "data:image/png;base64,iVBORw0KGgoA..."
-         }'
-```
diff --git a/ip-adapter/config.yaml b/ip-adapter/config.yaml
deleted file mode 100644
index e3434bf26..000000000
--- a/ip-adapter/config.yaml
+++ /dev/null
@@ -1,15 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: IP Adapter
-python_version: py311
-requirements:
-- torch==2.1.1
-- diffusers==0.24.0
-- transformers==4.35.2
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 15Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/jsonformatter/config.yaml b/jsonformatter/config.yaml
deleted file mode 100644
index 152d89d42..000000000
--- a/jsonformatter/config.yaml
+++ /dev/null
@@ -1,14 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  llm_model: databricks/dolly-v2-3b
-model_name: JsonFormatter
-python_version: py311
-requirements:
-- jsonformer
-- transformers
-- accelerate
-resources:
-  accelerator: A10G
-secrets: {}
-system_packages: []
diff --git a/kokoro/README.md b/kokoro/README.md
deleted file mode 100644
index b7cf9d7fa..000000000
--- a/kokoro/README.md
+++ /dev/null
@@ -1,13 +0,0 @@
-Kokoro is a frontier TTS model for its size of 82 million parameters (text in/audio out).
-API:
-```bash
-request:
-{"text": "Hello", "voice": "af", "speed": 1.0}
-
-text: str = defaults to "Hi, I'm kokoro"
-voice: str = defaults to "af", available options: "af", "af_bella", "af_sarah", "am_adam", "am_michael", "bf_emma", "bf_isabella", "bm_george", "bm_lewis", "af_nicole", "af_sky"
-speed: float = defaults to 1.0. The speed of the audio generated
-
-reponse:
-{"base64": "base64 encoded bytestring"}
-```
diff --git a/kokoro/config.yaml b/kokoro/config.yaml
deleted file mode 100644
index 3d645e206..000000000
--- a/kokoro/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-build_commands:
-- python3 -c "import nltk; nltk.download('punkt'); nltk.download('punkt_tab')"
-environment_variables: {}
-model_metadata:
-  example_model_input: {"text": "Kokoro is a frontier TTS model for its size of 82 million parameters (text in/audio out). On 25 Dec 2024, Kokoro v0.19 weights were permissively released in full fp32 precision under an Apache 2.0 license. As of 2 Jan 2025, 10 unique Voicepacks have been released, and a .onnx version of v0.19 is available.In the weeks leading up to its release, Kokoro v0.19 was the #1🥇 ranked model in TTS Spaces Arena. Kokoro had achieved higher Elo in this single-voice Arena setting over other models, using fewer parameters and less data. Kokoro's ability to top this Elo ladder suggests that the scaling law (Elo vs compute/data/params) for traditional TTS models might have a steeper slope than previously expected.", "voice": "af", "speed": 1.0}
-model_name: kokoro
-python_version: py311
-requirements:
-- torch==2.5.1
-- transformers==4.48.0
-- scipy==1.15.1
-- phonemizer==3.3.0
-- nltk==3.9.1
-- numpy
-- huggingface_hub[hf_transfer]
-- hf_transfer==0.1.9
-- munch==4.0.0
-resources:
-  accelerator: T4
-  use_gpu: true
-runtime:
-  predict_concurrency: 1
-secrets: {}
-system_packages:
-- espeak-ng
diff --git a/layoutlm-document-qa/README.md b/layoutlm-document-qa/README.md
deleted file mode 100644
index 7f6107773..000000000
--- a/layoutlm-document-qa/README.md
+++ /dev/null
@@ -1,76 +0,0 @@
- # LayoutLM Document QA Truss
-
-This repository packages [LayoutLM Document QA](https://huggingface.co/impira/layoutlm-document-qa) as a [Truss](https://truss.baseten.co).
-
-This multimodal model takes an image of an invoice (PNG or JPEG) and extracts information from it in response to natural language prompts.
-
-## Deploying LayoutLM Document QA
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd layoutlm-document-qa-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `layoutlm-document-qa-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking LayoutLM Document QA
-
-LayoutLM takes a dictionary with:
-
-* `url`: The URL of a PNG or JPEG of an invoice
-* `prompt`: The question to ask of the invoice
-
-Example invocation:
-
-```sh
-truss predict -d '{"url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png", "prompt": "What is the invoice number?"}'
-```
-
-Expected response:
-
-```python
-[{'answer': '9.06', 'end': 73, 'score': 0.9910207986831665, 'start': 73}]
-```
-
-You can also invoke your model via a REST API
-
-```
-curl -X POST "https://app.baseten.co/models/YOUR_MODEL_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png",
-           "prompt": "What is the invoice number?"
-         }'
-```
-
-## Hardware notes
-
-We found this model runs reasonably fast with 4 vCPUs and 16 GiB of RAM, no GPU needed. Invocation times are usually <10 seconds for the first prompt on an image (to account for image download times) then <3 seconds thereafter.
-
-Default config:
-
-```yaml
-...
-resources:
-  cpu: "4"
-  memory: 16Gi
-  use_gpu: false
-  accelerator: null
-...
-```
diff --git a/layoutlm-document-qa/config.yaml b/layoutlm-document-qa/config.yaml
deleted file mode 100644
index d3e151e41..000000000
--- a/layoutlm-document-qa/config.yaml
+++ /dev/null
@@ -1,26 +0,0 @@
-description: Extract information from images of invoices
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/impira-logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/document-qa.png
-  example_model_input:
-    prompt: What is the invoice number?
-    url: https://templates.invoicehome.com/invoice-template-us-neat-750px.png
-  tags:
-  - text-generation
-model_name: LayoutLM Document QA
-python_version: py39
-requirements:
-- Pillow==10.0.0
-- pytesseract==0.3.10
-- torch==2.0.1
-- transformers==4.30.2
-resources:
-  accelerator: null
-  cpu: '4'
-  memory: 16Gi
-  use_gpu: false
-secrets: {}
-system_packages:
-- tesseract-ocr
diff --git a/llama-cpp-server/README.md b/llama-cpp-server/README.md
deleted file mode 100644
index 01cdb580b..000000000
--- a/llama-cpp-server/README.md
+++ /dev/null
@@ -1,26 +0,0 @@
-# Llama.cpp Baseten model server
-
-Deploying llama.cpp requires a llama.cpp image with python installed and a config.yaml to deploy it. A sample config.yaml is provided in this repository. This sample deploys llama.cpp with Gemma 3 27B Instruct int4 QAT with the 1B model as a draft model.
-
-The following are instructions in case you need to build a image from source. We enable some non-default flags to optimize for performance, so please use the provided Dockerfile.
-
-### Building the docker image from source
-
-#### Prerequisites
-
-- Docker
-- NVIDIA Docker runtime
-- CUDA-capable GPU
-
-#### Building the docker image
-
-To build the docker image, use the following command:
-
-```bash
-git clone https://github.com/ggml-org/llama.cpp.git
-cp cuda.Dockerfile llama.cpp/.devops/cuda.Dockerfile
-cd llama.cpp
-docker build -t local/llama.cpp:server-cuda --target server -f .devops/cuda.Dockerfile .
-```
-
-You can then push this image to a container registry of your choice and then replace the base_image in the config.yaml
diff --git a/llama-cpp-server/config.yaml b/llama-cpp-server/config.yaml
deleted file mode 100644
index 64ebb7a85..000000000
--- a/llama-cpp-server/config.yaml
+++ /dev/null
@@ -1,46 +0,0 @@
-base_image:
-  image: alphatozeta/llama-cpp-server:0.4
-build_commands:
-  - pip install git+https://github.com/huggingface/transformers.git hf-xet
-model_metadata:
-  repo_id: google/gemma-3-27b-it-qat-q4_0-gguf
-  example_model_input: {
-    "model": "gemma",
-    "messages": [
-      {
-      "role": "user",
-      "content": [
-        {
-        "type": "text",
-        "text": "Describe this image in one sentence."
-        },
-        {
-        "type": "image_url",
-        "image_url": {
-          "url": "https://picsum.photos/id/237/200/300"
-        }
-        }
-      ]
-      }
-    ],
-    "stream": true,
-    "max_tokens": 512,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) huggingface-cli download google/gemma-3-27b-it-qat-q4_0-gguf --local-dir /app/gemma-3-27b-it-qat-q4_0-gguf && HF_TOKEN=$(cat /secrets/hf_access_token) huggingface-cli download google/gemma-3-1b-it-qat-q4_0-gguf --local-dir /app/gemma-3-1b-it-qat-q4_0-gguf && /app/llama-server -m /app/gemma-3-27b-it-qat-q4_0-gguf/gemma-3-27b-it-q4_0.gguf -md /app/gemma-3-1b-it-qat-q4_0-gguf/gemma-3-1b-it-q4_0.gguf --port 8000 -c 32768 -cd 32768 -ngl 999 -ngld 999 --draft-max 16 --draft-min 0 --prio 3 -fa --no-webui"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-requirements: []
-secrets:
-  hf_access_token: null
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency : 8
-model_name: llama cpp gemma 3 27b it qat q4_0
diff --git a/llama/engine-llama-3-1-70b-instruct/README.md b/llama/engine-llama-3-1-70b-instruct/README.md
deleted file mode 100644
index 312373439..000000000
--- a/llama/engine-llama-3-1-70b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Llama 3.1 70B Instruct
-
-This deployment of Llama 3.1 70B Instruct uses the TensorRT-LLM Engine Builder.
-
-For details, see: https://docs.baseten.co/performance/examples/llama-trt
diff --git a/llama/engine-llama-3-1-70b-instruct/config.yaml b/llama/engine-llama-3-1-70b-instruct/config.yaml
deleted file mode 100644
index 04727e12c..000000000
--- a/llama/engine-llama-3-1-70b-instruct/config.yaml
+++ /dev/null
@@ -1,56 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are a knowledgable, engaging, history teacher.",
-          },
-          {
-            role: "user",
-            content: "What was the role of Llamas in the Inca empire?",
-          },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.6,
-      top_p: 1.0,
-      top_k: 40,
-      frequency_penalty: 1,
-    }
-  repo_id: meta-llama/Llama-3.1-70B-Instruct
-model_name: Llama 3.1 70B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:2
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.1-70B-Instruct
-      source: HF
-    num_builder_gpus: 4
-    quantization_type: fp8_kv
-    max_seq_len: 131072
-    tensor_parallel_count: 2
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: true
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 131072
diff --git a/llama/engine-llama-3-1-8b-instruct/README.md b/llama/engine-llama-3-1-8b-instruct/README.md
deleted file mode 100644
index 7f8d872b3..000000000
--- a/llama/engine-llama-3-1-8b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Llama 3.1 8B Instruct
-
-This deployment of Llama 3.1 8B Instruct uses the TensorRT-LLM Engine Builder.
-
-For details, see: https://docs.baseten.co/performance/examples/llama-trt
diff --git a/llama/engine-llama-3-1-8b-instruct/config.yaml b/llama/engine-llama-3-1-8b-instruct/config.yaml
deleted file mode 100644
index 3ba5e8020..000000000
--- a/llama/engine-llama-3-1-8b-instruct/config.yaml
+++ /dev/null
@@ -1,56 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are a knowledgable, engaging, history teacher.",
-          },
-          {
-            role: "user",
-            content: "What was the role of Llamas in the Inca empire?",
-          },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.6,
-      top_p: 1.0,
-      top_k: 40,
-      frequency_penalty: 1,
-    }
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-model_name: Llama 3.1 8B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.1-8B-Instruct
-      source: HF
-    max_seq_len: 131072
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 131072
diff --git a/llama/engine-llama-3-3-70b-instruct/README.md b/llama/engine-llama-3-3-70b-instruct/README.md
deleted file mode 100644
index cd90bb5f1..000000000
--- a/llama/engine-llama-3-3-70b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Llama 3.3 70B Instruct
-
-This deployment of Llama 3.3 70B Instruct uses the TensorRT-LLM Engine Builder.
-
-For details, see: https://docs.baseten.co/performance/examples/llama-trt
diff --git a/llama/engine-llama-3-3-70b-instruct/config.yaml b/llama/engine-llama-3-3-70b-instruct/config.yaml
deleted file mode 100644
index 7492dd6f0..000000000
--- a/llama/engine-llama-3-3-70b-instruct/config.yaml
+++ /dev/null
@@ -1,56 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are a knowledgable, engaging, history teacher.",
-          },
-          {
-            role: "user",
-            content: "What was the role of Llamas in the Inca empire?",
-          },
-        ],
-      stream: true,
-      max_tokens: 1024,
-      temperature: 0.6,
-      top_p: 1.0,
-      top_k: 40,
-      frequency_penalty: 1,
-    }
-  repo_id: meta-llama/Llama-3.3-70B-Instruct
-model_name: Llama 3.3 70B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:2
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.3-70B-Instruct
-      source: HF
-    num_builder_gpus: 4
-    quantization_type: fp8_kv
-    max_seq_len: 131072
-    tensor_parallel_count: 2
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: true
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 131072
diff --git a/llama/engine-llama-3.1-405b-instruct/README.md b/llama/engine-llama-3.1-405b-instruct/README.md
deleted file mode 100644
index 51e71f95a..000000000
--- a/llama/engine-llama-3.1-405b-instruct/README.md
+++ /dev/null
@@ -1,179 +0,0 @@
-# TensorRT-LLM Briton with meta-llama/Llama-3.1-405B
-
-This is a Deployment for TensorRT-LLM Briton with meta-llama/Llama-3.1-405B. Briton is Baseten's solution for production-grade deployments via TensorRT-LLM for Causal Language Models models. (e.g. LLama, Qwen, Mistral)
-
-With Briton you get the following benefits by default:
-- *Lowest-latency* latency, beating frameworks such as vllm
-- *Highest-throughput* inference, automatically using XQA kernels, paged kv caching and inflight batching.
-- *distributed inference* run large models (such as LLama-405B) tensor-parallel
-- *json-schema based structured output for any model*
-- *chunked prefilling* for long generation tasks
-
-Optionally, you can also enable:
-- *speculative decoding* using an external draft model or self-speculative decoding
-- *fp8 quantization* deployments on H100, H200 and L4 GPUs
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [meta-llama/Llama-3.1-405B](https://huggingface.co/meta-llama/Llama-3.1-405B).
-Suitable models can be identified by the `ForCausalLM` suffix in the model name. Currently we support e.g. LLama, Qwen, Mistral models.
-
-meta-llama/Llama-3.1-405B  is a text-generation model, used to generate text given a prompt. \nIt is frequently used in chatbots, text completion, structured output and more.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/Briton-meta-llama-llama-3.1-405b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model Briton-meta-llama-llama-3.1-405b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### OpenAI compatible inference
-This solution is OpenAI compatible, which means you can use the OpenAI client library to interact with the model.
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-# Default completion
-response_completion = client.completions.create(
-    model="not_required",
-    prompt="Q: Tell me everything about Baseten.co! A:",
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Chat completion
-response_chat = client.chat.completions.create(
-    model="",
-    messages=[
-        {"role": "user", "content": "Tell me everything about Baseten.co!"}
-    ],
-    temperature=0.3,
-    max_tokens=100,
-)
-
-# Structured output
-from pydantic import BaseModel
-
-class CalendarEvent(BaseModel):
-    name: str
-    date: str
-    participants: list[str]
-
-completion = client.beta.chat.completions.parse(
-    model="not_required",
-    messages=[
-        {"role": "system", "content": "Extract the event information."},
-        {"role": "user", "content": "Alice and Bob are going to a science fair on Friday."},
-    ],
-    response_format=CalendarEvent,
-)
-
-event = completion.choices[0].message.parsed
-
-# If you model supports tool-calling, you can use the following example:
-tools = [{
-    "type": "function",
-    "function": {
-        "name": "get_weather",
-        "description": "Get current temperature for a given location.",
-        "parameters": {
-            "type": "object",
-            "properties": {
-                "location": {
-                    "type": "string",
-                    "description": "City and country e.g. Bogotá, Colombia"
-                }
-            },
-            "required": [
-                "location"
-            ],
-            "additionalProperties": False
-        },
-        "strict": True
-    }
-}]
-
-completion = client.chat.completions.create(
-    model="not_required",
-    messages=[{"role": "user", "content": "What is the weather like in Paris today?"}],
-    tools=tools
-)
-
-print(completion.choices[0].message.tool_calls)
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8_kv`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-Note: [This is a gated/private model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Do not set the actual value of key in the config.yaml. `hf_access_token: null` is fine - the true value will be fetched from the secret store.
-```yaml
-build_commands: []
-environment_variables:
-  ENABLE_EXECUTOR_API: 1
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: Tell me everything you know about optimized inference.
-      role: user
-    stream: true
-    temperature: 0.5
-  tags:
-  - openai-compatible
-model_name: Briton-meta-llama-llama-3.1-405b-fp8-truss-example
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:8
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: meta-llama/Llama-3.1-405B
-      revision: main
-      source: HF
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 8
-  runtime:
-    enable_chunked_context: true
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/llama/engine-llama-3.1-405b-instruct/config.yaml b/llama/engine-llama-3.1-405b-instruct/config.yaml
deleted file mode 100644
index f1d063fe9..000000000
--- a/llama/engine-llama-3.1-405b-instruct/config.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-build_commands: []
-environment_variables:
-  ENABLE_EXECUTOR_API: 1
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-      - content: Tell me everything you know about optimized inference.
-        role: user
-    stream: true
-    temperature: 0.5
-  tags:
-    - openai-compatible
-model_name: Briton-meta-llama-llama-3.1-405b-fp8-truss-example
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:8
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      # presigned url from: https://us-east-2.console.aws.amazon.com/s3/buckets/mp-model-weights-public?bucketType=general&region=us-east-2&tab=objects#
-      # feel free to reach out to us if you need access to this bucket
-      repo: https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar
-      source: REMOTE_URL
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 8
-  runtime:
-    enable_chunked_context: true
diff --git a/llama/llama-3-70b-instruct/README.md b/llama/llama-3-70b-instruct/README.md
deleted file mode 100644
index 2a81e4680..000000000
--- a/llama/llama-3-70b-instruct/README.md
+++ /dev/null
@@ -1,63 +0,0 @@
-# Llama 3 70B Instruct
-
-This is a [Truss](https://truss.baseten.co/) for Llama 3 70B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Llama 3 70B.
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp)). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd llama-3-70b-instruct
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `llama-3-70b-instruct` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Llama 3 70B Instruct API documentation
-
-This section provides an overview of the Llama 2 7B API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __messages__: The input text that you want the model to generate a response for.
-- __max_tokens__ (optional, default=512): The maximum number of tokens to return, counting input tokens. Maximum of 4096.
-- __temperature__ (optional, default=0.1): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results.
-- __top_p__ (optional, default=0.75): The cumulative probability threshold for token sampling. The model will only consider tokens whose cumulative probability is below this threshold.
-- __top_k__ (optional, default=40): The number of top tokens to consider when sampling. The model will only consider the top_k highest-probability tokens.
-- __num_beams__ (optional, default=4): The number of beams used for beam search. Increasing this value can result in higher-quality output but will increase the computational cost.
-
-The API also supports passing any parameter supported by HuggingFace's `Transformers.generate`.
-
-## Example usage
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "messages": [{"role": "user", "content": "What even is AGI?"}],
-           "max_tokens": 128
-         }'
-```
diff --git a/llama/llama-3-70b-instruct/config.yaml b/llama/llama-3-70b-instruct/config.yaml
deleted file mode 100644
index 22f91ab1f..000000000
--- a/llama/llama-3-70b-instruct/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/llama.png
-  repo_id: meta-llama/Meta-Llama-3-8B-Instruct
-  tags:
-  - text-generation
-model_name: Llama 3 70B Instruct
-python_version: py310
-requirements:
-  - accelerate
-  - einops
-  - transformers
-  - torch
-resources:
-  accelerator: H100:2
-  use_gpu: true
-secrets:
-  hf_access_token: "your api key"
-system_packages: []
diff --git a/llama/llama-3-8b-instruct/README.md b/llama/llama-3-8b-instruct/README.md
deleted file mode 100644
index 714a6f689..000000000
--- a/llama/llama-3-8b-instruct/README.md
+++ /dev/null
@@ -1,63 +0,0 @@
-# Llama 3 8B Instruct
-
-This is a [Truss](https://truss.baseten.co/) for Llama 3 8B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Llama 3 8B.
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp)). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd llama-3-8b-instruct
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `llama-3-8b-instruct` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Llama 3 8B Instruct API documentation
-
-This section provides an overview of the Llama 2 7B API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __messages__: The input text that you want the model to generate a response for.
-- __max_tokens__ (optional, default=512): The maximum number of tokens to return, counting input tokens. Maximum of 4096.
-- __temperature__ (optional, default=0.1): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results.
-- __top_p__ (optional, default=0.75): The cumulative probability threshold for token sampling. The model will only consider tokens whose cumulative probability is below this threshold.
-- __top_k__ (optional, default=40): The number of top tokens to consider when sampling. The model will only consider the top_k highest-probability tokens.
-- __num_beams__ (optional, default=4): The number of beams used for beam search. Increasing this value can result in higher-quality output but will increase the computational cost.
-
-The API also supports passing any parameter supported by HuggingFace's `Transformers.generate`.
-
-## Example usage
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "messages": [{"role": "user", "content": "What even is AGI?"}],
-           "max_tokens": 128
-         }'
-```
diff --git a/llama/llama-3-8b-instruct/config.yaml b/llama/llama-3-8b-instruct/config.yaml
deleted file mode 100644
index 938962ff0..000000000
--- a/llama/llama-3-8b-instruct/config.yaml
+++ /dev/null
@@ -1,24 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/llama.png
-  repo_id: meta-llama/Meta-Llama-3-8B-Instruct
-  tags:
-  - text-generation
-model_name: Llama 3 8B Instruct
-python_version: py310
-model_cache:
-  - repo_id: meta-llama/Meta-Llama-3-8B-Instruct
-    use_volume: false
-requirements:
-  - accelerate
-  - einops
-  - transformers
-  - torch
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets:
-  hf_access_token: "your-hf-access-token"
-system_packages: []
diff --git a/llama/llama-3_1-405b-instruct/README.md b/llama/llama-3_1-405b-instruct/README.md
deleted file mode 100644
index c4659a033..000000000
--- a/llama/llama-3_1-405b-instruct/README.md
+++ /dev/null
@@ -1,10 +0,0 @@
-# Llama 3.1 405B Instruct
-
-This is an implementation of Llama 3.1 405B for deployment on Baseten.
-
-- VLLM for faster inference
-- FP8 model weights
-- Runs on an 8xH100 instance
-
-Baseten offers private, secure deployments for LLMs like Llama 3.1 405B, including deployments to your own VPC.
-To deploy this model on Baseten, contact us at [support@baseten.co](support@baseten.co).
diff --git a/llama/llama-3_1-405b-instruct/config.yaml b/llama/llama-3_1-405b-instruct/config.yaml
deleted file mode 100644
index 61d9a3a59..000000000
--- a/llama/llama-3_1-405b-instruct/config.yaml
+++ /dev/null
@@ -1,16 +0,0 @@
-model_name: "Llama 3.1 405B Instruct VLLM"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: meta-llama/Llama-3.1-405B-Instruct-FP8
-  tensor_parallel: 8
-requirements:
-  - vllm==0.5.3post1
-  - transformers==4.43.1
-resources:
-  accelerator: H100:8
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
diff --git a/llama/llama-3_1-8b-instruct-sglang/config.yaml b/llama/llama-3_1-8b-instruct-sglang/config.yaml
deleted file mode 100644
index dbbd25d85..000000000
--- a/llama/llama-3_1-8b-instruct-sglang/config.yaml
+++ /dev/null
@@ -1,22 +0,0 @@
-model_name: "Llama 3.1 8B Instruct SGLang"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-  tensor_parallel: 1
-requirements:
-  - sglang[all]==0.3.0
-  - https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp311-cp311-linux_x86_64.whl
-model_cache:
-  - repo_id: meta-llama/Llama-3.1-8B-Instruct
-    use_volume: false
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
diff --git a/llama/llama-3_1-8b-instruct/config.yaml b/llama/llama-3_1-8b-instruct/config.yaml
deleted file mode 100644
index 3ab0cadd2..000000000
--- a/llama/llama-3_1-8b-instruct/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-model_name: "Llama 3.1 8B Instruct VLLM"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-  tensor_parallel: 1
-requirements:
-  - vllm==0.5.3post1
-model_cache:
-  - repo_id: meta-llama/Llama-3.1-8B-Instruct
-    use_volume: false
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
diff --git a/llama/llama-3_1_70b-instruct/config.yaml b/llama/llama-3_1_70b-instruct/config.yaml
deleted file mode 100644
index a6776abf7..000000000
--- a/llama/llama-3_1_70b-instruct/config.yaml
+++ /dev/null
@@ -1,17 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata: {}
-model_name: Llama 3.1 70B vLLM
-python_version: py310
-requirements:
-  - vllm==0.5.3post1
-  - accelerate
-resources:
-  accelerator: A100:4
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: ""
-system_packages:
-  - python3.10-venv
diff --git a/llama/llama-3_2-11b-vision-instruct/README.md b/llama/llama-3_2-11b-vision-instruct/README.md
deleted file mode 100644
index 3598138ed..000000000
--- a/llama/llama-3_2-11b-vision-instruct/README.md
+++ /dev/null
@@ -1,72 +0,0 @@
-# Llama 3.2 11B Vision Instruct vLLM Truss
-
-This is a [Truss](https://truss.baseten.co/) for Llama 3.2 11B Vision Instruct with vLLM. Llama 3.2 11B Vision Instruct is a multimodal (text + vision) LLM. This README will walk you through how to deploy this Truss on Baseten to get your own instance of it.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd llama/llama-3_2-11b-vision-instruct
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Apply for access to the Llama 3.2 11B Vision Instruct model on hugging face [here](https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct).
-4. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-5. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_token`. Note that you will *not* be able to successfully deploy this model without doing this.
-
-With `llama-3_2-11b-vision-instruct` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Notes
-
-Limitations from vLLM allow for a maximum of 1 image as input. You will get a memory error otherwise. You can keep track of the issue [here](https://github.com/vllm-project/vllm/issues/8826).
-
-## Example usage
-
-```sh
-truss predict -d '{model: "llama-3.2-11b-vision-instruct", "messages": [{"role": "user", "content": "Tell me about yourself"}]}'
-```
-
-Here's another example of invoking your model via a REST API but for image input:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "model: "llama-3.2-11b-vision-instruct",
-           "messages": [
-            {
-                "role": "user",
-                "content": [
-                {
-                    "type": "text",
-                    "text": "What type of animal is this? Answer in French only"
-                },
-                {
-                    "type": "image_url",
-                    "image_url": {
-                    "url": "https://vetmed.illinois.edu/wp-content/uploads/2021/04/pc-keller-hedgehog.jpg"
-                    }
-                }
-                ]
-            }
-            ],
-           "stream": true,
-           "max_tokens": 64,
-           "temperature": 0.2
-         }'
-```
diff --git a/llama/llama-3_2-11b-vision-instruct/config.yaml b/llama/llama-3_2-11b-vision-instruct/config.yaml
deleted file mode 100644
index eaf5a4973..000000000
--- a/llama/llama-3_2-11b-vision-instruct/config.yaml
+++ /dev/null
@@ -1,46 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.6.3.post1
-model_metadata:
-  tags:
-    - openai-compatible
-  repo_id: meta-llama/Llama-3.2-11B-Vision-Instruct
-  example_model_input: {
-    model: "llama-3.2-11b-vision-instruct",
-    messages: [
-      {
-        role: "user",
-        content: [
-          {
-            type: "text",
-            text: "Describe this image in one sentence."
-          },
-          {
-            type: "image_url",
-            image_url: {
-              url: "https://picsum.photos/id/237/200/300"
-            }
-          }
-        ]
-      }
-    ],
-    stream: true,
-    max_tokens: 512,
-    temperature: 0.5
-  }
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve meta-llama/Llama-3.2-11B-Vision-Instruct --dtype half --served-model-name llama-3.2-11b-vision-instruct --tensor-parallel-size 1 --gpu-memory-utilization 0.90 --max-model-len 4000 --max-num-seqs 8 --distributed-executor-backend mp --disable-custom-all-reduce --use-v2-block-manager --trust-remote-code --enforce-eager"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: A100
-  use_gpu: true
-model_name: Llama 3.2 11B Vision Instruct
-secrets:
-  hf_access_token: null
-environment_variables:
-  VLLM_LOGGING_LEVEL: WARNING
-  hf_access_token: null
-runtime:
-  predict_concurrency: 64
diff --git a/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/config.yaml b/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/config.yaml
deleted file mode 100644
index f75d0583f..000000000
--- a/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.8.4
-build_commands:
-  - pip install git+https://github.com/huggingface/transformers.git hf-xet
-model_metadata:
-  repo_id: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
-  example_model_input: {
-    "model": "llama",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    "stream": true,
-    "max_tokens": 512,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: sh -c /app/data/do.sh
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
-  hf_access_token: null
-resources:
-  accelerator: H100:8
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  predict_concurrency : 256
-model_name: Llama 4 Maverick 17B 128E Instruct H100 TP8
diff --git a/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/config.yaml b/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/config.yaml
deleted file mode 100755
index 96da2d062..000000000
--- a/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/config.yaml
+++ /dev/null
@@ -1,36 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.8.4
-build_commands:
-  - pip install git+https://github.com/huggingface/transformers.git hf-xet
-model_metadata:
-  repo_id: meta-llama/Llama-4-Scout-17B-16E-Instruct
-  example_model_input: {
-    "model": "llama",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    "stream": true,
-    "max_tokens": 512,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: sh -c /app/data/do.sh
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-environment_variables:
-  hf_access_token: null
-resources:
-  accelerator: H100:4
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  predict_concurrency : 256
-model_name: Llama 4 Scout 17B 16E Instruct H100 TP 4
diff --git a/llama/tinyllama-1.1B-chat-v1.0/config.yaml b/llama/tinyllama-1.1B-chat-v1.0/config.yaml
deleted file mode 100644
index 0d33221e6..000000000
--- a/llama/tinyllama-1.1B-chat-v1.0/config.yaml
+++ /dev/null
@@ -1,19 +0,0 @@
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    prompt: How tall is a tiny llama?
-model_name: tinyllama-trt
-python_version: py310
-resources:
-  accelerator: A10G
-  memory: 24Gi
-  use_gpu: true
-trt_llm:
-  build:
-    max_seq_len: 2048
-    base_model: decoder
-    quantization_type: no_quant
-    checkpoint_repository:
-      repo: TinyLlama/TinyLlama-1.1B-Chat-v1.0
-      source: HF
diff --git a/llava/llava-1.6-sgl/README.md b/llava/llava-1.6-sgl/README.md
deleted file mode 100644
index 132775d10..000000000
--- a/llava/llava-1.6-sgl/README.md
+++ /dev/null
@@ -1,77 +0,0 @@
-# LLaVA v1.6 SGL Truss
-
-This is a truss to run [Llava 1.6 using SGL](https://github.com/sgl-project/sglang)
-
-## Deploying LLaVA
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd llava/llava-v1.6-sgl
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `llava/llava-v1.6-sgl` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking LLaVA
-
-LLaVA takes in the following inputs:
- __prompt__(required): Piece of text used as instruction for the LLM
- __image__(required): Input image in the form of a base64 string used by the model
- __max_new_tokens__(optional): Max number of output tokens generated by the LLM
- __temperature__(optional): Configuration for LLM
-
-LLaVA will respond to the `prompt` conditioned on the `image`. The output is is a stream of tokens containing the model response.
-
-
-```python
-from PIL import Image
-from io import BytesIO
-import base64
-import requests
-
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-data = {
-    "prompt": "What is this a picture of?",
-    "image": pil_to_b64(Image.open("/path/to/image/mountain.jpeg")),
-}
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-<model-id>.api.baseten.co/production/predict",
-    headers=headers,
-    json=data,
-    stream=True
-)
-
-# Print the generated tokens as they get streamed
-for content in res.iter_content():
-    print(content.decode("utf-8"), end="", flush=True)
-```
-
-Sample Input:
-![mountain](https://github.com/basetenlabs/truss-examples/assets/15642666/5eb63370-0296-40ab-9387-428bf5e3cd53)
-
-Sample output:
-```
-This is a picture of Half Dome, a granite dome located in Yosemite National Park in the Sierra Nevada of California. It is one of the most iconic rock formations in the park and a popular destination for hikers and climbers. The image shows the dome with a clear blue sky and some clouds, highlighting the natural beauty of the area.
-```
diff --git a/llava/llava-1.6-sgl/config.yaml b/llava/llava-1.6-sgl/config.yaml
deleted file mode 100644
index ae2eb6956..000000000
--- a/llava/llava-1.6-sgl/config.yaml
+++ /dev/null
@@ -1,13 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: llava 1.6 SGL
-python_version: py310
-requirements: []
-requirements_file: ./requirements.txt
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets: {}
-system_packages: []
diff --git a/llava/llava-v1.5-7b/README.md b/llava/llava-v1.5-7b/README.md
deleted file mode 100644
index 48a685526..000000000
--- a/llava/llava-v1.5-7b/README.md
+++ /dev/null
@@ -1,48 +0,0 @@
-# LLaVA v1.5 Truss
-
-This repository packages [LLaVA 1.5](https://github.com/haotian-liu/LLaVA/) as a [Truss](https://truss.baseten.co/).
-
-LLaVA (Large Language and Vision Assistant) is a highly performant open-source vision language model with capabilities similar to GPT-4V.
-
-## Deploying LLaVA
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd llava/llava-v1.5-7b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `llava-v1.5-7b` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking LLaVA
-
-LLaVA takes in two inputs, a `query` and a base64 encoded `image`. LLaVA will respond to the `query` conditioned on the `image`. The output is a JSON blob with a single key, `result`, that answers the `query`.
-
-
-```sh
-truss predict -d '{"query": "Describe this picture in detail.", "image": "data:image/png;base64,iVBORw0KGgoA..."}'
-```
-
-You can also invoke your model via a REST API
-```
-curl -X POST https://app.baseten.co/model_versions/<YOUR_MODEL_VERSION_ID>/predict \
-     -H "Content-Type: application/json" \
-     -d '{
-           "query": "Describe this picture in detail.",
-           "image": "data:image/png;base64,iVBORw0KGgoA..."
-         }'
-```
diff --git a/llava/llava-v1.5-7b/config.yaml b/llava/llava-v1.5-7b/config.yaml
deleted file mode 100644
index 120a72be5..000000000
--- a/llava/llava-v1.5-7b/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: llava-v1.5-7b
-python_version: py311
-requirements:
-- torch==2.0.1
-- torchvision==0.15.2
-- transformers==4.31.0
-- tokenizers>=0.12.1,<0.14
-- sentencepiece==0.1.99
-- shortuuid==1.0.11
-- scipy==1.11.4
-- accelerate==0.21.0
-- peft==0.4.0
-- bitsandbytes==0.41.0
-- einops==0.6.1
-- einops-exts==0.0.4
-- timm==0.6.13
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 15Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/llava/llava-v1.6-34b/README.md b/llava/llava-v1.6-34b/README.md
deleted file mode 100644
index c8176989d..000000000
--- a/llava/llava-v1.6-34b/README.md
+++ /dev/null
@@ -1,48 +0,0 @@
-# LLaVA v1.6 Truss
-
-This repository packages [LLaVA 1.6](https://github.com/haotian-liu/LLaVA/) as a [Truss](https://truss.baseten.co/).
-
-LLaVA (Large Language and Vision Assistant) is a highly performant open-source vision language model with capabilities similar to GPT-4V.
-
-## Deploying LLaVA
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd llava/llava-v1.6-34b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `llava-v1.6-34b` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking LLaVA
-
-LLaVA takes in two inputs, a `query` and a base64 encoded `image`. LLaVA will respond to the `query` conditioned on the `image`. The output is a JSON blob with a single key, `result`, that answers the `query`.
-
-
-```sh
-truss predict -f input.json
-```
-
-You can also invoke your model via a REST API
-```
-curl -X POST https://app.baseten.co/model_versions/<YOUR_MODEL_VERSION_ID>/predict \
-     -H "Content-Type: application/json" \
-     -d '{
-           "query": "Describe this picture in detail.",
-           "image": "data:image/png;base64,iVBORw0KGgoA..."
-         }'
-```
diff --git a/llava/llava-v1.6-34b/config.yaml b/llava/llava-v1.6-34b/config.yaml
deleted file mode 100644
index a5beab7d8..000000000
--- a/llava/llava-v1.6-34b/config.yaml
+++ /dev/null
@@ -1,11 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: llava-v1.6-34b
-python_version: py311
-requirements:
-- git+https://github.com/haotian-liu/LLaVA.git
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/llm/README.md b/llm/README.md
new file mode 100644
index 000000000..22e6a0e31
--- /dev/null
+++ b/llm/README.md
@@ -0,0 +1,33 @@
+# Large Language Models
+
+Production-ready Truss configurations for large language models spanning a wide range of model families, sizes, and serving engines (vLLM, SGLang, TRT-LLM). Many directories contain multiple variants optimized for different hardware or quantization levels.
+
+| Directory | Models | Description |
+|-----------|--------|-------------|
+| [llama](llama/) | 14 | Meta Llama 3.x and 4.x models including 8B, 70B, 405B, vision, and TRT-LLM engine builds |
+| [qwen](qwen/) | 28 | Alibaba Qwen 2.5 and Qwen 3 models including coder, math, vision, and MoE variants |
+| [mistral](mistral/) | 15 | Mistral and Mixtral models with vLLM, TRT-LLM, and Devstral engine builds |
+| [deepseek](deepseek/) | 7 | DeepSeek R1 distilled models and vision/OCR variants |
+| [nemotron](nemotron/) | 7 | NVIDIA Nemotron models including Nano, Ultra, and vision variants |
+| [z-ai](z-ai/) | 5 | Zhipu GLM-4 models in various sizes and quantizations |
+| [cogito](cogito/) | 4 | Deep Cogito v2 Preview models on Llama and DeepSeek backbones |
+| [gemma](gemma/) | 3 | Google Gemma 2 and 3 models served with vLLM |
+| [phi](phi/) | 3 | Microsoft Phi-3 and Phi-3.5 mini instruction-tuned models |
+| [llava](llava/) | 3 | LLaVA multimodal vision-language models (v1.5, v1.6) |
+| [lora](lora/) | 3 | LoRA adapter serving with vLLM, SGLang, and TRT-LLM engines |
+| [falcon](falcon/) | 1 | TII Falcon 3 model with TRT-LLM engine |
+| [openai](openai/) | 2 | GPT-OSS 20B and 120B open-source reproductions |
+| [minimax](minimax/) | 1 | MiniMax M2-1 model |
+| [cogvlm](cogvlm/) | 1 | CogVLM visual question answering model |
+| [midnight](midnight/) | 1 | Midnight model for text generation |
+| [nsql](nsql/) | 1 | NSQL natural language to SQL model |
+| [personaplex-7b-v1](personaplex-7b-v1/) | 1 | PersonaPlex 7B persona-driven chat model |
+| [seed](seed/) | 1 | Seed LLM model |
+
+## Deploying
+
+Each model can be deployed to Baseten with:
+
+```bash
+truss push
+```
diff --git a/llm/cogito/README.md b/llm/cogito/README.md
new file mode 100644
index 000000000..968b7fe2c
--- /dev/null
+++ b/llm/cogito/README.md
@@ -0,0 +1,46 @@
+# Cogito v2 Preview Models
+
+Deploy [Deep Cogito](https://www.deepcogito.com/research/cogito-v2-preview) v2 preview models using vLLM's OpenAI-compatible server. These models feature tool calling and reasoning capabilities built on Llama and DeepSeek architectures.
+
+| Variant | Size | Architecture | GPU | Path |
+|---------|------|--------------|-----|------|
+| Cogito v2 Llama 70B | 70B | Dense | 2x H100 | [`cogito-v2-preview-llama-70B-vllm/`](cogito-v2-preview-llama-70B-vllm/) |
+| Cogito v2 Llama 109B MoE | 109B | MoE | 4x H100 | [`cogito-v2-preview-llama-109B-MoE-vllm/`](cogito-v2-preview-llama-109B-MoE-vllm/) |
+| Cogito v2 Llama 405B | 405B | Dense | 8x B200 | [`cogito-v2-preview-llama-405B-vllm/`](cogito-v2-preview-llama-405B-vllm/) |
+| Cogito v2 DeepSeek 671B MoE | 671B | MoE | 8x B200 | [`cogito-v2-preview-deepseek-671B-MoE-vllm/`](cogito-v2-preview-deepseek-671B-MoE-vllm/) |
+
+## Deploy
+
+> **Note:** These models require a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push llm/cogito/cogito-v2-preview-llama-70B-vllm
+```
+
+## Invoke
+
+All models use the OpenAI ChatCompletion format:
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/environments/production/sync/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepcogito/cogito-v2-preview-llama-70B",
+    messages=[{"role": "user", "content": "What is today's temperature in celsius? I'm in Paris."}],
+    max_tokens=1000,
+)
+
+print(response.choices[0].message.content)
+```
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepcogito/cogito-v2-preview-llama-70B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
diff --git a/llm/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/README.md b/llm/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/README.md
new file mode 100644
index 000000000..521085c82
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/README.md
@@ -0,0 +1,57 @@
+# Cogito V2 Preview DeepSeek 671B MoE FP8 vLLM
+
+Deploy [deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8](https://huggingface.co/deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8](https://huggingface.co/deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | B200:8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.9.2`
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/config.yaml b/llm/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/config.yaml
new file mode 100644
index 000000000..e84b85284
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-deepseek-671B-MoE-vllm/config.yaml
@@ -0,0 +1,37 @@
+description: "deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8 for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.9.2
+model_metadata:
+  example_model_input: {
+    model: "deepseek",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    stream: true,
+    max_tokens: 10000,
+    temperature: 0.6
+  }
+  repo_id: deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8 --served-model-name deepseek --max-model-len 131072 --port 8000 --gpu-memory-utilization 0.90 --disable-custom-all-reduce --trust-remote-code --tensor-parallel-size 8 --distributed-executor-backend mp --enable-auto-tool-choice --tool-call-parser deepseek_v3"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+cache_internal:
+  - repo_id: deepcogito/cogito-v2-preview-deepseek-671B-MoE-FP8
+resources:
+  accelerator: B200:8
+  cpu: '1'
+  memory: 24Gi
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: Cogito V2 Preview DeepSeek 671B MoE FP8 vLLM
+environment_variables:
+  hf_access_token: null
diff --git a/llm/cogito/cogito-v2-preview-llama-109B-MoE-vllm/README.md b/llm/cogito/cogito-v2-preview-llama-109B-MoE-vllm/README.md
new file mode 100644
index 000000000..568ab5357
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-llama-109B-MoE-vllm/README.md
@@ -0,0 +1,58 @@
+# Cogito V2 Preview Llama 109B MoE vLLM
+
+Deploy [deepcogito/cogito-v2-preview-llama-109B-MoE](https://huggingface.co/deepcogito/cogito-v2-preview-llama-109B-MoE) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepcogito/cogito-v2-preview-llama-109B-MoE](https://huggingface.co/deepcogito/cogito-v2-preview-llama-109B-MoE) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:4 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepcogito/cogito-v2-preview-llama-109B-MoE",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepcogito/cogito-v2-preview-llama-109B-MoE", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.10.0`
+- Predict concurrency: **32**
+- Streaming: **enabled**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/cogito/cogito-v2-preview-llama-109B-MoE-vllm/config.yaml b/llm/cogito/cogito-v2-preview-llama-109B-MoE-vllm/config.yaml
new file mode 100644
index 000000000..ab1430702
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-llama-109B-MoE-vllm/config.yaml
@@ -0,0 +1,36 @@
+description: "deepcogito/cogito-v2-preview-llama-109B-MoE for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.10.0
+model_metadata:
+  repo_id: deepcogito/cogito-v2-preview-llama-109B-MoE
+  example_model_input: {
+    "model": "llama",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    "stream": true,
+    "max_tokens": 10000,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-llama-109B-MoE --served-model-name llama --max-model-len 32000 --tensor-parallel-size 4 --distributed-executor-backend mp --enable-auto-tool-choice --tool-call-parser llama3_json"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
+  hf_access_token: null
+resources:
+  accelerator: H100:4
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  predict_concurrency : 32
+model_name: Cogito V2 Preview Llama 109B MoE vLLM
diff --git a/llm/cogito/cogito-v2-preview-llama-405B-vllm/README.md b/llm/cogito/cogito-v2-preview-llama-405B-vllm/README.md
new file mode 100644
index 000000000..c041e4d3b
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-llama-405B-vllm/README.md
@@ -0,0 +1,58 @@
+# Cogito V2 Preview Llama 405B vLLM
+
+Deploy [deepcogito/cogito-v2-preview-llama-405B](https://huggingface.co/deepcogito/cogito-v2-preview-llama-405B) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepcogito/cogito-v2-preview-llama-405B](https://huggingface.co/deepcogito/cogito-v2-preview-llama-405B) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | B200:8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepcogito/cogito-v2-preview-llama-405B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepcogito/cogito-v2-preview-llama-405B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.10.0`
+- Predict concurrency: **32**
+- Streaming: **enabled**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/cogito/cogito-v2-preview-llama-405B-vllm/config.yaml b/llm/cogito/cogito-v2-preview-llama-405B-vllm/config.yaml
new file mode 100644
index 000000000..3a39d5d77
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-llama-405B-vllm/config.yaml
@@ -0,0 +1,38 @@
+description: "deepcogito/cogito-v2-preview-llama-405B for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.10.0
+model_metadata:
+  repo_id: deepcogito/cogito-v2-preview-llama-405B
+  example_model_input: {
+    "model": "llama",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    "stream": true,
+    "max_tokens": 10000,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-llama-405B --served-model-name llama --max-model-len 32000 --tensor-parallel-size 8 --enable-chunked-prefill --enable-prefix-caching --max-num-seqs 8 --distributed-executor-backend mp --enable-auto-tool-choice --tool-call-parser llama3_json "
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
+  hf_access_token: null
+cache_internal:
+  - repo_id: deepcogito/cogito-v2-preview-llama-405B
+resources:
+  accelerator: B200:8
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  predict_concurrency : 32
+model_name: Cogito V2 Preview Llama 405B vLLM
diff --git a/llm/cogito/cogito-v2-preview-llama-70B-vllm/README.md b/llm/cogito/cogito-v2-preview-llama-70B-vllm/README.md
new file mode 100644
index 000000000..54cba8e41
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-llama-70B-vllm/README.md
@@ -0,0 +1,58 @@
+# Cogito V2 Preview Llama 70B vLLM
+
+Deploy [deepcogito/cogito-v2-preview-llama-70B](https://huggingface.co/deepcogito/cogito-v2-preview-llama-70B) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepcogito/cogito-v2-preview-llama-70B](https://huggingface.co/deepcogito/cogito-v2-preview-llama-70B) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepcogito/cogito-v2-preview-llama-70B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepcogito/cogito-v2-preview-llama-70B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.10.0`
+- Predict concurrency: **32**
+- Streaming: **enabled**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/cogito/cogito-v2-preview-llama-70B-vllm/config.yaml b/llm/cogito/cogito-v2-preview-llama-70B-vllm/config.yaml
new file mode 100755
index 000000000..fe75fe12a
--- /dev/null
+++ b/llm/cogito/cogito-v2-preview-llama-70B-vllm/config.yaml
@@ -0,0 +1,36 @@
+description: "deepcogito/cogito-v2-preview-llama-70B for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.10.0
+model_metadata:
+  repo_id: deepcogito/cogito-v2-preview-llama-70B
+  example_model_input: {
+    "model": "llama",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    "stream": true,
+    "max_tokens": 10000,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve deepcogito/cogito-v2-preview-llama-70B --served-model-name llama --max-model-len 32000 --tensor-parallel-size 2 --distributed-executor-backend mp --gpu-memory-utilization 0.95 --enable-auto-tool-choice --tool-call-parser llama3_json"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
+  hf_access_token: null
+resources:
+  accelerator: H100:2
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  predict_concurrency : 32
+model_name: Cogito V2 Preview Llama 70B vLLM
diff --git a/llm/cogvlm/README.md b/llm/cogvlm/README.md
new file mode 100644
index 000000000..4a1708d62
--- /dev/null
+++ b/llm/cogvlm/README.md
@@ -0,0 +1,29 @@
+# CogVLM
+
+Deploy CogVLM for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/cogvlm/config.yaml b/llm/cogvlm/config.yaml
new file mode 100644
index 000000000..3e26abc1a
--- /dev/null
+++ b/llm/cogvlm/config.yaml
@@ -0,0 +1,25 @@
+description: "CogVLM for vision-language tasks"
+environment_variables: {}
+external_package_dirs: []
+model_name: CogVLM
+python_version: py311
+requirements:
+- torch==2.0.1
+- sentencepiece==0.1.99
+- protobuf==4.25.1
+- transformers==4.35.2
+- einops==0.7.0
+- torchvision==0.15.2
+- Pillow==10.1.0
+- xformers==0.0.22
+- accelerate==0.25.0
+model_metadata:
+  repo_id: "THUDM/cogvlm-chat-hf"
+  example_model_input: {"query": "Describe this image in detail", "image": "https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg"}
+resources:
+  accelerator: A100
+  cpu: '3'
+  memory: 15Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py b/llm/cogvlm/model/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py
rename to llm/cogvlm/model/__init__.py
diff --git a/cogvlm/model/model.py b/llm/cogvlm/model/model.py
similarity index 100%
rename from cogvlm/model/model.py
rename to llm/cogvlm/model/model.py
diff --git a/deepseek-ocr/Bad-Handwriting.png b/llm/deepseek/deepseek-ocr/Bad-Handwriting.png
similarity index 100%
rename from deepseek-ocr/Bad-Handwriting.png
rename to llm/deepseek/deepseek-ocr/Bad-Handwriting.png
diff --git a/llm/deepseek/deepseek-ocr/README.md b/llm/deepseek/deepseek-ocr/README.md
new file mode 100644
index 000000000..b796c9112
--- /dev/null
+++ b/llm/deepseek/deepseek-ocr/README.md
@@ -0,0 +1,54 @@
+# deepseek-ocr-latest
+
+Deploy [deepseek-ai/DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-OCR](https://huggingface.co/deepseek-ai/DeepSeek-OCR) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100_40GB |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-OCR",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-OCR", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang@sha256:bb19265cdc61a65a158b84fb69d84f885f4c5f55e12e4515be88223ec067cf50`
+- Predict concurrency: **256**
diff --git a/llm/deepseek/deepseek-ocr/config.yaml b/llm/deepseek/deepseek-ocr/config.yaml
new file mode 100644
index 000000000..6d629e107
--- /dev/null
+++ b/llm/deepseek/deepseek-ocr/config.yaml
@@ -0,0 +1,31 @@
+description: "DeepSeek OCR for optical character recognition"
+model_metadata:
+  repo_id: "deepseek-ai/DeepSeek-OCR"
+  example_model_input:
+    model: "deepseek-ai/DeepSeek-OCR"
+    messages:
+      - role: user
+        content:
+          - type: image_url
+            image_url:
+              url: "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg=="
+          - type: text
+            text: "<|grounding|>Convert the document to markdown."
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: deepseek-ocr-latest
+base_image:
+  image: lmsysorg/sglang@sha256:bb19265cdc61a65a158b84fb69d84f885f4c5f55e12e4515be88223ec067cf50
+docker_server:
+  start_command: sh -c "python3 -m sglang.launch_server --model deepseek-ai/DeepSeek-OCR --served-model-name deepseek-ai/DeepSeek-OCR --host 0.0.0.0 --port 8000"
+  readiness_endpoint: /health_generate
+  liveness_endpoint: /health_generate
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+runtime:
+  predict_concurrency: 256
diff --git a/deepseek-ocr/model/__init__.py b/llm/deepseek/deepseek-ocr/model/__init__.py
similarity index 100%
rename from deepseek-ocr/model/__init__.py
rename to llm/deepseek/deepseek-ocr/model/__init__.py
diff --git a/deepseek-ocr/model/model.py b/llm/deepseek/deepseek-ocr/model/model.py
similarity index 100%
rename from deepseek-ocr/model/model.py
rename to llm/deepseek/deepseek-ocr/model/model.py
diff --git a/deepseek-ocr/test_document_ocr.py b/llm/deepseek/deepseek-ocr/test_document_ocr.py
similarity index 100%
rename from deepseek-ocr/test_document_ocr.py
rename to llm/deepseek/deepseek-ocr/test_document_ocr.py
diff --git a/deepseek-ocr/visualization_1.png b/llm/deepseek/deepseek-ocr/visualization_1.png
similarity index 100%
rename from deepseek-ocr/visualization_1.png
rename to llm/deepseek/deepseek-ocr/visualization_1.png
diff --git a/deepseek-ocr/visualization_2.png b/llm/deepseek/deepseek-ocr/visualization_2.png
similarity index 100%
rename from deepseek-ocr/visualization_2.png
rename to llm/deepseek/deepseek-ocr/visualization_2.png
diff --git a/deepseek-ocr/visualizer.py b/llm/deepseek/deepseek-ocr/visualizer.py
similarity index 100%
rename from deepseek-ocr/visualizer.py
rename to llm/deepseek/deepseek-ocr/visualizer.py
diff --git a/llm/deepseek/deepseek-vl2/README.md b/llm/deepseek/deepseek-vl2/README.md
new file mode 100644
index 000000000..19cb0dfda
--- /dev/null
+++ b/llm/deepseek/deepseek-vl2/README.md
@@ -0,0 +1,54 @@
+# deepseek vl2
+
+Deploy [deepseek-ai/deepseek-vl2](https://huggingface.co/deepseek-ai/deepseek-vl2) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/deepseek-vl2](https://huggingface.co/deepseek-ai/deepseek-vl2) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/deepseek-vl2",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/deepseek-vl2", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.5.5`
+- Predict concurrency: **256**
diff --git a/llm/deepseek/deepseek-vl2/config.yaml b/llm/deepseek/deepseek-vl2/config.yaml
new file mode 100644
index 000000000..215398d22
--- /dev/null
+++ b/llm/deepseek/deepseek-vl2/config.yaml
@@ -0,0 +1,31 @@
+description: "DeepSeek VL2 for vision-language tasks"
+model_metadata:
+  repo_id: "deepseek-ai/deepseek-vl2"
+  example_model_input:
+    model: "deepseek-ai/deepseek-vl2"
+    messages:
+      - role: user
+        content:
+          - type: image_url
+            image_url:
+              url: "data:image/png;base64,iVBORw0KGgoAAAANSUhEUgAAAAEAAAABCAYAAAAfFcSJAAAADUlEQVR42mNk+M9QDwADhgGAWjR9awAAAABJRU5ErkJggg=="
+          - type: text
+            text: "<|grounding|>Convert the document to markdown."
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: deepseek vl2
+base_image:
+  image: lmsysorg/sglang:v0.5.5
+docker_server:
+  start_command: sh -c "python3 -m sglang.launch_server --model deepseek-ai/deepseek-vl2 --served-model-name deepseek-ai/deepseek-vl2 --host 0.0.0.0 --port 8000"
+  readiness_endpoint: /health_generate
+  liveness_endpoint: /health_generate
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency: 256
diff --git a/llm/deepseek/engine-deepseek-r1-distill-llama-70b/README.md b/llm/deepseek/engine-deepseek-r1-distill-llama-70b/README.md
new file mode 100644
index 000000000..f80e01a53
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-llama-70b/README.md
@@ -0,0 +1,65 @@
+# DeepSeek R1 Distill Llama 70B
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **use_fp8_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/deepseek/engine-deepseek-r1-distill-llama-70b/config.yaml b/llm/deepseek/engine-deepseek-r1-distill-llama-70b/config.yaml
new file mode 100644
index 000000000..ae853c96d
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-llama-70b/config.yaml
@@ -0,0 +1,53 @@
+description: "deepseek-ai/DeepSeek-R1-Distill-Llama-70B for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "user",
+            content: "Which is heavier, a pound of bricks or a pound of feathers?",
+          },
+        ],
+      stream: true,
+      max_tokens: 1024,
+      temperature: 0.6,
+      top_p: 1.0,
+      top_k: 40,
+      frequency_penalty: 1,
+    }
+  repo_id: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
+model_name: DeepSeek R1 Distill Llama 70B
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:2
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
+      source: HF
+    num_builder_gpus: 4
+    quantization_type: fp8_kv
+    max_seq_len: 131072
+    tensor_parallel_count: 2
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: true
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 131072
diff --git a/llm/deepseek/engine-deepseek-r1-distill-llama-8b/README.md b/llm/deepseek/engine-deepseek-r1-distill-llama-8b/README.md
new file mode 100644
index 000000000..2804c2ece
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-llama-8b/README.md
@@ -0,0 +1,63 @@
+# DeepSeek R1 Distill Llama 8B
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Llama-8B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-8B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Llama-8B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Llama-8B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/deepseek/engine-deepseek-r1-distill-llama-8b/config.yaml b/llm/deepseek/engine-deepseek-r1-distill-llama-8b/config.yaml
new file mode 100644
index 000000000..ccecc83b0
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-llama-8b/config.yaml
@@ -0,0 +1,53 @@
+description: "deepseek-ai/DeepSeek-R1-Distill-Llama-8B for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "user",
+            content: "Which is heavier, a pound of bricks or a pound of feathers?",
+          },
+        ],
+      stream: true,
+      max_tokens: 1024,
+      temperature: 0.6,
+      top_p: 1.0,
+      top_k: 40,
+      frequency_penalty: 1,
+    }
+  repo_id: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
+model_name: DeepSeek R1 Distill Llama 8B
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-8B
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    max_seq_len: 131072
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 131072
diff --git a/llm/deepseek/engine-deepseek-r1-distill-qwen-14b/README.md b/llm/deepseek/engine-deepseek-r1-distill-qwen-14b/README.md
new file mode 100644
index 000000000..c4c52db3c
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-qwen-14b/README.md
@@ -0,0 +1,61 @@
+# DeepSeek R1 Distill Qwen 14B
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Qwen-14B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-14B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Qwen-14B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml b/llm/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml
new file mode 100644
index 000000000..c49363fb7
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-qwen-14b/config.yaml
@@ -0,0 +1,49 @@
+description: "deepseek-ai/DeepSeek-R1-Distill-Qwen-14B for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "user",
+            content: "Which is heavier, a pound of bricks or a pound of feathers?",
+          },
+        ],
+      stream: true,
+      max_tokens: 1024,
+      temperature: 0.6,
+    }
+  repo_id: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
+model_name: DeepSeek R1 Distill Qwen 14B
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-14B
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/deepseek/engine-deepseek-r1-distill-qwen-32b/README.md b/llm/deepseek/engine-deepseek-r1-distill-qwen-32b/README.md
new file mode 100644
index 000000000..beb82dada
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-qwen-32b/README.md
@@ -0,0 +1,61 @@
+# DeepSeek R1 Distill Qwen 32B
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml b/llm/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml
new file mode 100644
index 000000000..70260fbfc
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-qwen-32b/config.yaml
@@ -0,0 +1,49 @@
+description: "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "user",
+            content: "Which is heavier, a pound of bricks or a pound of feathers?",
+          },
+        ],
+      stream: true,
+      max_tokens: 1024,
+      temperature: 0.6,
+    }
+  repo_id: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
+model_name: DeepSeek R1 Distill Qwen 32B
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
+      source: HF
+    num_builder_gpus: 2
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/deepseek/engine-deepseek-r1-distill-qwen-7b/README.md b/llm/deepseek/engine-deepseek-r1-distill-qwen-7b/README.md
new file mode 100644
index 000000000..769e306a3
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-qwen-7b/README.md
@@ -0,0 +1,61 @@
+# DeepSeek R1 Distill Qwen 7B
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Qwen-7B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Qwen-7B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/deepseek/engine-deepseek-r1-distill-qwen-7b/config.yaml b/llm/deepseek/engine-deepseek-r1-distill-qwen-7b/config.yaml
new file mode 100644
index 000000000..1bd546125
--- /dev/null
+++ b/llm/deepseek/engine-deepseek-r1-distill-qwen-7b/config.yaml
@@ -0,0 +1,49 @@
+description: "deepseek-ai/DeepSeek-R1-Distill-Qwen-7B for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "user",
+            content: "Which is heavier, a pound of bricks or a pound of feathers?",
+          },
+        ],
+      stream: true,
+      max_tokens: 1024,
+      temperature: 0.6,
+    }
+  repo_id: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
+model_name: DeepSeek R1 Distill Qwen 7B
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-7B
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/falcon/_archive/falcon3-10B-trt-llm-spec-dec/README.md b/llm/falcon/_archive/falcon3-10B-trt-llm-spec-dec/README.md
new file mode 100644
index 000000000..72e9e0944
--- /dev/null
+++ b/llm/falcon/_archive/falcon3-10B-trt-llm-spec-dec/README.md
@@ -0,0 +1,51 @@
+# Falcon 3 10B Instruct
+
+Deploy [tiiuae/Falcon3-10B-Instruct](https://huggingface.co/tiiuae/Falcon3-10B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [tiiuae/Falcon3-10B-Instruct](https://huggingface.co/tiiuae/Falcon3-10B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | NO QUANT |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "frequency_penalty": 1,
+  "max_tokens": 512,
+  "messages": [
+    {
+      "content": "You are a knowledgable, engaging, biology teacher.",
+      "role": "system"
+    },
+    {
+      "content": "What makes falcons effective hunters?",
+      "role": "user"
+    }
+  ],
+  "stream": true,
+  "temperature": 0.6
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Speculative decoding: **DRAFT_TOKENS_EXTERNAL**
+- Max sequence length: **8,192**
+- Chunked context: **enabled**
+- Plugin: **paged_kv_cache**
+- Plugin: **use_paged_context_fmha**
+- Streaming: **enabled**
diff --git a/falcon/falcon3-10B-trt-llm-spec-dec/config.yaml b/llm/falcon/_archive/falcon3-10B-trt-llm-spec-dec/config.yaml
similarity index 100%
rename from falcon/falcon3-10B-trt-llm-spec-dec/config.yaml
rename to llm/falcon/_archive/falcon3-10B-trt-llm-spec-dec/config.yaml
diff --git a/llm/falcon/falcon3-3B-trt-llm-engine-high-throughput/README.md b/llm/falcon/falcon3-3B-trt-llm-engine-high-throughput/README.md
new file mode 100644
index 000000000..b50eebeed
--- /dev/null
+++ b/llm/falcon/falcon3-3B-trt-llm-engine-high-throughput/README.md
@@ -0,0 +1,52 @@
+# Falcon 3 3B Instruct
+
+Deploy [tiiuae/Falcon3-3B-Instruct](https://huggingface.co/tiiuae/Falcon3-3B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [tiiuae/Falcon3-3B-Instruct](https://huggingface.co/tiiuae/Falcon3-3B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | A10G |
+| Quantization | NO QUANT |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "frequency_penalty": 1,
+  "max_tokens": 512,
+  "messages": [
+    {
+      "content": "You are a knowledgable, engaging, biology teacher.",
+      "role": "system"
+    },
+    {
+      "content": "What makes falcons effective hunters?",
+      "role": "user"
+    }
+  ],
+  "stream": true,
+  "temperature": 0.6
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **8,192**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **paged_kv_cache**
+- Plugin: **use_paged_context_fmha**
+- Streaming: **enabled**
diff --git a/llm/falcon/falcon3-3B-trt-llm-engine-high-throughput/config.yaml b/llm/falcon/falcon3-3B-trt-llm-engine-high-throughput/config.yaml
new file mode 100644
index 000000000..0d606c4e3
--- /dev/null
+++ b/llm/falcon/falcon3-3B-trt-llm-engine-high-throughput/config.yaml
@@ -0,0 +1,45 @@
+description: "tiiuae/Falcon3-3B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    frequency_penalty: 1
+    max_tokens: 512
+    messages:
+      - content: You are a knowledgable, engaging, biology teacher.
+        role: system
+      - content: What makes falcons effective hunters?
+        role: user
+    stream: true
+    temperature: 0.6
+  repo_id: tiiuae/Falcon3-3B-Instruct
+model_name: Falcon 3 3B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: A10G
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: tiiuae/Falcon3-3B-Instruct
+      source: HF
+    max_seq_len: 8192
+    num_builder_gpus: 1
+    plugin_configuration:
+      paged_kv_cache: true
+      use_paged_context_fmha: true
+    quantization_type: no_quant
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    kv_cache_free_gpu_mem_fraction: 0.85
+    request_default_max_tokens: 8192
diff --git a/llm/gemma/gemma-2-27b-it-vllm/README.md b/llm/gemma/gemma-2-27b-it-vllm/README.md
new file mode 100644
index 000000000..de26a50ee
--- /dev/null
+++ b/llm/gemma/gemma-2-27b-it-vllm/README.md
@@ -0,0 +1,34 @@
+# Gemma 2 27B Instruct VLLM
+
+Deploy [google/gemma-2-27b-it](https://huggingface.co/google/gemma-2-27b-it) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [google/gemma-2-27b-it](https://huggingface.co/google/gemma-2-27b-it) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "what is the meaning of life"
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/gemma/gemma-2-27b-it-vllm/config.yaml b/llm/gemma/gemma-2-27b-it-vllm/config.yaml
new file mode 100644
index 000000000..6efea777c
--- /dev/null
+++ b/llm/gemma/gemma-2-27b-it-vllm/config.yaml
@@ -0,0 +1,18 @@
+description: "google/gemma-2-27b-it for text generation"
+model_name: "Gemma 2 27B Instruct VLLM"
+python_version: py311
+model_metadata:
+  example_model_input: {"prompt": "what is the meaning of life"}
+  repo_id: google/gemma-2-27b-it
+  tensor_parallel: 1
+  max_num_seqs: 16
+requirements:
+  - vllm==0.5.1
+  - https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp311-cp311-linux_x86_64.whl
+resources:
+  accelerator: A100
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets:
+  hf_access_token: null
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py b/llm/gemma/gemma-2-27b-it-vllm/model/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py
rename to llm/gemma/gemma-2-27b-it-vllm/model/__init__.py
diff --git a/gemma/gemma-2-27b-it-vllm/model/model.py b/llm/gemma/gemma-2-27b-it-vllm/model/model.py
similarity index 100%
rename from gemma/gemma-2-27b-it-vllm/model/model.py
rename to llm/gemma/gemma-2-27b-it-vllm/model/model.py
diff --git a/llm/gemma/gemma-2-9b-it-vllm/README.md b/llm/gemma/gemma-2-9b-it-vllm/README.md
new file mode 100644
index 000000000..761165113
--- /dev/null
+++ b/llm/gemma/gemma-2-9b-it-vllm/README.md
@@ -0,0 +1,34 @@
+# Gemma 2 9B Instruct VLLM
+
+Deploy [google/gemma-2-9b-it](https://huggingface.co/google/gemma-2-9b-it) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [google/gemma-2-9b-it](https://huggingface.co/google/gemma-2-9b-it) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "what is the meaning of life"
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/gemma/gemma-2-9b-it-vllm/config.yaml b/llm/gemma/gemma-2-9b-it-vllm/config.yaml
new file mode 100644
index 000000000..f0d3f16d5
--- /dev/null
+++ b/llm/gemma/gemma-2-9b-it-vllm/config.yaml
@@ -0,0 +1,17 @@
+description: "google/gemma-2-9b-it for text generation"
+model_name: "Gemma 2 9B Instruct VLLM"
+python_version: py311
+model_metadata:
+  example_model_input: {"prompt": "what is the meaning of life"}
+  repo_id: google/gemma-2-9b-it
+  tensor_parallel: 1
+requirements:
+  - vllm==0.5.1
+  - https://github.com/flashinfer-ai/flashinfer/releases/download/v0.0.8/flashinfer-0.0.8+cu121torch2.3-cp311-cp311-linux_x86_64.whl
+resources:
+  accelerator: A100
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets:
+  hf_access_token: null
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py b/llm/gemma/gemma-2-9b-it-vllm/model/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py
rename to llm/gemma/gemma-2-9b-it-vllm/model/__init__.py
diff --git a/gemma/gemma-2-9b-it-vllm/model/model.py b/llm/gemma/gemma-2-9b-it-vllm/model/model.py
similarity index 100%
rename from gemma/gemma-2-9b-it-vllm/model/model.py
rename to llm/gemma/gemma-2-9b-it-vllm/model/model.py
diff --git a/llm/gemma/gemma-3-27b-it/README.md b/llm/gemma/gemma-3-27b-it/README.md
new file mode 100644
index 000000000..c1704c5c4
--- /dev/null
+++ b/llm/gemma/gemma-3-27b-it/README.md
@@ -0,0 +1,59 @@
+# Gemma 27B Instruct
+
+Deploy [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [google/gemma-3-27b-it](https://huggingface.co/google/gemma-3-27b-it) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="google/gemma-3-27b-it",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "google/gemma-3-27b-it", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:8a4a2efc6fc32cdc30e4e35ba3f8c64dcd0aa1d0`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **8**
+- Streaming: **enabled**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/gemma/gemma-3-27b-it/config.yaml b/llm/gemma/gemma-3-27b-it/config.yaml
new file mode 100644
index 000000000..e2b5513f3
--- /dev/null
+++ b/llm/gemma/gemma-3-27b-it/config.yaml
@@ -0,0 +1,62 @@
+description: "google/gemma-3-27b-it for text generation"
+base_image:
+  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:8a4a2efc6fc32cdc30e4e35ba3f8c64dcd0aa1d0
+build_commands:
+  - pip install git+https://github.com/huggingface/transformers@071a161d3e38f56dbda2743b979f0afeed2cd4f1
+model_metadata:
+  repo_id: google/gemma-3-27b-it
+  example_model_input: {
+    "model": "gemma",
+    "messages": [
+      {
+      "role": "user",
+      "content": [
+        {
+        "type": "text",
+        "text": "Describe this image in one sentence."
+        },
+        {
+        "type": "image_url",
+        "image_url": {
+          "url": "https://picsum.photos/id/237/200/300"
+        }
+        }
+      ]
+      }
+    ],
+    "stream": true,
+    "max_tokens": 512,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: "sh -c \"truss-transfer-cli && VLLM_USE_V1=1 HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve /app/model_cache/gemma --served-model-name gemma --max-num-seqs 8 --max-model-len 16384 --limit_mm_per_prompt 'image=1' --hf-overrides '{\\\"do_pan_and_scan\\\": true}' --gpu-memory-utilization 0.95\""
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
+model_cache:
+  - repo_id: google/gemma-3-27b-it
+    revision: 005ad3404e59d6023443cb575daa05336842228a
+    use_volume: true
+    volume_folder: gemma
+requirements:
+- huggingface_hub==0.19.4
+- hf_transfer==0.1.4
+- datasets==2.16.1
+resources:
+  accelerator: H100
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  health_checks:
+    restart_check_delay_seconds: 300      # Waits 5 minutes after deployment before starting health checks
+    restart_threshold_seconds: 300       # Triggers a restart if health checks fail for 5 minutes
+    stop_traffic_threshold_seconds: 120  # Stops traffic if health checks fail for 2 minutes
+  predict_concurrency : 8
+  truss_server_version_override: "0.11.4"
+model_name: Gemma 27B Instruct
diff --git a/llama/llama-2-13b-chat/README.md b/llm/llama/_archive/llama-2-13b-chat/README.md
similarity index 100%
rename from llama/llama-2-13b-chat/README.md
rename to llm/llama/_archive/llama-2-13b-chat/README.md
diff --git a/llama/llama-2-13b-chat/config.yaml b/llm/llama/_archive/llama-2-13b-chat/config.yaml
similarity index 100%
rename from llama/llama-2-13b-chat/config.yaml
rename to llm/llama/_archive/llama-2-13b-chat/config.yaml
diff --git a/sana/sana_1600M/packages/Sana/diffusion/utils/__init__.py b/llm/llama/_archive/llama-2-13b-chat/model/__init__.py
old mode 100755
new mode 100644
similarity index 100%
rename from sana/sana_1600M/packages/Sana/diffusion/utils/__init__.py
rename to llm/llama/_archive/llama-2-13b-chat/model/__init__.py
diff --git a/llama/llama-2-13b-chat/model/model.py b/llm/llama/_archive/llama-2-13b-chat/model/model.py
similarity index 100%
rename from llama/llama-2-13b-chat/model/model.py
rename to llm/llama/_archive/llama-2-13b-chat/model/model.py
diff --git a/llama/llama-2-13b/README.md b/llm/llama/_archive/llama-2-13b/README.md
similarity index 100%
rename from llama/llama-2-13b/README.md
rename to llm/llama/_archive/llama-2-13b/README.md
diff --git a/llama/llama-2-13b/config.yaml b/llm/llama/_archive/llama-2-13b/config.yaml
similarity index 100%
rename from llama/llama-2-13b/config.yaml
rename to llm/llama/_archive/llama-2-13b/config.yaml
diff --git a/sana/sana_1600M/packages/Sana/tools/__init__.py b/llm/llama/_archive/llama-2-13b/model/__init__.py
similarity index 100%
rename from sana/sana_1600M/packages/Sana/tools/__init__.py
rename to llm/llama/_archive/llama-2-13b/model/__init__.py
diff --git a/llama/llama-2-13b/model/model.py b/llm/llama/_archive/llama-2-13b/model/model.py
similarity index 100%
rename from llama/llama-2-13b/model/model.py
rename to llm/llama/_archive/llama-2-13b/model/model.py
diff --git a/llama/llama-2-70b-chat/README.md b/llm/llama/_archive/llama-2-70b-chat/README.md
similarity index 100%
rename from llama/llama-2-70b-chat/README.md
rename to llm/llama/_archive/llama-2-70b-chat/README.md
diff --git a/llama/llama-2-70b-chat/config.yaml b/llm/llama/_archive/llama-2-70b-chat/config.yaml
similarity index 100%
rename from llama/llama-2-70b-chat/config.yaml
rename to llm/llama/_archive/llama-2-70b-chat/config.yaml
diff --git a/sana/sana_600M/model/__init__.py b/llm/llama/_archive/llama-2-70b-chat/model/__init__.py
similarity index 100%
rename from sana/sana_600M/model/__init__.py
rename to llm/llama/_archive/llama-2-70b-chat/model/__init__.py
diff --git a/llama/llama-2-70b-chat/model/model.py b/llm/llama/_archive/llama-2-70b-chat/model/model.py
similarity index 100%
rename from llama/llama-2-70b-chat/model/model.py
rename to llm/llama/_archive/llama-2-70b-chat/model/model.py
diff --git a/llama/llama-2-70b/README.md b/llm/llama/_archive/llama-2-70b/README.md
similarity index 100%
rename from llama/llama-2-70b/README.md
rename to llm/llama/_archive/llama-2-70b/README.md
diff --git a/llama/llama-2-70b/config.yaml b/llm/llama/_archive/llama-2-70b/config.yaml
similarity index 100%
rename from llama/llama-2-70b/config.yaml
rename to llm/llama/_archive/llama-2-70b/config.yaml
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/__init__.py b/llm/llama/_archive/llama-2-70b/model/__init__.py
old mode 100755
new mode 100644
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/__init__.py
rename to llm/llama/_archive/llama-2-70b/model/__init__.py
diff --git a/llama/llama-2-70b/model/model.py b/llm/llama/_archive/llama-2-70b/model/model.py
similarity index 100%
rename from llama/llama-2-70b/model/model.py
rename to llm/llama/_archive/llama-2-70b/model/model.py
diff --git a/llama/llama-2-7b-chat/README.md b/llm/llama/_archive/llama-2-7b-chat/README.md
similarity index 100%
rename from llama/llama-2-7b-chat/README.md
rename to llm/llama/_archive/llama-2-7b-chat/README.md
diff --git a/llama/llama-2-7b-chat/config.yaml b/llm/llama/_archive/llama-2-7b-chat/config.yaml
similarity index 100%
rename from llama/llama-2-7b-chat/config.yaml
rename to llm/llama/_archive/llama-2-7b-chat/config.yaml
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py b/llm/llama/_archive/llama-2-7b-chat/model/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/__init__.py
rename to llm/llama/_archive/llama-2-7b-chat/model/__init__.py
diff --git a/llama/llama-2-7b-chat/model/model.py b/llm/llama/_archive/llama-2-7b-chat/model/model.py
similarity index 100%
rename from llama/llama-2-7b-chat/model/model.py
rename to llm/llama/_archive/llama-2-7b-chat/model/model.py
diff --git a/llama/llama-2-7b/README.md b/llm/llama/_archive/llama-2-7b/README.md
similarity index 100%
rename from llama/llama-2-7b/README.md
rename to llm/llama/_archive/llama-2-7b/README.md
diff --git a/llama/llama-2-7b/config.yaml b/llm/llama/_archive/llama-2-7b/config.yaml
similarity index 100%
rename from llama/llama-2-7b/config.yaml
rename to llm/llama/_archive/llama-2-7b/config.yaml
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py b/llm/llama/_archive/llama-2-7b/model/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/apps/__init__.py
rename to llm/llama/_archive/llama-2-7b/model/__init__.py
diff --git a/llama/llama-2-7b/model/model.py b/llm/llama/_archive/llama-2-7b/model/model.py
similarity index 100%
rename from llama/llama-2-7b/model/model.py
rename to llm/llama/_archive/llama-2-7b/model/model.py
diff --git a/llama/llama-7b-exllama-streaming/README.md b/llm/llama/_archive/llama-7b-exllama-streaming/README.md
similarity index 100%
rename from llama/llama-7b-exllama-streaming/README.md
rename to llm/llama/_archive/llama-7b-exllama-streaming/README.md
diff --git a/llama/llama-7b-exllama-streaming/config.yaml b/llm/llama/_archive/llama-7b-exllama-streaming/config.yaml
similarity index 100%
rename from llama/llama-7b-exllama-streaming/config.yaml
rename to llm/llama/_archive/llama-7b-exllama-streaming/config.yaml
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py b/llm/llama/_archive/llama-7b-exllama-streaming/model/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/model/dc_ae/efficientvit/models/__init__.py
rename to llm/llama/_archive/llama-7b-exllama-streaming/model/__init__.py
diff --git a/llama/llama-7b-exllama-streaming/model/model.py b/llm/llama/_archive/llama-7b-exllama-streaming/model/model.py
similarity index 100%
rename from llama/llama-7b-exllama-streaming/model/model.py
rename to llm/llama/_archive/llama-7b-exllama-streaming/model/model.py
diff --git a/llama/llama-7b-exllama/README.md b/llm/llama/_archive/llama-7b-exllama/README.md
similarity index 100%
rename from llama/llama-7b-exllama/README.md
rename to llm/llama/_archive/llama-7b-exllama/README.md
diff --git a/llama/llama-7b-exllama/config.yaml b/llm/llama/_archive/llama-7b-exllama/config.yaml
similarity index 100%
rename from llama/llama-7b-exllama/config.yaml
rename to llm/llama/_archive/llama-7b-exllama/config.yaml
diff --git a/sana/sana_600M/packages/Sana/diffusion/utils/__init__.py b/llm/llama/_archive/llama-7b-exllama/model/__init__.py
old mode 100755
new mode 100644
similarity index 100%
rename from sana/sana_600M/packages/Sana/diffusion/utils/__init__.py
rename to llm/llama/_archive/llama-7b-exllama/model/__init__.py
diff --git a/llama/llama-7b-exllama/model/model.py b/llm/llama/_archive/llama-7b-exllama/model/model.py
similarity index 100%
rename from llama/llama-7b-exllama/model/model.py
rename to llm/llama/_archive/llama-7b-exllama/model/model.py
diff --git a/llama/llama-7b-vllm/config.yaml b/llm/llama/_archive/llama-7b-vllm/config.yaml
similarity index 100%
rename from llama/llama-7b-vllm/config.yaml
rename to llm/llama/_archive/llama-7b-vllm/config.yaml
diff --git a/sana/sana_600M/packages/Sana/tools/__init__.py b/llm/llama/_archive/llama-7b-vllm/model/__init__.py
similarity index 100%
rename from sana/sana_600M/packages/Sana/tools/__init__.py
rename to llm/llama/_archive/llama-7b-vllm/model/__init__.py
diff --git a/llama/llama-7b-vllm/model/model.py b/llm/llama/_archive/llama-7b-vllm/model/model.py
similarity index 100%
rename from llama/llama-7b-vllm/model/model.py
rename to llm/llama/_archive/llama-7b-vllm/model/model.py
diff --git a/llama/llama-7b/README.md b/llm/llama/_archive/llama-7b/README.md
similarity index 100%
rename from llama/llama-7b/README.md
rename to llm/llama/_archive/llama-7b/README.md
diff --git a/llama/llama-7b/config.yaml b/llm/llama/_archive/llama-7b/config.yaml
similarity index 100%
rename from llama/llama-7b/config.yaml
rename to llm/llama/_archive/llama-7b/config.yaml
diff --git a/llama/llama-7b/data/config.json b/llm/llama/_archive/llama-7b/data/config.json
similarity index 100%
rename from llama/llama-7b/data/config.json
rename to llm/llama/_archive/llama-7b/data/config.json
diff --git a/llama/llama-7b/data/generation_config.json b/llm/llama/_archive/llama-7b/data/generation_config.json
similarity index 100%
rename from llama/llama-7b/data/generation_config.json
rename to llm/llama/_archive/llama-7b/data/generation_config.json
diff --git a/llama/llama-7b/data/pytorch_model.bin.index.json b/llm/llama/_archive/llama-7b/data/pytorch_model.bin.index.json
similarity index 100%
rename from llama/llama-7b/data/pytorch_model.bin.index.json
rename to llm/llama/_archive/llama-7b/data/pytorch_model.bin.index.json
diff --git a/segment-anything/model/__init__.py b/llm/llama/_archive/llama-7b/model/__init__.py
similarity index 100%
rename from segment-anything/model/__init__.py
rename to llm/llama/_archive/llama-7b/model/__init__.py
diff --git a/llama/llama-7b/model/model.py b/llm/llama/_archive/llama-7b/model/model.py
similarity index 100%
rename from llama/llama-7b/model/model.py
rename to llm/llama/_archive/llama-7b/model/model.py
diff --git a/llm/llama/engine-llama-3-1-70b-instruct/README.md b/llm/llama/engine-llama-3-1-70b-instruct/README.md
new file mode 100644
index 000000000..75cbf93ff
--- /dev/null
+++ b/llm/llama/engine-llama-3-1-70b-instruct/README.md
@@ -0,0 +1,65 @@
+# Llama 3.1 70B Instruct
+
+Deploy [meta-llama/Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-70B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.1-70B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-70B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **use_fp8_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/llama/engine-llama-3-1-70b-instruct/config.yaml b/llm/llama/engine-llama-3-1-70b-instruct/config.yaml
new file mode 100644
index 000000000..d86a44901
--- /dev/null
+++ b/llm/llama/engine-llama-3-1-70b-instruct/config.yaml
@@ -0,0 +1,57 @@
+description: "meta-llama/Llama-3.1-70B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are a knowledgable, engaging, history teacher.",
+          },
+          {
+            role: "user",
+            content: "What was the role of Llamas in the Inca empire?",
+          },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.6,
+      top_p: 1.0,
+      top_k: 40,
+      frequency_penalty: 1,
+    }
+  repo_id: meta-llama/Llama-3.1-70B-Instruct
+model_name: Llama 3.1 70B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:2
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.1-70B-Instruct
+      source: HF
+    num_builder_gpus: 4
+    quantization_type: fp8_kv
+    max_seq_len: 131072
+    tensor_parallel_count: 2
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: true
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 131072
diff --git a/llm/llama/engine-llama-3-1-8b-instruct/README.md b/llm/llama/engine-llama-3-1-8b-instruct/README.md
new file mode 100644
index 000000000..1c5cadba0
--- /dev/null
+++ b/llm/llama/engine-llama-3-1-8b-instruct/README.md
@@ -0,0 +1,63 @@
+# Llama 3.1 8B Instruct
+
+Deploy [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.1-8B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-8B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/llama/engine-llama-3-1-8b-instruct/config.yaml b/llm/llama/engine-llama-3-1-8b-instruct/config.yaml
new file mode 100644
index 000000000..b9a43b774
--- /dev/null
+++ b/llm/llama/engine-llama-3-1-8b-instruct/config.yaml
@@ -0,0 +1,57 @@
+description: "meta-llama/Llama-3.1-8B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are a knowledgable, engaging, history teacher.",
+          },
+          {
+            role: "user",
+            content: "What was the role of Llamas in the Inca empire?",
+          },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.6,
+      top_p: 1.0,
+      top_k: 40,
+      frequency_penalty: 1,
+    }
+  repo_id: meta-llama/Llama-3.1-8B-Instruct
+model_name: Llama 3.1 8B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.1-8B-Instruct
+      source: HF
+    max_seq_len: 131072
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 131072
diff --git a/llm/llama/engine-llama-3-3-70b-instruct/README.md b/llm/llama/engine-llama-3-3-70b-instruct/README.md
new file mode 100644
index 000000000..2b9a33ead
--- /dev/null
+++ b/llm/llama/engine-llama-3-3-70b-instruct/README.md
@@ -0,0 +1,65 @@
+# Llama 3.3 70B Instruct
+
+Deploy [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.3-70B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.3-70B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **use_fp8_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/llama/engine-llama-3-3-70b-instruct/config.yaml b/llm/llama/engine-llama-3-3-70b-instruct/config.yaml
new file mode 100644
index 000000000..12524bf21
--- /dev/null
+++ b/llm/llama/engine-llama-3-3-70b-instruct/config.yaml
@@ -0,0 +1,57 @@
+description: "meta-llama/Llama-3.3-70B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are a knowledgable, engaging, history teacher.",
+          },
+          {
+            role: "user",
+            content: "What was the role of Llamas in the Inca empire?",
+          },
+        ],
+      stream: true,
+      max_tokens: 1024,
+      temperature: 0.6,
+      top_p: 1.0,
+      top_k: 40,
+      frequency_penalty: 1,
+    }
+  repo_id: meta-llama/Llama-3.3-70B-Instruct
+model_name: Llama 3.3 70B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:2
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.3-70B-Instruct
+      source: HF
+    num_builder_gpus: 4
+    quantization_type: fp8_kv
+    max_seq_len: 131072
+    tensor_parallel_count: 2
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: true
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 131072
diff --git a/llm/llama/engine-llama-3.1-405b-instruct/README.md b/llm/llama/engine-llama-3.1-405b-instruct/README.md
new file mode 100644
index 000000000..2a2f8d456
--- /dev/null
+++ b/llm/llama/engine-llama-3.1-405b-instruct/README.md
@@ -0,0 +1,63 @@
+# Briton-meta-llama-llama-3.1-405b-fp8-truss-example
+
+Deploy [https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar](https://huggingface.co/https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar](https://huggingface.co/https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:8 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **8** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
+- Environment variables: `ENABLE_EXECUTOR_API`
diff --git a/llm/llama/engine-llama-3.1-405b-instruct/config.yaml b/llm/llama/engine-llama-3.1-405b-instruct/config.yaml
new file mode 100644
index 000000000..9eecc559c
--- /dev/null
+++ b/llm/llama/engine-llama-3.1-405b-instruct/config.yaml
@@ -0,0 +1,41 @@
+description: "Briton-meta-llama-llama-3.1-405b-fp8-truss-example for text generation"
+build_commands: []
+environment_variables:
+  ENABLE_EXECUTOR_API: 1
+external_package_dirs: []
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+      - content: Tell me everything you know about optimized inference.
+        role: user
+    stream: true
+    temperature: 0.5
+  tags:
+    - openai-compatible
+model_name: Briton-meta-llama-llama-3.1-405b-fp8-truss-example
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:8
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+secrets:
+  hf_access_token: null
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      # presigned url from: https://us-east-2.console.aws.amazon.com/s3/buckets/mp-model-weights-public?bucketType=general&region=us-east-2&tab=objects#
+      # feel free to reach out to us if you need access to this bucket
+      repo: https://mp-model-weights-public.s3.us-east-2.amazonaws.com/llama-405b-tp8-fp8kv-tllm.tar
+      source: REMOTE_URL
+    max_seq_len: 131072
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 8
+  runtime:
+    enable_chunked_context: true
diff --git a/llm/llama/llama-3-1-405b-instruct/README.md b/llm/llama/llama-3-1-405b-instruct/README.md
new file mode 100644
index 000000000..0aba694f9
--- /dev/null
+++ b/llm/llama/llama-3-1-405b-instruct/README.md
@@ -0,0 +1,34 @@
+# Llama 3.1 405B Instruct VLLM
+
+Deploy [meta-llama/Llama-3.1-405B-Instruct-FP8](https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct-FP8) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-405B-Instruct-FP8](https://huggingface.co/meta-llama/Llama-3.1-405B-Instruct-FP8) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | H100:8 |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "what is the meaning of life"
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/llama/llama-3-1-405b-instruct/config.yaml b/llm/llama/llama-3-1-405b-instruct/config.yaml
new file mode 100644
index 000000000..7d7ce18be
--- /dev/null
+++ b/llm/llama/llama-3-1-405b-instruct/config.yaml
@@ -0,0 +1,17 @@
+description: "meta-llama/Llama-3.1-405B-Instruct-FP8 for text generation"
+model_name: "Llama 3.1 405B Instruct VLLM"
+python_version: py311
+model_metadata:
+  example_model_input: {"prompt": "what is the meaning of life"}
+  repo_id: meta-llama/Llama-3.1-405B-Instruct-FP8
+  tensor_parallel: 8
+requirements:
+  - vllm==0.5.3post1
+  - transformers==4.43.1
+resources:
+  accelerator: H100:8
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets:
+  hf_access_token: null
diff --git a/sesame-csm-1b/model/__init__.py b/llm/llama/llama-3-1-405b-instruct/model/__init__.py
similarity index 100%
rename from sesame-csm-1b/model/__init__.py
rename to llm/llama/llama-3-1-405b-instruct/model/__init__.py
diff --git a/llama/llama-3_1-405b-instruct/model/model.py b/llm/llama/llama-3-1-405b-instruct/model/model.py
similarity index 100%
rename from llama/llama-3_1-405b-instruct/model/model.py
rename to llm/llama/llama-3-1-405b-instruct/model/model.py
diff --git a/llama/llama-3_1-405b-instruct/model/sighelper.py b/llm/llama/llama-3-1-405b-instruct/model/sighelper.py
similarity index 100%
rename from llama/llama-3_1-405b-instruct/model/sighelper.py
rename to llm/llama/llama-3-1-405b-instruct/model/sighelper.py
diff --git a/llm/llama/llama-3-1-70b-instruct/README.md b/llm/llama/llama-3-1-70b-instruct/README.md
new file mode 100644
index 000000000..625566964
--- /dev/null
+++ b/llm/llama/llama-3-1-70b-instruct/README.md
@@ -0,0 +1,32 @@
+# Llama 3.1 70B vLLM
+
+Deploy Llama 3.1 70B vLLM for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100:4 |
+| Python | py310 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
+- System packages: `python3.10-venv`
diff --git a/llm/llama/llama-3-1-70b-instruct/config.yaml b/llm/llama/llama-3-1-70b-instruct/config.yaml
new file mode 100644
index 000000000..bc59de7e0
--- /dev/null
+++ b/llm/llama/llama-3-1-70b-instruct/config.yaml
@@ -0,0 +1,20 @@
+description: "Llama 3.1 70B vLLM for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "meta-llama/Llama-3.1-70B-Instruct"
+  example_model_input: {"prompt": "What is the meaning of life?"}
+model_name: Llama 3.1 70B vLLM
+python_version: py310
+requirements:
+  - vllm==0.5.3post1
+  - accelerate==0.25.0
+resources:
+  accelerator: A100:4
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets:
+  hf_access_token: ""
+system_packages:
+  - python3.10-venv
diff --git a/stable-diffusion/dreamshaper-lcm/model/__init__.py b/llm/llama/llama-3-1-70b-instruct/model/__init__.py
similarity index 100%
rename from stable-diffusion/dreamshaper-lcm/model/__init__.py
rename to llm/llama/llama-3-1-70b-instruct/model/__init__.py
diff --git a/llama/llama-3_1_70b-instruct/model/model.py b/llm/llama/llama-3-1-70b-instruct/model/model.py
similarity index 100%
rename from llama/llama-3_1_70b-instruct/model/model.py
rename to llm/llama/llama-3-1-70b-instruct/model/model.py
diff --git a/llama/llama-3_1_70b-instruct/model/sighelper.py b/llm/llama/llama-3-1-70b-instruct/model/sighelper.py
similarity index 100%
rename from llama/llama-3_1_70b-instruct/model/sighelper.py
rename to llm/llama/llama-3-1-70b-instruct/model/sighelper.py
diff --git a/llm/llama/llama-3-1-8b-instruct-sglang/README.md b/llm/llama/llama-3-1-8b-instruct-sglang/README.md
new file mode 100644
index 000000000..2704a90e5
--- /dev/null
+++ b/llm/llama/llama-3-1-8b-instruct-sglang/README.md
@@ -0,0 +1,34 @@
+# Llama 3.1 8B Instruct SGLang
+
+Deploy [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100 |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "what is the meaning of life"
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/llama/llama-3-1-8b-instruct-sglang/config.yaml b/llm/llama/llama-3-1-8b-instruct-sglang/config.yaml
new file mode 100644
index 000000000..b55650a94
--- /dev/null
+++ b/llm/llama/llama-3-1-8b-instruct-sglang/config.yaml
@@ -0,0 +1,23 @@
+description: "meta-llama/Llama-3.1-8B-Instruct for text generation"
+model_name: "Llama 3.1 8B Instruct SGLang"
+python_version: py311
+model_metadata:
+  example_model_input: {"prompt": "what is the meaning of life"}
+  repo_id: meta-llama/Llama-3.1-8B-Instruct
+  tensor_parallel: 1
+requirements:
+  - sglang[all]==0.3.0
+  - https://github.com/flashinfer-ai/flashinfer/releases/download/v0.1.6/flashinfer-0.1.6+cu121torch2.4-cp311-cp311-linux_x86_64.whl
+model_cache:
+  - repo_id: meta-llama/Llama-3.1-8B-Instruct
+    use_volume: false
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets:
+  hf_access_token: null
diff --git a/stable-diffusion/playground-v2-trt/model/__init__.py b/llm/llama/llama-3-1-8b-instruct-sglang/model/__init__.py
similarity index 100%
rename from stable-diffusion/playground-v2-trt/model/__init__.py
rename to llm/llama/llama-3-1-8b-instruct-sglang/model/__init__.py
diff --git a/llama/llama-3_1-8b-instruct-sglang/model/model.py b/llm/llama/llama-3-1-8b-instruct-sglang/model/model.py
similarity index 100%
rename from llama/llama-3_1-8b-instruct-sglang/model/model.py
rename to llm/llama/llama-3-1-8b-instruct-sglang/model/model.py
diff --git a/llm/llama/llama-3-1-8b-instruct/README.md b/llm/llama/llama-3-1-8b-instruct/README.md
new file mode 100644
index 000000000..61abf3733
--- /dev/null
+++ b/llm/llama/llama-3-1-8b-instruct/README.md
@@ -0,0 +1,34 @@
+# Llama 3.1 8B Instruct VLLM
+
+Deploy [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | H100_40GB |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "what is the meaning of life"
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/llama/llama-3-1-8b-instruct/config.yaml b/llm/llama/llama-3-1-8b-instruct/config.yaml
new file mode 100644
index 000000000..7778ec92b
--- /dev/null
+++ b/llm/llama/llama-3-1-8b-instruct/config.yaml
@@ -0,0 +1,22 @@
+description: "meta-llama/Llama-3.1-8B-Instruct for text generation"
+model_name: "Llama 3.1 8B Instruct VLLM"
+python_version: py311
+model_metadata:
+  example_model_input: {"prompt": "what is the meaning of life"}
+  repo_id: meta-llama/Llama-3.1-8B-Instruct
+  tensor_parallel: 1
+requirements:
+  - vllm==0.5.3post1
+model_cache:
+  - repo_id: meta-llama/Llama-3.1-8B-Instruct
+    use_volume: false
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+resources:
+  accelerator: H100_40GB
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets:
+  hf_access_token: null
diff --git a/stable-diffusion/sd-textual-inversion/model/__init__.py b/llm/llama/llama-3-1-8b-instruct/model/__init__.py
similarity index 100%
rename from stable-diffusion/sd-textual-inversion/model/__init__.py
rename to llm/llama/llama-3-1-8b-instruct/model/__init__.py
diff --git a/llama/llama-3_1-8b-instruct/model/model.py b/llm/llama/llama-3-1-8b-instruct/model/model.py
similarity index 100%
rename from llama/llama-3_1-8b-instruct/model/model.py
rename to llm/llama/llama-3-1-8b-instruct/model/model.py
diff --git a/llm/llama/llama-3-2-11b-vision-instruct/README.md b/llm/llama/llama-3-2-11b-vision-instruct/README.md
new file mode 100644
index 000000000..7a6cc6be6
--- /dev/null
+++ b/llm/llama/llama-3-2-11b-vision-instruct/README.md
@@ -0,0 +1,58 @@
+# Llama 3.2 11B Vision Instruct
+
+Deploy [meta-llama/Llama-3.2-11B-Vision-Instruct](https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.2-11B-Vision-Instruct](https://huggingface.co/meta-llama/Llama-3.2-11B-Vision-Instruct) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | A100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.2-11B-Vision-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.2-11B-Vision-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.6.3.post1`
+- Predict concurrency: **64**
+- Streaming: **enabled**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/llama/llama-3-2-11b-vision-instruct/config.yaml b/llm/llama/llama-3-2-11b-vision-instruct/config.yaml
new file mode 100644
index 000000000..756897ef5
--- /dev/null
+++ b/llm/llama/llama-3-2-11b-vision-instruct/config.yaml
@@ -0,0 +1,47 @@
+description: "Llama 3.2 11B Vision Instruct for vision-language tasks"
+base_image:
+  image: vllm/vllm-openai:v0.6.3.post1
+model_metadata:
+  tags:
+    - openai-compatible
+  repo_id: meta-llama/Llama-3.2-11B-Vision-Instruct
+  example_model_input: {
+    model: "llama-3.2-11b-vision-instruct",
+    messages: [
+      {
+        role: "user",
+        content: [
+          {
+            type: "text",
+            text: "Describe this image in one sentence."
+          },
+          {
+            type: "image_url",
+            image_url: {
+              url: "https://picsum.photos/id/237/200/300"
+            }
+          }
+        ]
+      }
+    ],
+    stream: true,
+    max_tokens: 512,
+    temperature: 0.5
+  }
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve meta-llama/Llama-3.2-11B-Vision-Instruct --dtype half --served-model-name llama-3.2-11b-vision-instruct --tensor-parallel-size 1 --gpu-memory-utilization 0.90 --max-model-len 4000 --max-num-seqs 8 --distributed-executor-backend mp --disable-custom-all-reduce --use-v2-block-manager --trust-remote-code --enforce-eager"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: A100
+  use_gpu: true
+model_name: Llama 3.2 11B Vision Instruct
+secrets:
+  hf_access_token: null
+environment_variables:
+  VLLM_LOGGING_LEVEL: WARNING
+  hf_access_token: null
+runtime:
+  predict_concurrency: 64
diff --git a/llm/llama/llama-3-70b-instruct/README.md b/llm/llama/llama-3-70b-instruct/README.md
new file mode 100644
index 000000000..09f81552e
--- /dev/null
+++ b/llm/llama/llama-3-70b-instruct/README.md
@@ -0,0 +1,32 @@
+# Llama 3 70B Instruct
+
+Deploy [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | H100:2 |
+| Python | py310 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/llama/llama-3-70b-instruct/config.yaml b/llm/llama/llama-3-70b-instruct/config.yaml
new file mode 100644
index 000000000..08b49bc23
--- /dev/null
+++ b/llm/llama/llama-3-70b-instruct/config.yaml
@@ -0,0 +1,23 @@
+description: "meta-llama/Meta-Llama-3-8B-Instruct for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/llama.png
+  repo_id: meta-llama/Meta-Llama-3-8B-Instruct
+  tags:
+  - text-generation
+  example_model_input: {"messages": [{"role": "user", "content": "What is the meaning of life?"}]}
+model_name: Llama 3 70B Instruct
+python_version: py310
+requirements:
+  - accelerate==0.25.0
+  - einops==0.7.0
+  - transformers==4.36.0
+  - torch==2.1.0
+resources:
+  accelerator: H100:2
+  use_gpu: true
+secrets:
+  hf_access_token: "your api key"
+system_packages: []
diff --git a/stable-diffusion/sd-turbo/model/__init__.py b/llm/llama/llama-3-70b-instruct/model/__init__.py
similarity index 100%
rename from stable-diffusion/sd-turbo/model/__init__.py
rename to llm/llama/llama-3-70b-instruct/model/__init__.py
diff --git a/llama/llama-3-70b-instruct/model/model.py b/llm/llama/llama-3-70b-instruct/model/model.py
similarity index 100%
rename from llama/llama-3-70b-instruct/model/model.py
rename to llm/llama/llama-3-70b-instruct/model/model.py
diff --git a/llm/llama/llama-3-8b-instruct/README.md b/llm/llama/llama-3-8b-instruct/README.md
new file mode 100644
index 000000000..46974edf5
--- /dev/null
+++ b/llm/llama/llama-3-8b-instruct/README.md
@@ -0,0 +1,32 @@
+# Llama 3 8B Instruct
+
+Deploy [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Meta-Llama-3-8B-Instruct](https://huggingface.co/meta-llama/Meta-Llama-3-8B-Instruct) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py310 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/llama/llama-3-8b-instruct/config.yaml b/llm/llama/llama-3-8b-instruct/config.yaml
new file mode 100644
index 000000000..3119f054e
--- /dev/null
+++ b/llm/llama/llama-3-8b-instruct/config.yaml
@@ -0,0 +1,26 @@
+description: "meta-llama/Meta-Llama-3-8B-Instruct for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/llama.png
+  repo_id: meta-llama/Meta-Llama-3-8B-Instruct
+  tags:
+  - text-generation
+  example_model_input: {"messages": [{"role": "user", "content": "What is the meaning of life?"}]}
+model_name: Llama 3 8B Instruct
+python_version: py310
+model_cache:
+  - repo_id: meta-llama/Meta-Llama-3-8B-Instruct
+    use_volume: false
+requirements:
+  - accelerate==0.25.0
+  - einops==0.7.0
+  - transformers==4.36.0
+  - torch==2.1.0
+resources:
+  accelerator: A100
+  use_gpu: true
+secrets:
+  hf_access_token: "your-hf-access-token"
+system_packages: []
diff --git a/stable-diffusion/sdxl-controlnet-canny/model/__init__.py b/llm/llama/llama-3-8b-instruct/model/__init__.py
similarity index 100%
rename from stable-diffusion/sdxl-controlnet-canny/model/__init__.py
rename to llm/llama/llama-3-8b-instruct/model/__init__.py
diff --git a/llama/llama-3-8b-instruct/model/model.py b/llm/llama/llama-3-8b-instruct/model/model.py
similarity index 100%
rename from llama/llama-3-8b-instruct/model/model.py
rename to llm/llama/llama-3-8b-instruct/model/model.py
diff --git a/llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/README.md b/llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/README.md
new file mode 100644
index 000000000..118bf3e1a
--- /dev/null
+++ b/llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/README.md
@@ -0,0 +1,59 @@
+# Llama 4 Maverick 17B 128E Instruct H100 TP8
+
+Deploy [meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8](https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8](https://huggingface.co/meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:8 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.8.4`
+- Predict concurrency: **256**
+- Streaming: **enabled**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/config.yaml b/llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/config.yaml
new file mode 100644
index 000000000..a98aa5e2e
--- /dev/null
+++ b/llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/config.yaml
@@ -0,0 +1,38 @@
+description: "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8 for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.8.4
+build_commands:
+  - pip install git+https://github.com/huggingface/transformers.git hf-xet
+model_metadata:
+  repo_id: meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8
+  example_model_input: {
+    "model": "meta-llama/Llama-4-Maverick-17B-128E-Instruct-FP8",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    "stream": true,
+    "max_tokens": 512,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: sh -c /app/data/do.sh
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
+  hf_access_token: null
+resources:
+  accelerator: H100:8
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  predict_concurrency : 256
+model_name: Llama 4 Maverick 17B 128E Instruct H100 TP8
diff --git a/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/data/do.sh b/llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/data/do.sh
similarity index 100%
rename from llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/data/do.sh
rename to llm/llama/llama-4-maverick-17b-128e-instruct-fp8-vllm/data/do.sh
diff --git a/llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/README.md b/llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/README.md
new file mode 100644
index 000000000..2ba9618d2
--- /dev/null
+++ b/llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/README.md
@@ -0,0 +1,57 @@
+# Llama 4 Scout 17B 16E Instruct H100 TP 4
+
+Deploy [meta-llama/Llama-4-Scout-17B-16E-Instruct](https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-4-Scout-17B-16E-Instruct](https://huggingface.co/meta-llama/Llama-4-Scout-17B-16E-Instruct) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:4 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-4-Scout-17B-16E-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-4-Scout-17B-16E-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.8.4`
+- Predict concurrency: **256**
+- Streaming: **enabled**
diff --git a/llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/config.yaml b/llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/config.yaml
new file mode 100755
index 000000000..c836a3fae
--- /dev/null
+++ b/llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/config.yaml
@@ -0,0 +1,37 @@
+description: "meta-llama/Llama-4-Scout-17B-16E-Instruct for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.8.4
+build_commands:
+  - pip install git+https://github.com/huggingface/transformers.git hf-xet
+model_metadata:
+  repo_id: meta-llama/Llama-4-Scout-17B-16E-Instruct
+  example_model_input: {
+    "model": "llama",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    "stream": true,
+    "max_tokens": 512,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: sh -c /app/data/do.sh
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+environment_variables:
+  hf_access_token: null
+resources:
+  accelerator: H100:4
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  predict_concurrency : 256
+model_name: Llama 4 Scout 17B 16E Instruct H100 TP 4
diff --git a/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/data/do.sh b/llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/data/do.sh
similarity index 100%
rename from llama/llama-4-scout-17b-16e-instruct-bf16-vllm/data/do.sh
rename to llm/llama/llama-4-scout-17b-16e-instruct-bf16-vllm/data/do.sh
diff --git a/llm/llama/tinyllama-1.1B-chat-v1.0/README.md b/llm/llama/tinyllama-1.1B-chat-v1.0/README.md
new file mode 100644
index 000000000..5992e39c9
--- /dev/null
+++ b/llm/llama/tinyllama-1.1B-chat-v1.0/README.md
@@ -0,0 +1,56 @@
+# tinyllama-trt
+
+Deploy [TinyLlama/TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [TinyLlama/TinyLlama-1.1B-Chat-v1.0](https://huggingface.co/TinyLlama/TinyLlama-1.1B-Chat-v1.0) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | A10G |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="TinyLlama/TinyLlama-1.1B-Chat-v1.0",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "TinyLlama/TinyLlama-1.1B-Chat-v1.0", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **2,048**
diff --git a/llm/llama/tinyllama-1.1B-chat-v1.0/config.yaml b/llm/llama/tinyllama-1.1B-chat-v1.0/config.yaml
new file mode 100644
index 000000000..24d464ef1
--- /dev/null
+++ b/llm/llama/tinyllama-1.1B-chat-v1.0/config.yaml
@@ -0,0 +1,20 @@
+description: "TinyLlama/TinyLlama-1.1B-Chat-v1.0 for text generation"
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    prompt: How tall is a tiny llama?
+model_name: tinyllama-trt
+python_version: py310
+resources:
+  accelerator: A10G
+  memory: 24Gi
+  use_gpu: true
+trt_llm:
+  build:
+    max_seq_len: 2048
+    base_model: decoder
+    quantization_type: no_quant
+    checkpoint_repository:
+      repo: TinyLlama/TinyLlama-1.1B-Chat-v1.0
+      source: HF
diff --git a/llm/llava/llava-1.6-sgl/README.md b/llm/llava/llava-1.6-sgl/README.md
new file mode 100644
index 000000000..320ad500b
--- /dev/null
+++ b/llm/llava/llava-1.6-sgl/README.md
@@ -0,0 +1,29 @@
+# llava 1.6 SGL
+
+Deploy llava 1.6 SGL for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/llava/llava-1.6-sgl/config.yaml b/llm/llava/llava-1.6-sgl/config.yaml
new file mode 100644
index 000000000..54a0fe5b7
--- /dev/null
+++ b/llm/llava/llava-1.6-sgl/config.yaml
@@ -0,0 +1,17 @@
+description: "LLaVA 1.6 for vision-language tasks"
+environment_variables: {}
+external_package_dirs: []
+model_name: llava 1.6 SGL
+python_version: py310
+requirements: []
+requirements_file: ./requirements.txt
+model_metadata:
+  repo_id: "liuhaotian/llava-v1.6-34b"
+  example_model_input: {"prompt": "Describe this image in detail", "image": "data:image/png;base64,iVBORw0KGgo..."}
+resources:
+  accelerator: A100
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets: {}
+system_packages: []
diff --git a/stable-diffusion/sdxl-controlnet-depth/model/__init__.py b/llm/llava/llava-1.6-sgl/model/__init__.py
similarity index 100%
rename from stable-diffusion/sdxl-controlnet-depth/model/__init__.py
rename to llm/llava/llava-1.6-sgl/model/__init__.py
diff --git a/llava/llava-1.6-sgl/model/model.py b/llm/llava/llava-1.6-sgl/model/model.py
similarity index 100%
rename from llava/llava-1.6-sgl/model/model.py
rename to llm/llava/llava-1.6-sgl/model/model.py
diff --git a/llava/llava-1.6-sgl/requirements.txt b/llm/llava/llava-1.6-sgl/requirements.txt
similarity index 100%
rename from llava/llava-1.6-sgl/requirements.txt
rename to llm/llava/llava-1.6-sgl/requirements.txt
diff --git a/llm/llava/llava-v1.5-7b/README.md b/llm/llava/llava-v1.5-7b/README.md
new file mode 100644
index 000000000..801b5c6c4
--- /dev/null
+++ b/llm/llava/llava-v1.5-7b/README.md
@@ -0,0 +1,29 @@
+# llava-v1.5-7b
+
+Deploy llava-v1.5-7b for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/llava/llava-v1.5-7b/config.yaml b/llm/llava/llava-v1.5-7b/config.yaml
new file mode 100644
index 000000000..c942f50ed
--- /dev/null
+++ b/llm/llava/llava-v1.5-7b/config.yaml
@@ -0,0 +1,29 @@
+description: "LLaVA v1.5 7B for vision-language tasks"
+environment_variables: {}
+external_package_dirs: []
+model_name: llava-v1.5-7b
+python_version: py311
+requirements:
+- torch==2.0.1
+- torchvision==0.15.2
+- transformers==4.31.0
+- tokenizers>=0.12.1,<0.14
+- sentencepiece==0.1.99
+- shortuuid==1.0.11
+- scipy==1.11.4
+- accelerate==0.21.0
+- peft==0.4.0
+- bitsandbytes==0.41.0
+- einops==0.6.1
+- einops-exts==0.0.4
+- timm==0.6.13
+model_metadata:
+  repo_id: "liuhaotian/llava-v1.5-7b"
+  example_model_input: {"query": "Describe this image in detail", "image": "https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg"}
+resources:
+  accelerator: A10G
+  cpu: '3'
+  memory: 15Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/stable-diffusion/sdxl-controlnet/model/__init__.py b/llm/llava/llava-v1.5-7b/model/__init__.py
similarity index 100%
rename from stable-diffusion/sdxl-controlnet/model/__init__.py
rename to llm/llava/llava-v1.5-7b/model/__init__.py
diff --git a/llava/llava-v1.5-7b/model/model.py b/llm/llava/llava-v1.5-7b/model/model.py
similarity index 100%
rename from llava/llava-v1.5-7b/model/model.py
rename to llm/llava/llava-v1.5-7b/model/model.py
diff --git a/llava/llava-v1.5-7b/packages/llava/__init__.py b/llm/llava/llava-v1.5-7b/packages/llava/__init__.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/__init__.py
rename to llm/llava/llava-v1.5-7b/packages/llava/__init__.py
diff --git a/llava/llava-v1.5-7b/packages/llava/constants.py b/llm/llava/llava-v1.5-7b/packages/llava/constants.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/constants.py
rename to llm/llava/llava-v1.5-7b/packages/llava/constants.py
diff --git a/llava/llava-v1.5-7b/packages/llava/conversation.py b/llm/llava/llava-v1.5-7b/packages/llava/conversation.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/conversation.py
rename to llm/llava/llava-v1.5-7b/packages/llava/conversation.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_bench.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_bench.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_bench.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_bench.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_visual.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_visual.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_visual.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_gpt_review_visual.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_pope.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_pope.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_pope.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_pope.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4_requery.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4_requery.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4_requery.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_science_qa_gpt4_requery.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/eval_textvqa.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/eval_textvqa.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/eval_textvqa.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/eval_textvqa.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/generate_webpage_data_from_table.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/generate_webpage_data_from_table.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/generate_webpage_data_from_table.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/generate_webpage_data_from_table.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/m4c_evaluator.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/m4c_evaluator.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/m4c_evaluator.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/m4c_evaluator.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/model_qa.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/model_qa.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/model_qa.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/model_qa.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/model_vqa.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/model_vqa.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_loader.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_loader.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/model_vqa_loader.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_loader.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_mmbench.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_mmbench.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/model_vqa_mmbench.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_mmbench.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_qbench.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_qbench.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/model_vqa_qbench.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_qbench.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_science.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_science.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/model_vqa_science.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/model_vqa_science.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/qa_baseline_gpt35.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/qa_baseline_gpt35.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/qa_baseline_gpt35.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/qa_baseline_gpt35.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/run_llava.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/run_llava.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/run_llava.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/run_llava.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/summarize_gpt_review.py b/llm/llava/llava-v1.5-7b/packages/llava/eval/summarize_gpt_review.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/summarize_gpt_review.py
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/summarize_gpt_review.py
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_alpaca-13b.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_alpaca-13b.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_alpaca-13b.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_alpaca-13b.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_bard.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_bard.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_bard.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_bard.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_gpt35.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_gpt35.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_gpt35.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_gpt35.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_llama-13b.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_llama-13b.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_llama-13b.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_llama-13b.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_vicuna-13b.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_vicuna-13b.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_vicuna-13b.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/answer/answer_vicuna-13b.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/caps_boxes_coco2014_val_80.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/caps_boxes_coco2014_val_80.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/caps_boxes_coco2014_val_80.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/caps_boxes_coco2014_val_80.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/model.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/model.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/model.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/model.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/prompt.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/prompt.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/prompt.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/prompt.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/question.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/question.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/question.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/question.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_13b_v0.json b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_13b_v0.json
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_13b_v0.json
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_13b_v0.json
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_lcs_558k_sqa_12e_vicuna_v1_3_13b.json b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_lcs_558k_sqa_12e_vicuna_v1_3_13b.json
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_lcs_558k_sqa_12e_vicuna_v1_3_13b.json
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/results/test_sqa_llava_lcs_558k_sqa_12e_vicuna_v1_3_13b.json
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_alpaca-13b_vicuna-13b.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_alpaca-13b_vicuna-13b.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/review/review_alpaca-13b_vicuna-13b.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_alpaca-13b_vicuna-13b.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_bard_vicuna-13b.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_bard_vicuna-13b.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/review/review_bard_vicuna-13b.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_bard_vicuna-13b.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_gpt35_vicuna-13b.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_gpt35_vicuna-13b.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/review/review_gpt35_vicuna-13b.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_gpt35_vicuna-13b.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_llama-13b_vicuna-13b.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_llama-13b_vicuna-13b.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/review/review_llama-13b_vicuna-13b.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/review/review_llama-13b_vicuna-13b.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/reviewer.jsonl b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/reviewer.jsonl
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/reviewer.jsonl
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/reviewer.jsonl
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/table/rule.json b/llm/llava/llava-v1.5-7b/packages/llava/eval/table/rule.json
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/table/rule.json
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/table/rule.json
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/alpaca.png b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/alpaca.png
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/alpaca.png
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/alpaca.png
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/bard.jpg b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/bard.jpg
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/bard.jpg
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/bard.jpg
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/chatgpt.svg b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/chatgpt.svg
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/chatgpt.svg
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/chatgpt.svg
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/llama.jpg b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/llama.jpg
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/llama.jpg
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/llama.jpg
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/swords_FILL0_wght300_GRAD0_opsz48.svg b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/swords_FILL0_wght300_GRAD0_opsz48.svg
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/swords_FILL0_wght300_GRAD0_opsz48.svg
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/swords_FILL0_wght300_GRAD0_opsz48.svg
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/vicuna.jpeg b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/vicuna.jpeg
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/vicuna.jpeg
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/figures/vicuna.jpeg
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/index.html b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/index.html
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/index.html
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/index.html
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/script.js b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/script.js
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/script.js
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/script.js
diff --git a/llava/llava-v1.5-7b/packages/llava/eval/webpage/styles.css b/llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/styles.css
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/eval/webpage/styles.css
rename to llm/llava/llava-v1.5-7b/packages/llava/eval/webpage/styles.css
diff --git a/llava/llava-v1.5-7b/packages/llava/mm_utils.py b/llm/llava/llava-v1.5-7b/packages/llava/mm_utils.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/mm_utils.py
rename to llm/llava/llava-v1.5-7b/packages/llava/mm_utils.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/__init__.py b/llm/llava/llava-v1.5-7b/packages/llava/model/__init__.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/__init__.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/__init__.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/apply_delta.py b/llm/llava/llava-v1.5-7b/packages/llava/model/apply_delta.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/apply_delta.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/apply_delta.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/builder.py b/llm/llava/llava-v1.5-7b/packages/llava/model/builder.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/builder.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/builder.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/consolidate.py b/llm/llava/llava-v1.5-7b/packages/llava/model/consolidate.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/consolidate.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/consolidate.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/llava_llama.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/llava_llama.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/llava_llama.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/llava_llama.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/llava_mpt.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/llava_mpt.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/llava_mpt.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/llava_mpt.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/adapt_tokenizer.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/adapt_tokenizer.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/adapt_tokenizer.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/adapt_tokenizer.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/attention.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/attention.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/attention.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/attention.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/blocks.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/blocks.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/blocks.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/blocks.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/configuration_mpt.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/configuration_mpt.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/configuration_mpt.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/configuration_mpt.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/custom_embedding.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/custom_embedding.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/custom_embedding.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/custom_embedding.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/flash_attn_triton.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/flash_attn_triton.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/flash_attn_triton.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/flash_attn_triton.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/hf_prefixlm_converter.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/hf_prefixlm_converter.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/hf_prefixlm_converter.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/hf_prefixlm_converter.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/meta_init_context.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/meta_init_context.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/meta_init_context.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/meta_init_context.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/modeling_mpt.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/modeling_mpt.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/modeling_mpt.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/modeling_mpt.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/norm.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/norm.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/norm.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/norm.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/param_init_fns.py b/llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/param_init_fns.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/param_init_fns.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/language_model/mpt/param_init_fns.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/llava_arch.py b/llm/llava/llava-v1.5-7b/packages/llava/model/llava_arch.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/llava_arch.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/llava_arch.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/make_delta.py b/llm/llava/llava-v1.5-7b/packages/llava/model/make_delta.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/make_delta.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/make_delta.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/builder.py b/llm/llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/builder.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/builder.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/builder.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/clip_encoder.py b/llm/llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/clip_encoder.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/clip_encoder.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/multimodal_encoder/clip_encoder.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/multimodal_projector/builder.py b/llm/llava/llava-v1.5-7b/packages/llava/model/multimodal_projector/builder.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/multimodal_projector/builder.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/multimodal_projector/builder.py
diff --git a/llava/llava-v1.5-7b/packages/llava/model/utils.py b/llm/llava/llava-v1.5-7b/packages/llava/model/utils.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/model/utils.py
rename to llm/llava/llava-v1.5-7b/packages/llava/model/utils.py
diff --git a/stable-diffusion/sdxl-lightning/model/__init__.py b/llm/llava/llava-v1.5-7b/packages/llava/serve/__init__.py
similarity index 100%
rename from stable-diffusion/sdxl-lightning/model/__init__.py
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/__init__.py
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/cli.py b/llm/llava/llava-v1.5-7b/packages/llava/serve/cli.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/cli.py
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/cli.py
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/controller.py b/llm/llava/llava-v1.5-7b/packages/llava/serve/controller.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/controller.py
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/controller.py
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/examples/extreme_ironing.jpg b/llm/llava/llava-v1.5-7b/packages/llava/serve/examples/extreme_ironing.jpg
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/examples/extreme_ironing.jpg
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/examples/extreme_ironing.jpg
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/examples/waterview.jpg b/llm/llava/llava-v1.5-7b/packages/llava/serve/examples/waterview.jpg
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/examples/waterview.jpg
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/examples/waterview.jpg
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/gradio_web_server.py b/llm/llava/llava-v1.5-7b/packages/llava/serve/gradio_web_server.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/gradio_web_server.py
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/gradio_web_server.py
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/model_worker.py b/llm/llava/llava-v1.5-7b/packages/llava/serve/model_worker.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/model_worker.py
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/model_worker.py
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/register_worker.py b/llm/llava/llava-v1.5-7b/packages/llava/serve/register_worker.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/register_worker.py
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/register_worker.py
diff --git a/llava/llava-v1.5-7b/packages/llava/serve/test_message.py b/llm/llava/llava-v1.5-7b/packages/llava/serve/test_message.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/serve/test_message.py
rename to llm/llava/llava-v1.5-7b/packages/llava/serve/test_message.py
diff --git a/llava/llava-v1.5-7b/packages/llava/train/llama_flash_attn_monkey_patch.py b/llm/llava/llava-v1.5-7b/packages/llava/train/llama_flash_attn_monkey_patch.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/train/llama_flash_attn_monkey_patch.py
rename to llm/llava/llava-v1.5-7b/packages/llava/train/llama_flash_attn_monkey_patch.py
diff --git a/llava/llava-v1.5-7b/packages/llava/train/llama_xformers_attn_monkey_patch.py b/llm/llava/llava-v1.5-7b/packages/llava/train/llama_xformers_attn_monkey_patch.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/train/llama_xformers_attn_monkey_patch.py
rename to llm/llava/llava-v1.5-7b/packages/llava/train/llama_xformers_attn_monkey_patch.py
diff --git a/llava/llava-v1.5-7b/packages/llava/train/llava_trainer.py b/llm/llava/llava-v1.5-7b/packages/llava/train/llava_trainer.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/train/llava_trainer.py
rename to llm/llava/llava-v1.5-7b/packages/llava/train/llava_trainer.py
diff --git a/llava/llava-v1.5-7b/packages/llava/train/train.py b/llm/llava/llava-v1.5-7b/packages/llava/train/train.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/train/train.py
rename to llm/llava/llava-v1.5-7b/packages/llava/train/train.py
diff --git a/llava/llava-v1.5-7b/packages/llava/train/train_mem.py b/llm/llava/llava-v1.5-7b/packages/llava/train/train_mem.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/train/train_mem.py
rename to llm/llava/llava-v1.5-7b/packages/llava/train/train_mem.py
diff --git a/llava/llava-v1.5-7b/packages/llava/train/train_xformers.py b/llm/llava/llava-v1.5-7b/packages/llava/train/train_xformers.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/train/train_xformers.py
rename to llm/llava/llava-v1.5-7b/packages/llava/train/train_xformers.py
diff --git a/llava/llava-v1.5-7b/packages/llava/utils.py b/llm/llava/llava-v1.5-7b/packages/llava/utils.py
similarity index 100%
rename from llava/llava-v1.5-7b/packages/llava/utils.py
rename to llm/llava/llava-v1.5-7b/packages/llava/utils.py
diff --git a/llm/llava/llava-v1.6-34b/README.md b/llm/llava/llava-v1.6-34b/README.md
new file mode 100644
index 000000000..833f6dccc
--- /dev/null
+++ b/llm/llava/llava-v1.6-34b/README.md
@@ -0,0 +1,29 @@
+# llava-v1.6-34b
+
+Deploy llava-v1.6-34b for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/llava/llava-v1.6-34b/config.yaml b/llm/llava/llava-v1.6-34b/config.yaml
new file mode 100644
index 000000000..705068009
--- /dev/null
+++ b/llm/llava/llava-v1.6-34b/config.yaml
@@ -0,0 +1,15 @@
+description: "LLaVA v1.6 34B for vision-language tasks"
+environment_variables: {}
+external_package_dirs: []
+model_name: llava-v1.6-34b
+python_version: py311
+requirements:
+- git+https://github.com/haotian-liu/LLaVA.git
+model_metadata:
+  repo_id: "liuhaotian/llava-v1.6-34b"
+  example_model_input: {"query": "Describe this image in detail", "image": "https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg"}
+resources:
+  accelerator: A100
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/llava/llava-v1.6-34b/input.json b/llm/llava/llava-v1.6-34b/input.json
similarity index 100%
rename from llava/llava-v1.6-34b/input.json
rename to llm/llava/llava-v1.6-34b/input.json
diff --git a/stable-diffusion/sdxl-lora-swapping/model/__init__.py b/llm/llava/llava-v1.6-34b/model/__init__.py
similarity index 100%
rename from stable-diffusion/sdxl-lora-swapping/model/__init__.py
rename to llm/llava/llava-v1.6-34b/model/__init__.py
diff --git a/llava/llava-v1.6-34b/model/model.py b/llm/llava/llava-v1.6-34b/model/model.py
similarity index 100%
rename from llava/llava-v1.6-34b/model/model.py
rename to llm/llava/llava-v1.6-34b/model/model.py
diff --git a/lora/README.md b/llm/lora/README.md
similarity index 100%
rename from lora/README.md
rename to llm/lora/README.md
diff --git a/llm/lora/engine-lora/README.md b/llm/lora/engine-lora/README.md
new file mode 100644
index 000000000..47d51e546
--- /dev/null
+++ b/llm/lora/engine-lora/README.md
@@ -0,0 +1,63 @@
+# Mistral 7B Instruct Engine Lora
+
+Deploy [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mistral-7B-Instruct-v0.3",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mistral-7B-Instruct-v0.3", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/lora/engine-lora/config.yaml b/llm/lora/engine-lora/config.yaml
new file mode 100644
index 000000000..82f10314d
--- /dev/null
+++ b/llm/lora/engine-lora/config.yaml
@@ -0,0 +1,57 @@
+description: "Mistral 7B with LoRA adapters via TRT-LLM engine"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      model: "finance",
+      messages:
+        [{ role: "user", content: "How would you choose back in 2008?" }],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: mistralai/Mistral-7B-Instruct-v0.3
+model_name: Mistral 7B Instruct Engine Lora
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: mistralai/Mistral-7B-Instruct-v0.3
+      source: HF
+    lora_adapters:
+      legal:
+        source: HF
+        repo: Aretoss/Lexgen
+      finance:
+        source: HF
+        repo: vaibhav1/lora-mistral-finance
+      medical:
+        source: HF
+        repo: Imsachinsingh00/Fine_tuned_LoRA_Mistral_MTSDialog_Summarization
+    max_seq_len: 32768
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
+    served_model_name: mistral
diff --git a/llm/lora/sglang-lora/README.md b/llm/lora/sglang-lora/README.md
new file mode 100644
index 000000000..b6297dfed
--- /dev/null
+++ b/llm/lora/sglang-lora/README.md
@@ -0,0 +1,45 @@
+# Mistral-7B-Instruct SGLang Lora
+
+Deploy [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100_40GB |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/generate \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "text": [
+    "What would you choose in 2008?",
+    "What would you choose in 2008?"
+  ],
+  "sampling_params": {
+    "max_new_tokens": 1000,
+    "temperature": 1.0
+  },
+  "lora_path": [
+    "legal",
+    "finance"
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.9.post6-cu126`
+- Predict concurrency: **32**
diff --git a/llm/lora/sglang-lora/config.yaml b/llm/lora/sglang-lora/config.yaml
new file mode 100644
index 000000000..ffedd5ce1
--- /dev/null
+++ b/llm/lora/sglang-lora/config.yaml
@@ -0,0 +1,31 @@
+description: "Mistral 7B with LoRA adapters via SGLang"
+base_image:
+  image: lmsysorg/sglang:v0.4.9.post6-cu126
+model_metadata:
+  example_model_input: {
+    "text": [
+        "What would you choose in 2008?",
+        "What would you choose in 2008?",
+    ],
+    "sampling_params": {"max_new_tokens": 1000, "temperature": 1.0},
+    "lora_path": ["legal", "finance"],
+  }
+  repo_id: mistralai/Mistral-7B-Instruct-v0.3
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path mistralai/Mistral-7B-Instruct-v0.3 --port 8000 --trust-remote-code --enable-lora --lora-paths legal=Aretoss/Lexgen finance=vaibhav1/lora-mistral-finance medical=Imsachinsingh00/Fine_tuned_LoRA_Mistral_MTSDialog_Summarization --disable-radix-cache"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /generate
+  server_port: 8000
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 24Gi
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: Mistral-7B-Instruct SGLang Lora
+environment_variables:
+  hf_access_token: null
+requirements:
+  - protobuf==4.25.1
diff --git a/llm/lora/vllm-lora/README.md b/llm/lora/vllm-lora/README.md
new file mode 100644
index 000000000..326166da2
--- /dev/null
+++ b/llm/lora/vllm-lora/README.md
@@ -0,0 +1,44 @@
+# Mistral-7B-Instruct VLLM Lora
+
+Deploy [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100_40GB |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "model": "finance",
+  "messages": [
+    {
+      "role": "user",
+      "content": "How would you choose back in 2008?"
+    }
+  ],
+  "stream": true,
+  "max_tokens": 512,
+  "temperature": 0.9
+}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.9.2`
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/lora/vllm-lora/config.yaml b/llm/lora/vllm-lora/config.yaml
new file mode 100644
index 000000000..be1b786db
--- /dev/null
+++ b/llm/lora/vllm-lora/config.yaml
@@ -0,0 +1,33 @@
+description: "Mistral 7B with LoRA adapters via vLLM"
+base_image:
+  image: vllm/vllm-openai:v0.9.2
+model_metadata:
+  example_model_input: {
+    model: "finance",
+    messages: [
+      {
+        role: "user",
+        content: "How would you choose back in 2008?"
+      }
+    ],
+    stream: true,
+    max_tokens: 512,
+    temperature: 0.9
+  }
+  repo_id: mistralai/Mistral-7B-Instruct-v0.3
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Mistral-7B-Instruct-v0.3 --tokenizer_mode mistral --config_format mistral --load_format mistral --served-model-name mistral --max-model-len 16384 --port 8000 --gpu-memory-utilization 0.90 --disable-custom-all-reduce --trust-remote-code --enable-lora --lora-modules finance=vaibhav1/lora-mistral-finance legal=Aretoss/Lexgen"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 24Gi
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+model_name: Mistral-7B-Instruct VLLM Lora
+environment_variables:
+  hf_access_token: null
diff --git a/llm/midnight/README.md b/llm/midnight/README.md
new file mode 100644
index 000000000..1444a676a
--- /dev/null
+++ b/llm/midnight/README.md
@@ -0,0 +1,34 @@
+# Kaiko Midnight
+
+Pathology foundation model for medical image analysis and classification
+
+| Property | Value |
+|----------|-------|
+| Task | Embeddings |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "image_url": "https://upload.wikimedia.org/wikipedia/commons/8/80/Breast_DCIS_histopathology_%281%29.jpg",
+  "task": "classification",
+  "batch_size": 1
+}'
+```
+
+## Configuration highlights
+
+- Base image: `nvcr.io/nvidia/pytorch:25.06-py3`
+- Predict concurrency: **32**
diff --git a/llm/midnight/config.yaml b/llm/midnight/config.yaml
new file mode 100644
index 000000000..f4ae65273
--- /dev/null
+++ b/llm/midnight/config.yaml
@@ -0,0 +1,24 @@
+model_name: Kaiko Midnight
+description: Pathology foundation model for medical image analysis and classification
+python_version: py39
+base_image:
+  image: nvcr.io/nvidia/pytorch:25.06-py3
+requirements_file: ./requirements.txt
+resources:
+  accelerator: T4
+  use_gpu: true
+  memory: 1Gi
+  cpu: '1'
+runtime:
+  predict_concurrency: 32
+model_metadata:
+  repo_id: "kaiko-ai/midnight"
+  example_model_input:
+    image_url: "https://upload.wikimedia.org/wikipedia/commons/8/80/Breast_DCIS_histopathology_%281%29.jpg"
+    task: "classification"  # or "segmentation"
+    batch_size: 1
+  tags:
+    - medical-imaging
+    - pathology
+    - computer-vision
+    - embeddings
diff --git a/stable-diffusion/sdxl-lora/model/__init__.py b/llm/midnight/model/__init__.py
similarity index 100%
rename from stable-diffusion/sdxl-lora/model/__init__.py
rename to llm/midnight/model/__init__.py
diff --git a/midnight/model/model.py b/llm/midnight/model/model.py
similarity index 100%
rename from midnight/model/model.py
rename to llm/midnight/model/model.py
diff --git a/midnight/requirements.txt b/llm/midnight/requirements.txt
similarity index 100%
rename from midnight/requirements.txt
rename to llm/midnight/requirements.txt
diff --git a/llm/minimax/minimax-m2-1/README.md b/llm/minimax/minimax-m2-1/README.md
new file mode 100644
index 000000000..cbe54b415
--- /dev/null
+++ b/llm/minimax/minimax-m2-1/README.md
@@ -0,0 +1,56 @@
+# minimax
+
+Deploy [MiniMaxAI/MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [MiniMaxAI/MiniMax-M2.1](https://huggingface.co/MiniMaxAI/MiniMax-M2.1) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="MiniMaxAI/MiniMax-M2.1",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "MiniMaxAI/MiniMax-M2.1", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:nightly-dev-20260126-48f4340b`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/minimax/minimax-m2-1/config.yaml b/llm/minimax/minimax-m2-1/config.yaml
new file mode 100644
index 000000000..cd4719ee9
--- /dev/null
+++ b/llm/minimax/minimax-m2-1/config.yaml
@@ -0,0 +1,52 @@
+description: "MiniMaxAI/MiniMax-M2.1 for text generation"
+base_image:
+  image: lmsysorg/sglang:nightly-dev-20260126-48f4340b
+docker_server:
+  liveness_endpoint: /health_generate
+  predict_endpoint: /v1/chat/completions
+  readiness_endpoint: /health_generate
+  server_port: 8000
+  start_command: sh -c "truss-transfer-cli && find /app/model_cache/checkpoint -type f -print0 | xargs -0 -P 0 -I {} dd if={} of=/dev/null bs=4M && python3 -m sglang.launch_server --model-path /app/model_cache/checkpoint --tp-size 8 --ep-size 8 --tool-call-parser minimax-m2 --trust-remote-code --host 0.0.0.0 --reasoning-parser minimax --port 8000 --mem-fraction-static 0.85"
+#environment_variables:
+#  SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN: 1
+model_cache:
+- allow_patterns:
+  - '*.json'
+  - '*.safetensors'
+  - '*.txt'
+  - '*.model'
+  - '*.py'
+  - '*.jinja'
+  repo_id: MiniMaxAI/MiniMax-M2.1
+  revision: 927ea2b64008fe4a1e31a4e107a6b75916b3b44a
+  use_volume: true
+  volume_folder: checkpoint
+model_metadata:
+  example_model_input:
+    max_tokens: 4096
+    messages:
+    - content: You are a helpful assistant.
+      role: system
+    - content: Who won the world series in 2020?
+      role: user
+    model: MiniMaxAI/MiniMax-M2
+    stream: true
+    temperature: 0.6
+  model_name: MiniMax-M2
+  tags:
+  - openai-compatible
+model_name: minimax
+resources:
+  accelerator: H100:8
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: true
+runtime:
+  health_checks:
+    restart_check_delay_seconds: 1200
+    restart_threshold_seconds: 600
+    stop_traffic_threshold_seconds: 1800
+  is_websocket_endpoint: false
+  predict_concurrency: 32
+  transport:
+    kind: http
diff --git a/mistral/engine-mistral-7b-instruct/README.md b/llm/mistral/_archive/engine-mistral-7b-instruct/README.md
similarity index 100%
rename from mistral/engine-mistral-7b-instruct/README.md
rename to llm/mistral/_archive/engine-mistral-7b-instruct/README.md
diff --git a/mistral/engine-mistral-7b-instruct/config.yaml b/llm/mistral/_archive/engine-mistral-7b-instruct/config.yaml
similarity index 100%
rename from mistral/engine-mistral-7b-instruct/config.yaml
rename to llm/mistral/_archive/engine-mistral-7b-instruct/config.yaml
diff --git a/mistral/mistral-7b-chat/README.md b/llm/mistral/_archive/mistral-7b-chat/README.md
similarity index 100%
rename from mistral/mistral-7b-chat/README.md
rename to llm/mistral/_archive/mistral-7b-chat/README.md
diff --git a/mistral/mistral-7b-chat/config.yaml b/llm/mistral/_archive/mistral-7b-chat/config.yaml
similarity index 100%
rename from mistral/mistral-7b-chat/config.yaml
rename to llm/mistral/_archive/mistral-7b-chat/config.yaml
diff --git a/stable-diffusion/sdxl-turbo/model/__init__.py b/llm/mistral/_archive/mistral-7b-chat/model/__init__.py
similarity index 100%
rename from stable-diffusion/sdxl-turbo/model/__init__.py
rename to llm/mistral/_archive/mistral-7b-chat/model/__init__.py
diff --git a/mistral/mistral-7b-chat/model/model.py b/llm/mistral/_archive/mistral-7b-chat/model/model.py
similarity index 100%
rename from mistral/mistral-7b-chat/model/model.py
rename to llm/mistral/_archive/mistral-7b-chat/model/model.py
diff --git a/mistral/mistral-7b-instruct-vllm/README.md b/llm/mistral/_archive/mistral-7b-instruct-vllm/README.md
similarity index 100%
rename from mistral/mistral-7b-instruct-vllm/README.md
rename to llm/mistral/_archive/mistral-7b-instruct-vllm/README.md
diff --git a/mistral/mistral-7b-instruct-vllm/config.yaml b/llm/mistral/_archive/mistral-7b-instruct-vllm/config.yaml
similarity index 100%
rename from mistral/mistral-7b-instruct-vllm/config.yaml
rename to llm/mistral/_archive/mistral-7b-instruct-vllm/config.yaml
diff --git a/stable-diffusion/stable-diffusion-3-medium/model/__init__.py b/llm/mistral/_archive/mistral-7b-instruct-vllm/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-3-medium/model/__init__.py
rename to llm/mistral/_archive/mistral-7b-instruct-vllm/model/__init__.py
diff --git a/mistral/mistral-7b-instruct-vllm/model/model.py b/llm/mistral/_archive/mistral-7b-instruct-vllm/model/model.py
similarity index 100%
rename from mistral/mistral-7b-instruct-vllm/model/model.py
rename to llm/mistral/_archive/mistral-7b-instruct-vllm/model/model.py
diff --git a/mistral/mistral-7b-instruct/README.md b/llm/mistral/_archive/mistral-7b-instruct/README.md
similarity index 100%
rename from mistral/mistral-7b-instruct/README.md
rename to llm/mistral/_archive/mistral-7b-instruct/README.md
diff --git a/mistral/mistral-7b-instruct/config.yaml b/llm/mistral/_archive/mistral-7b-instruct/config.yaml
similarity index 100%
rename from mistral/mistral-7b-instruct/config.yaml
rename to llm/mistral/_archive/mistral-7b-instruct/config.yaml
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/model/__init__.py b/llm/mistral/_archive/mistral-7b-instruct/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-inpainting-trt/model/__init__.py
rename to llm/mistral/_archive/mistral-7b-instruct/model/__init__.py
diff --git a/mistral/mistral-7b-instruct/model/model.py b/llm/mistral/_archive/mistral-7b-instruct/model/model.py
similarity index 100%
rename from mistral/mistral-7b-instruct/model/model.py
rename to llm/mistral/_archive/mistral-7b-instruct/model/model.py
diff --git a/llm/mistral/engine-devstral/README.md b/llm/mistral/engine-devstral/README.md
new file mode 100644
index 000000000..9b9a4dfe1
--- /dev/null
+++ b/llm/mistral/engine-devstral/README.md
@@ -0,0 +1,60 @@
+# Devstral Small 2505
+
+Deploy [mistralai/Devstral-Small-2505](https://huggingface.co/mistralai/Devstral-Small-2505) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Devstral-Small-2505](https://huggingface.co/mistralai/Devstral-Small-2505) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Devstral-Small-2505",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Devstral-Small-2505", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/llm/mistral/engine-devstral/config.yaml b/llm/mistral/engine-devstral/config.yaml
new file mode 100644
index 000000000..e8596e3e0
--- /dev/null
+++ b/llm/mistral/engine-devstral/config.yaml
@@ -0,0 +1,50 @@
+description: "Devstral Small for code generation"
+model_metadata:
+  example_model_input: {
+    messages: [
+      {
+        role: "system",
+        content: "<Load system prompt from Hugging Face>"
+      },
+      {
+        role: "user",
+        content: "<your-command>"
+      }
+    ],
+    stream: true,
+    max_tokens: 512,
+    temperature: 0.15,
+    top_p: 1.0,
+    top_k: 40,
+    frequency_penalty: 1
+  }
+  tags:
+    - openai-compatible
+model_name: Devstral Small 2505
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    checkpoint_repository:
+      repo: mistralai/Devstral-Small-2505
+      revision: "refs/pr/8"
+      source: HF
+    num_builder_gpus: 2
+    max_batch_size: 64
+    max_seq_len: 131072
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+    speculator: # optional: use speculative decoding
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 8
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+  runtime:
+    enable_chunked_context: true
diff --git a/llm/mistral/engine-mistral-small-3/README.md b/llm/mistral/engine-mistral-small-3/README.md
new file mode 100644
index 000000000..a11c49f08
--- /dev/null
+++ b/llm/mistral/engine-mistral-small-3/README.md
@@ -0,0 +1,64 @@
+# Mistral Small 3 Instruct FP8
+
+Deploy [mistralai/Mistral-Small-24B-Instruct-2501](https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mistral-Small-24B-Instruct-2501](https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mistral-Small-24B-Instruct-2501",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mistral-Small-24B-Instruct-2501", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **use_fp8_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/mistral/engine-mistral-small-3/config.yaml b/llm/mistral/engine-mistral-small-3/config.yaml
new file mode 100644
index 000000000..1ee570079
--- /dev/null
+++ b/llm/mistral/engine-mistral-small-3/config.yaml
@@ -0,0 +1,54 @@
+description: "mistralai/Mistral-Small-24B-Instruct-2501 for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are a knowledgable, engaging, meteorology teacher.",
+          },
+          {
+            role: "user",
+            content: "What is the impact of the Mistral wind on the French climate?",
+          },
+        ],
+      stream: true,
+      max_tokens: 1024,
+      temperature: 0.15,
+    }
+  repo_id: mistralai/Mistral-Small-24B-Instruct-2501
+model_name: Mistral Small 3 Instruct FP8
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: mistralai/Mistral-Small-24B-Instruct-2501
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: fp8_kv
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: true
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/mistral/engine-mixtral-8x22b-instruct/README.md b/llm/mistral/engine-mixtral-8x22b-instruct/README.md
new file mode 100644
index 000000000..e7dc3f019
--- /dev/null
+++ b/llm/mistral/engine-mixtral-8x22b-instruct/README.md
@@ -0,0 +1,60 @@
+# Mistral 8x22B Instruct
+
+Deploy [mistralai/Mixtral-8x22B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mixtral-8x22B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mixtral-8x22B-Instruct-v0.1",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mixtral-8x22B-Instruct-v0.1", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **8,192**
+- Streaming: **enabled**
diff --git a/llm/mistral/engine-mixtral-8x22b-instruct/config.yaml b/llm/mistral/engine-mixtral-8x22b-instruct/config.yaml
new file mode 100644
index 000000000..c568025cc
--- /dev/null
+++ b/llm/mistral/engine-mixtral-8x22b-instruct/config.yaml
@@ -0,0 +1,46 @@
+description: "mistralai/Mixtral-8x22B-Instruct-v0.1 for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are a knowledgable, engaging, geology teacher.",
+          },
+          {
+            role: "user",
+            content: "What is the impact of the Mistral wind on the French climate?",
+          },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: mistralai/Mixtral-8x22B-Instruct-v0.1
+model_name: Mistral 8x22B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:2
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: mistralai/Mixtral-8x22B-Instruct-v0.1
+      source: HF
+    max_seq_len: 8192
+    num_builder_gpus: 4
+    quantization_type: fp8_kv
+    tensor_parallel_count: 2
diff --git a/llm/mistral/engine-mixtral-8x7b-instruct/README.md b/llm/mistral/engine-mixtral-8x7b-instruct/README.md
new file mode 100644
index 000000000..a71ce490e
--- /dev/null
+++ b/llm/mistral/engine-mixtral-8x7b-instruct/README.md
@@ -0,0 +1,64 @@
+# Mistral 8x7B Instruct
+
+Deploy [mistralai/Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mixtral-8x7B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-Instruct-v0.1) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mixtral-8x7B-Instruct-v0.1",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mixtral-8x7B-Instruct-v0.1", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **use_fp8_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/mistral/engine-mixtral-8x7b-instruct/config.yaml b/llm/mistral/engine-mixtral-8x7b-instruct/config.yaml
new file mode 100644
index 000000000..5e8892002
--- /dev/null
+++ b/llm/mistral/engine-mixtral-8x7b-instruct/config.yaml
@@ -0,0 +1,54 @@
+description: "mistralai/Mixtral-8x7B-Instruct-v0.1 for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are a knowledgable, engaging, meteorology teacher.",
+          },
+          {
+            role: "user",
+            content: "What is the impact of the Mistral wind on the French climate?",
+          },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: mistralai/Mixtral-8x7B-Instruct-v0.1
+model_name: Mistral 8x7B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets:
+  hf_access_token: set token in baseten workspace
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: mistralai/Mixtral-8x7B-Instruct-v0.1
+      source: HF
+    num_builder_gpus: 2
+    quantization_type: fp8_kv
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: true
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/mistral/mistral-7b/README.md b/llm/mistral/mistral-7b/README.md
new file mode 100644
index 000000000..aa0ad9156
--- /dev/null
+++ b/llm/mistral/mistral-7b/README.md
@@ -0,0 +1,33 @@
+# mistral-7b
+
+Deploy mistral-7b for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "What is the Mistral wind?"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/mistral/mistral-7b/config.yaml b/llm/mistral/mistral-7b/config.yaml
new file mode 100644
index 000000000..c4291badf
--- /dev/null
+++ b/llm/mistral/mistral-7b/config.yaml
@@ -0,0 +1,27 @@
+description: "mistral-7b for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "mistralai/Mistral-7B-v0.1"
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  example_model_input:
+    prompt: What is the Mistral wind?
+  pretty_name: Mistral 7B
+  tags:
+  - text-generation
+model_name: mistral-7b
+python_version: py311
+requirements:
+- transformers==4.42.3
+- sentencepiece==0.1.99
+- accelerate==0.25.0
+- torch==2.0.1
+- numpy==1.26.4
+resources:
+  accelerator: A10G
+  memory: 25Gi
+  use_gpu: true
+secrets:
+  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
+system_packages: []
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/__init__.py b/llm/mistral/mistral-7b/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt-h100/model/__init__.py
rename to llm/mistral/mistral-7b/model/__init__.py
diff --git a/mistral/mistral-7b/model/model.py b/llm/mistral/mistral-7b/model/model.py
similarity index 100%
rename from mistral/mistral-7b/model/model.py
rename to llm/mistral/mistral-7b/model/model.py
diff --git a/llm/mistral/mistral-small-3.1/README.md b/llm/mistral/mistral-small-3.1/README.md
new file mode 100644
index 000000000..ff9cd21f3
--- /dev/null
+++ b/llm/mistral/mistral-small-3.1/README.md
@@ -0,0 +1,58 @@
+# Mistral Small 3.1
+
+Deploy [mistralai/Mistral-Small-3.1-24B-Instruct-2503](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mistral-Small-3.1-24B-Instruct-2503](https://huggingface.co/mistralai/Mistral-Small-3.1-24B-Instruct-2503) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mistral-Small-3.1-24B-Instruct-2503",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mistral-Small-3.1-24B-Instruct-2503", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.10.1.1`
+- Predict concurrency: **8**
+- Streaming: **enabled**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/mistral/mistral-small-3.1/config.yaml b/llm/mistral/mistral-small-3.1/config.yaml
new file mode 100644
index 000000000..d3dc77068
--- /dev/null
+++ b/llm/mistral/mistral-small-3.1/config.yaml
@@ -0,0 +1,57 @@
+#vllm serve mistralai/Mistral-Small-3.1-24B-Instruct-2503 --tokenizer_mode mistral --config_format mistral --load_format mistral --tool-call-parser mistral --enable-auto-tool-choice --limit_mm_per_prompt 'image=10' --tensor-parallel-size 2
+description: "mistralai/Mistral-Small-3.1-24B-Instruct-2503 for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.10.1.1
+model_metadata:
+  repo_id: mistralai/Mistral-Small-3.1-24B-Instruct-2503
+  example_model_input: {
+    "model": "mistral",
+    "messages": [
+      {
+      "role": "user",
+      "content": [
+        {
+        "type": "text",
+        "text": "Describe this image in one sentence."
+        },
+        {
+        "type": "image_url",
+        "image_url": {
+          "url": "https://picsum.photos/id/237/200/300"
+        }
+        }
+      ]
+      }
+    ],
+    "stream": true,
+    "max_tokens": 512,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+docker_server:
+  start_command: "sh -c \"b10-compile-cache & VLLM_USE_V1=1 HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Mistral-Small-3.1-24B-Instruct-2503 --tokenizer_mode mistral --config_format mistral --load_format mistral --tool-call-parser mistral --enable-auto-tool-choice --served-model-name mistral --max-num-seqs 8 --max-model-len 16384 --tensor-parallel-size 1 --gpu-memory-utilization 0.95\""
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+environment_variables:
+  VLLM_LOGGING_LEVEL: INFO
+  hf_access_token: null
+requirements:
+- huggingface_hub==0.19.4
+- hf_transfer==0.1.4
+- datasets==2.16.1
+- b10-transfer==0.0.5
+resources:
+  accelerator: H100:1
+  use_gpu: true
+secrets:
+  hf_access_token: null
+runtime:
+  health_checks:
+    restart_check_delay_seconds: 300      # Waits 5 minutes after deployment before starting health checks
+    restart_threshold_seconds: 300       # Triggers a restart if health checks fail for 5 minutes
+    stop_traffic_threshold_seconds: 120  # Stops traffic if health checks fail for 2 minutes
+  predict_concurrency : 8
+model_name: Mistral Small 3.1
diff --git a/llm/mistral/mixtral-8x22b-trt-int8-weights-only/README.md b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/README.md
new file mode 100644
index 000000000..73d3ad4cd
--- /dev/null
+++ b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/README.md
@@ -0,0 +1,57 @@
+# Mixtral 8x22B Instruct TRT-LLM Weights Only Quantized
+
+Mixtral 8x22B Instruct, with INT8 weights only quantization, optimized with TRT-LLM!
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100:4 |
+| Quantization | INT8 |
+| OpenAI compatible | Yes |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="model",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "model", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `docker.io/baseten/triton_trt_llm:4062d46_20240401`
+- Predict concurrency: **256**
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/TRT-LLM-README.md b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/TRT-LLM-README.md
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/TRT-LLM-README.md
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/TRT-LLM-README.md
diff --git a/llm/mistral/mixtral-8x22b-trt-int8-weights-only/config.yaml b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/config.yaml
new file mode 100644
index 000000000..55e61dc74
--- /dev/null
+++ b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/config.yaml
@@ -0,0 +1,42 @@
+base_image:
+  image: docker.io/baseten/triton_trt_llm:4062d46_20240401
+  python_executable_path: /usr/bin/python3
+description: Mixtral 8x22B Instruct, with INT8 weights only quantization, optimized
+  with TRT-LLM!
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "mistralai/Mixtral-8x22B-Instruct-v0.1"
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  engine_repository: baseten/mixtral-8x22B_i60000_o4000_bs2_tp4_int8_weights_only_A100-tllm_0.9.0.dev2024022000
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: What is your favourite condiment?
+      role: user
+    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
+        adds just the right amount of zesty flavour to whatever I'm cooking up in
+        the kitchen!
+      role: assistant
+    - content: Do you have mayonnaise recipes?
+      role: user
+  tags:
+  - text-generation
+  - openai-compatible
+  tensor_parallelism: 4
+  tokenizer_repository: mistralai/Mixtral-8x22B-Instruct-v0.1
+model_name: Mixtral 8x22B Instruct TRT-LLM Weights Only Quantized
+python_version: py311
+requirements:
+- tritonclient[all]==2.42.0
+- transformers==4.42.3
+resources:
+  accelerator: A100:4
+  use_gpu: true
+runtime:
+  num_workers: 1
+  predict_concurrency: 256
+secrets:
+  hf_access_token: "your-hf-access-token"
+system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/data/.gitattributes b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/data/.gitattributes
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/data/.gitattributes
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/data/.gitattributes
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt/model/__init__.py b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0-trt/model/__init__.py
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/model/__init__.py
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/model/model.py b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/model/model.py
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/model/model.py
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/model/model.py
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/client.py b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/client.py
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/client.py
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/client.py
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/ensemble/config.pbtxt b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/ensemble/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/ensemble/config.pbtxt
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/ensemble/config.pbtxt
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/1/model.py b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/1/model.py
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/1/model.py
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/config.pbtxt b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/postprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/1/model.py b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/1/model.py
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/1/model.py
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/config.pbtxt b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/preprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/utils.py b/llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/utils.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/utils.py
rename to llm/mistral/mixtral-8x22b-trt-int8-weights-only/packages/utils.py
diff --git a/llm/mistral/mixtral-8x22b/README.md b/llm/mistral/mixtral-8x22b/README.md
new file mode 100644
index 000000000..0595821b0
--- /dev/null
+++ b/llm/mistral/mixtral-8x22b/README.md
@@ -0,0 +1,34 @@
+# Mixtral 8x22
+
+Deploy [mistralai/Mixtral-8x22B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mixtral-8x22B-Instruct-v0.1](https://huggingface.co/mistralai/Mixtral-8x22B-Instruct-v0.1) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100:4 |
+| Python | py310 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "What is the Mistral wind?"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/mistral/mixtral-8x22b/config.yaml b/llm/mistral/mixtral-8x22b/config.yaml
new file mode 100644
index 000000000..15307ea68
--- /dev/null
+++ b/llm/mistral/mixtral-8x22b/config.yaml
@@ -0,0 +1,24 @@
+description: "mistralai/Mixtral-8x22B-Instruct-v0.1 for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: mistralai/Mixtral-8x22B-Instruct-v0.1
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  example_model_input:
+    prompt: What is the Mistral wind?
+  pretty_name: Mistral 8x22B
+  tags:
+  - text-generation
+model_name: Mixtral 8x22
+python_version: py310
+requirements:
+  - accelerate==0.25.0
+  - transformers==4.42.3
+  - torch==2.2.0
+resources:
+  accelerator: A100:4
+  use_gpu: true
+secrets:
+  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
+system_packages: []
diff --git a/stable-diffusion/stable-diffusion-xl-1.0/model/__init__.py b/llm/mistral/mixtral-8x22b/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-diffusion-xl-1.0/model/__init__.py
rename to llm/mistral/mixtral-8x22b/model/__init__.py
diff --git a/mistral/mixtral-8x22b/model/model.py b/llm/mistral/mixtral-8x22b/model/model.py
similarity index 100%
rename from mistral/mixtral-8x22b/model/model.py
rename to llm/mistral/mixtral-8x22b/model/model.py
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/README.md
new file mode 100644
index 000000000..fd159da11
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/README.md
@@ -0,0 +1,55 @@
+# Mixtral 8x7B Instruct TRT-LLM for H100
+
+Mixtral 8x7B Instruct optimized with TRT-LLM! Compatible with OpenAI Client
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="model",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "model", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `docker.io/baseten/trtllm-server:r23.12_baseten_v0.7.1`
+- Predict concurrency: **256**
+- Environment variables: `HF_HUB_ENABLE_HF_TRANSFER`
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/TRT-LLM-README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/TRT-LLM-README.md
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/TRT-LLM-README.md
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/TRT-LLM-README.md
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/config.yaml b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/config.yaml
new file mode 100644
index 000000000..838027a75
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/config.yaml
@@ -0,0 +1,43 @@
+base_image:
+  image: docker.io/baseten/trtllm-server:r23.12_baseten_v0.7.1
+  python_executable_path: /usr/bin/python3
+description: Mixtral 8x7B Instruct optimized with TRT-LLM! Compatible with OpenAI
+  Client
+environment_variables:
+  HF_HUB_ENABLE_HF_TRANSFER: 1
+external_package_dirs: []
+model_metadata:
+  repo_id: "mistralai/Mixtral-8x7B-Instruct-v0.1"
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  engine_repository: baseten/mixtral-h100-0.7.1
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: What is your favourite condiment?
+      role: user
+    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
+        adds just the right amount of zesty flavour to whatever I'm cooking up in
+        the kitchen!
+      role: assistant
+    - content: Do you have mayonnaise recipes?
+      role: user
+  tags:
+  - text-generation
+  - openai-compatible
+  tensor_parallelism: 2
+  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
+model_name: Mixtral 8x7B Instruct TRT-LLM for H100
+python_version: py311
+requirements:
+- tritonclient[all]==2.42.0
+- transformers==4.42.3
+- jinja2==3.1.3
+- hf_transfer==0.1.5
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 256
+secrets: {}
+system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/data/.gitattributes b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/data/.gitattributes
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/data/.gitattributes
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/data/.gitattributes
diff --git a/stable-diffusion/stable-diffusion/model/__init__.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-diffusion/model/__init__.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/model/__init__.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/model/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/model/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/client.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/client.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/client.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/client.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/utils.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/utils.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/utils.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-h100/packages/utils.py
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/README.md
new file mode 100644
index 000000000..2507441d0
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/README.md
@@ -0,0 +1,55 @@
+# Mixtral 8x7B Instruct TRT-LLM Weights Only Quantized for H100
+
+Mixtral 8x7B Instruct, with INT8 weights only quantization, optimized with TRT-LLM! Compatible with OpenAI Client
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="model",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "model", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `docker.io/baseten/trtllm-server:r23.12_baseten_v0.7.1`
+- Predict concurrency: **256**
+- Environment variables: `HF_HUB_ENABLE_HF_TRANSFER`
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/TRT-LLM-README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/TRT-LLM-README.md
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/TRT-LLM-README.md
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/TRT-LLM-README.md
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/config.yaml b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/config.yaml
new file mode 100644
index 000000000..2868bb999
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/config.yaml
@@ -0,0 +1,43 @@
+base_image:
+  image: docker.io/baseten/trtllm-server:r23.12_baseten_v0.7.1
+  python_executable_path: /usr/bin/python3
+description: Mixtral 8x7B Instruct, with INT8 weights only quantization, optimized
+  with TRT-LLM! Compatible with OpenAI Client
+environment_variables:
+  HF_HUB_ENABLE_HF_TRANSFER: 1
+external_package_dirs: []
+model_metadata:
+  repo_id: "mistralai/Mixtral-8x7B-Instruct-v0.1"
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  engine_repository: baseten/mixtral-weights-only-quantized-h100-0.7.1
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: What is your favourite condiment?
+      role: user
+    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
+        adds just the right amount of zesty flavour to whatever I'm cooking up in
+        the kitchen!
+      role: assistant
+    - content: Do you have mayonnaise recipes?
+      role: user
+  tags:
+  - text-generation
+  - openai-compatible
+  tensor_parallelism: 1
+  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
+model_name: Mixtral 8x7B Instruct TRT-LLM Weights Only Quantized for H100
+python_version: py311
+requirements:
+- tritonclient[all]==2.42.0
+- transformers==4.42.3
+- jinja2==3.1.3
+- hf_transfer==0.1.5
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency: 256
+secrets: {}
+system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/data/.gitattributes b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/data/.gitattributes
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/data/.gitattributes
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/data/.gitattributes
diff --git a/stable-diffusion/stable-video-diffusion/model/__init__.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/__init__.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/__init__.py
diff --git a/templates/trt-llm/model/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/model.py
similarity index 100%
rename from templates/trt-llm/model/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/model/model.py
diff --git a/templates/trt-llm/packages/client.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/client.py
similarity index 100%
rename from templates/trt-llm/packages/client.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/client.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/ensemble/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/postprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/preprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/utils.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/utils.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/utils.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/packages/utils.py
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/README.md
new file mode 100644
index 000000000..b44d0b754
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/README.md
@@ -0,0 +1,57 @@
+# Mixtral 8x7B Instruct TRT-LLM Weights Only Quantized
+
+Mixtral 8x7B Instruct, with INT8 weights only quantization, optimized with TRT-LLM! Compatible with OpenAI Client
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mixtral-8x7B-v0.1](https://huggingface.co/mistralai/Mixtral-8x7B-v0.1) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| OpenAI compatible | Yes |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mixtral-8x7B-v0.1",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mixtral-8x7B-v0.1", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `docker.io/baseten/triton_trt_llm:main-20231215`
+- Predict concurrency: **256**
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/TRT-LLM-README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/TRT-LLM-README.md
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/TRT-LLM-README.md
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/TRT-LLM-README.md
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/config.yaml b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/config.yaml
new file mode 100644
index 000000000..4fd886745
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/config.yaml
@@ -0,0 +1,42 @@
+base_image:
+  image: docker.io/baseten/triton_trt_llm:main-20231215
+  python_executable_path: /usr/bin/python3
+description: Mixtral 8x7B Instruct, with INT8 weights only quantization, optimized
+  with TRT-LLM! Compatible with OpenAI Client
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  engine_repository: baseten/mixtral-weights-only-quantized
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: What is your favourite condiment?
+      role: user
+    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
+        adds just the right amount of zesty flavour to whatever I'm cooking up in
+        the kitchen!
+      role: assistant
+    - content: Do you have mayonnaise recipes?
+      role: user
+  tags:
+  - text-generation
+  - openai-compatible
+  tensor_parallelism: 1
+  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
+  repo_id: mistralai/Mixtral-8x7B-v0.1
+model_name: Mixtral 8x7B Instruct TRT-LLM Weights Only Quantized
+python_version: py311
+requirements:
+- tritonclient[all]==2.42.0
+- transformers==4.42.3
+resources:
+  accelerator: A100
+  use_gpu: true
+runtime:
+  num_workers: 1
+  predict_concurrency: 256
+secrets:
+  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
+system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/data/.gitattributes b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/data/.gitattributes
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/data/.gitattributes
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/data/.gitattributes
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/__init__.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/__init__.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/__init__.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/model/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/client.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/client.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/client.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/client.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/ensemble/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/ensemble/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/ensemble/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/ensemble/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/postprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/1/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/preprocessing/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/utils.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/utils.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/utils.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/packages/utils.py
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm/README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm/README.md
new file mode 100644
index 000000000..9b191755b
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm/README.md
@@ -0,0 +1,54 @@
+# Mixtral 8x7B Instruct TRT-LLM
+
+Mixtral 8x7B Instruct optimized with TRT-LLM! Compatible with OpenAI Client
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100:2 |
+| OpenAI compatible | Yes |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="model",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "model", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `docker.io/baseten/triton_trt_llm:main-20231215`
+- Predict concurrency: **256**
diff --git a/templates/trt-llm/TRT-LLM-README.md b/llm/mistral/mixtral-8x7b-instruct-trt-llm/TRT-LLM-README.md
similarity index 100%
rename from templates/trt-llm/TRT-LLM-README.md
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/TRT-LLM-README.md
diff --git a/llm/mistral/mixtral-8x7b-instruct-trt-llm/config.yaml b/llm/mistral/mixtral-8x7b-instruct-trt-llm/config.yaml
new file mode 100644
index 000000000..39a343bf7
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-trt-llm/config.yaml
@@ -0,0 +1,41 @@
+base_image:
+  image: docker.io/baseten/triton_trt_llm:main-20231215
+  python_executable_path: /usr/bin/python3
+description: Mixtral 8x7B Instruct optimized with TRT-LLM! Compatible with OpenAI
+  Client
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "mistralai/Mixtral-8x7B-Instruct-v0.1"
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  engine_repository: baseten/mixtral
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: What is your favourite condiment?
+      role: user
+    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
+        adds just the right amount of zesty flavour to whatever I'm cooking up in
+        the kitchen!
+      role: assistant
+    - content: Do you have mayonnaise recipes?
+      role: user
+  tags:
+  - text-generation
+  - openai-compatible
+  tensor_parallelism: 2
+  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
+model_name: Mixtral 8x7B Instruct TRT-LLM
+python_version: py311
+requirements:
+- tritonclient[all]==2.42.0
+- transformers==4.42.3
+resources:
+  accelerator: A100:2
+  use_gpu: true
+runtime:
+  num_workers: 1
+  predict_concurrency: 256
+secrets: {}
+system_packages: []
diff --git a/templates/trt-llm/data/.gitattributes b/llm/mistral/mixtral-8x7b-instruct-trt-llm/data/.gitattributes
similarity index 100%
rename from templates/trt-llm/data/.gitattributes
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/data/.gitattributes
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/util/__init__.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/util/__init__.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/model/__init__.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/model/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm/model/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/model/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/model/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/client.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/client.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/client.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/client.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/ensemble/config.pbtxt
diff --git a/templates/trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py
similarity index 100%
rename from templates/trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/postprocessing/config.pbtxt
diff --git a/templates/trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py
similarity index 100%
rename from templates/trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/1/model.py
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/preprocessing/config.pbtxt
diff --git a/templates/trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
similarity index 100%
rename from templates/trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/inflight_batcher_llm/tensorrt_llm/config.pbtxt
diff --git a/templates/trt-llm/packages/utils.py b/llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/utils.py
similarity index 100%
rename from templates/trt-llm/packages/utils.py
rename to llm/mistral/mixtral-8x7b-instruct-trt-llm/packages/utils.py
diff --git a/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/README.md b/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/README.md
new file mode 100644
index 000000000..92667b52c
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/README.md
@@ -0,0 +1,29 @@
+# Mixtral 8x7B — VLLM TP2 — A100:2
+
+Deploy Mixtral 8x7B — VLLM TP2 — A100:2 for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100:2 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/config.yaml b/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/config.yaml
new file mode 100644
index 000000000..70305d99a
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/config.yaml
@@ -0,0 +1,17 @@
+description: "Mixtral 8x7B — VLLM TP2 — A100:2 for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_name: Mixtral 8x7B — VLLM TP2 — A100:2
+python_version: py310
+requirements:
+- vllm==0.5.3.post1
+model_metadata:
+  repo_id: "mistralai/Mixtral-8x7B-Instruct-v0.1"
+  example_model_input: {"prompt": "What is machine learning?"}
+resources:
+  accelerator: A100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets: {}
+system_packages: []
diff --git a/stable-diffusion/stable-video-diffusion/model/scripts/util/detection/__init__.py b/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/__init__.py
similarity index 100%
rename from stable-diffusion/stable-video-diffusion/model/scripts/util/detection/__init__.py
rename to llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/__init__.py
diff --git a/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/model.py b/llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/model.py
rename to llm/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/model/model.py
diff --git a/llm/mistral/mixtral-8x7b-instruct-vllm/README.md b/llm/mistral/mixtral-8x7b-instruct-vllm/README.md
new file mode 100644
index 000000000..054520085
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-vllm/README.md
@@ -0,0 +1,29 @@
+# Mixtral 8x7B
+
+Deploy Mixtral 8x7B for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100:2 |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/mistral/mixtral-8x7b-instruct-vllm/config.yaml b/llm/mistral/mixtral-8x7b-instruct-vllm/config.yaml
new file mode 100644
index 000000000..1078b91b4
--- /dev/null
+++ b/llm/mistral/mixtral-8x7b-instruct-vllm/config.yaml
@@ -0,0 +1,17 @@
+description: "Mixtral 8x7B for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_name: Mixtral 8x7B
+python_version: py310
+requirements:
+- vllm==0.2.5
+model_metadata:
+  repo_id: "mistralai/Mixtral-8x7B-Instruct-v0.1"
+  example_model_input: {"prompt": "What is machine learning?"}
+resources:
+  accelerator: A100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets: {}
+system_packages: []
diff --git a/templates/faster-whisper-truss/model/__init__.py b/llm/mistral/mixtral-8x7b-instruct-vllm/model/__init__.py
similarity index 100%
rename from templates/faster-whisper-truss/model/__init__.py
rename to llm/mistral/mixtral-8x7b-instruct-vllm/model/__init__.py
diff --git a/mistral/mixtral-8x7b-instruct-vllm/model/model.py b/llm/mistral/mixtral-8x7b-instruct-vllm/model/model.py
similarity index 100%
rename from mistral/mixtral-8x7b-instruct-vllm/model/model.py
rename to llm/mistral/mixtral-8x7b-instruct-vllm/model/model.py
diff --git a/llm/mistral/pixtral-12b/README.md b/llm/mistral/pixtral-12b/README.md
new file mode 100644
index 000000000..2e6afc2df
--- /dev/null
+++ b/llm/mistral/pixtral-12b/README.md
@@ -0,0 +1,53 @@
+# Pixtral 12B
+
+Deploy [mistral-community/pixtral-12b-240910](https://huggingface.co/mistral-community/pixtral-12b-240910) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistral-community/pixtral-12b-240910](https://huggingface.co/mistral-community/pixtral-12b-240910) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A100 |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "messages": [
+    {
+      "role": "user",
+      "content": [
+        {
+          "type": "text",
+          "text": "Describe this image in one sentence."
+        },
+        {
+          "type": "image_url",
+          "image_url": {
+            "url": "https://picsum.photos/id/237/200/300"
+          }
+        }
+      ]
+    }
+  ],
+  "stream": false,
+  "max_tokens": 512,
+  "temperature": 0.5
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/mistral/pixtral-12b/config.yaml b/llm/mistral/pixtral-12b/config.yaml
new file mode 100644
index 000000000..cc1ea8a0d
--- /dev/null
+++ b/llm/mistral/pixtral-12b/config.yaml
@@ -0,0 +1,44 @@
+description: "Pixtral 12B for vision-language tasks"
+model_metadata:
+  repo_id: mistral-community/pixtral-12b-240910
+  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
+  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
+  example_model_input: {
+    messages: [
+      {
+        role: user,
+        content: [
+          {
+            type: text,
+            text: "Describe this image in one sentence."
+          },
+          {
+            type: image_url,
+            image_url: {
+              url: "https://picsum.photos/id/237/200/300"
+            }
+          }
+        ]
+      }
+    ],
+    stream: false,
+    max_tokens: 512,
+    temperature: 0.5
+  }
+  vllm_config:
+    tensor_parallel_size: 1
+    max_model_len: 16384
+    max_num_batched_tokens: 16384
+    limit_mm_per_prompt: {image: 5}
+  tags:
+    - text-generation
+    - multimodal
+model_name: Pixtral 12B
+python_version: py311
+secrets:
+  hf_access_token: null
+requirements:
+  - vllm==0.6.1
+resources:
+  accelerator: A100
+  use_gpu: true
diff --git a/templates/transformers-openai-compatible/model/__init__.py b/llm/mistral/pixtral-12b/model/__init__.py
similarity index 100%
rename from templates/transformers-openai-compatible/model/__init__.py
rename to llm/mistral/pixtral-12b/model/__init__.py
diff --git a/mistral/pixtral-12b/model/model.py b/llm/mistral/pixtral-12b/model/model.py
similarity index 100%
rename from mistral/pixtral-12b/model/model.py
rename to llm/mistral/pixtral-12b/model/model.py
diff --git a/llm/nemotron/llama-3-1-nemotron-70b-instruct/README.md b/llm/nemotron/llama-3-1-nemotron-70b-instruct/README.md
new file mode 100644
index 000000000..0aa316d21
--- /dev/null
+++ b/llm/nemotron/llama-3-1-nemotron-70b-instruct/README.md
@@ -0,0 +1,49 @@
+# Llama-3.1-Nemotron-70B-Instruct
+
+Deploy [nvidia/Llama-3.1-Nemotron-70B-Instruct-HF](https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/Llama-3.1-Nemotron-70B-Instruct-HF](https://huggingface.co/nvidia/Llama-3.1-Nemotron-70B-Instruct-HF) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 KV |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "messages": [
+    {
+      "role": "user",
+      "content": "How many r in strawberry?"
+    }
+  ],
+  "stream": true,
+  "max_tokens": 512,
+  "temperature": 0.6
+}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **use_fp8_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/nemotron/llama-3-1-nemotron-70b-instruct/config.yaml b/llm/nemotron/llama-3-1-nemotron-70b-instruct/config.yaml
new file mode 100644
index 000000000..90953b26d
--- /dev/null
+++ b/llm/nemotron/llama-3-1-nemotron-70b-instruct/config.yaml
@@ -0,0 +1,43 @@
+description: "nvidia/Llama-3.1-Nemotron-70B-Instruct-HF for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages: [{ role: "user", content: "How many r in strawberry?" }],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.6,
+    }
+  repo_id: nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
+model_name: Llama-3.1-Nemotron-70B-Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:2
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
+      source: HF
+    num_builder_gpus: 4
+    quantization_type: fp8_kv
+    max_seq_len: 131072
+    tensor_parallel_count: 2
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: true
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 131072
diff --git a/llm/nemotron/llama-3-1-nemotron-nano-vl-8b-v1/README.md b/llm/nemotron/llama-3-1-nemotron-nano-vl-8b-v1/README.md
new file mode 100644
index 000000000..150a57989
--- /dev/null
+++ b/llm/nemotron/llama-3-1-nemotron-nano-vl-8b-v1/README.md
@@ -0,0 +1,57 @@
+# Llama 3.1 Nemotron Nano VL 8B V1
+
+Deploy [nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1](https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1](https://huggingface.co/nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+| Python | py312 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.11.0`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **16**
+- Streaming: **enabled**
diff --git a/llm/nemotron/llama-3-1-nemotron-nano-vl-8b-v1/config.yaml b/llm/nemotron/llama-3-1-nemotron-nano-vl-8b-v1/config.yaml
new file mode 100644
index 000000000..4ad38520a
--- /dev/null
+++ b/llm/nemotron/llama-3-1-nemotron-nano-vl-8b-v1/config.yaml
@@ -0,0 +1,48 @@
+description: "Llama 3.1 Nemotron Nano VL 8B for vision-language tasks"
+base_image:
+  image: vllm/vllm-openai:v0.11.0
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful vision-language assistant."
+      - role: user
+        content:
+          - type: image
+            url: "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+          - type: text
+            text: "Describe this image in detail."
+    stream: true
+    model: "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
+    max_tokens: 1024
+    temperature: 0.7
+  tags:
+    - openai-compatible
+model_name: Llama 3.1 Nemotron Nano VL 8B V1
+requirements:
+  - transformers>=4.55.0
+  - accelerate==1.2.1
+  - timm==1.0.12
+  - einops==0.8.0
+  - open-clip-torch==2.29.0
+  - pillow==10.4.0
+python_version: py312
+model_cache:
+  - repo_id: nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1
+    revision: main
+    use_volume: true
+    volume_folder: "llama-3-1-nemotron-nano-vl-8b-v1"
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+docker_server:
+  start_command: vllm serve nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1 --tensor-parallel-size 1 --served-model-name llama-3-1-nemotron-nano-vl-8b-v1 --trust-remote-code --max-model-len 16384 --gpu-memory-utilization 0.9
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 16
diff --git a/llm/nemotron/llama-nemoretriever-colembed-3b-v1/README.md b/llm/nemotron/llama-nemoretriever-colembed-3b-v1/README.md
new file mode 100644
index 000000000..e19ffc5b7
--- /dev/null
+++ b/llm/nemotron/llama-nemoretriever-colembed-3b-v1/README.md
@@ -0,0 +1,61 @@
+# Llama NemoRetriever ColEmbed 3B V1
+
+NVIDIA's ColEmbed cross-modal embedding model for text and image retrieval
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/llama-nemoretriever-colembed-3b-v1](https://huggingface.co/nvidia/llama-nemoretriever-colembed-3b-v1) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | L4 |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "queries": [
+    "What is machine learning?",
+    "What is the capital of France?"
+  ],
+  "passages": [
+    "Machine learning is a subset of artificial intelligence.",
+    "The capital of France is Paris.",
+    {
+      "type": "image",
+      "content": "data:image/jpeg;base64,/9j/4AAQSkZJRgABAQAAAQABAAD/2wCEAAkGBxMTEhUTEhMVFRUXFxUXFxgYGBgYFhgdFxgWGRYZFxcaHyggGh0lHhYXITEhJSkrLi4uFx8zODMsNygtLysBCgoKDg0OGBAQGy0dHSUtLS0rLS0tKy0tLS0tLS0rLS0tLS0tKy0tLS0tLS0tLS0tLS0tLS0tLS0tLS0tKy0tLf/AABEIALEBHAMBIgACEQEDEQH/xAAbAAACAwEBAQAAAAAAAAAAAAAEBQIDBgEAB//EAEgQAAIBAwMCBAIGBgcGBAcAAAECEQADIQQSMQVBEyJRYQZxFDJCgZGhFSMzscHRUlNicrPC8DRDc4Ky4RaSovEHJDVEY4OT/8QAGAEAAwEBAAAAAAAAAAAAAAAAAAECAwT/xAAjEQACAgIDAQACAwEAAAAAAAAAAQIREiEDMVFBE2EicYEy/9oADAMBAAIRAxEAPwD6DrmJJYeme340oXcxgfdXL+pk5/8Aeu6Fv1in0rsSpHM3bNNY0yWrbZJJj5z91eS4oxyfWqLuu2iIz3oD6eOIj3rNRbNW0g/Vaog1Zo7qt9YSKUX9Wp49e9DjUQcGKvC0Rns0+tccjFKbWv2mRQTdQJ5M1UdYD86a46JlyDm1qt5G7iimue+O1Z9dZV30otzxQ4bBcg9sopmfau2tAAxJ79qU3tfttXDtJhD7fnTO4lxiYU8ms2qNFTGVqyqiIHrQdxVjyc+lBXVuj2++gpuehpxj9sTl8od2tTAg81Yt7cQBQWlJOCD6UZZ0Z7EVLopNjCzaC0FrzmIx60Yz1RrcDMRWcey30KXq5c9oqDQea9YPIitmZILswP59wKNt3MUCxWMVBtURUNWaXRZr9SKC8X1/OuXroJoS43pmrjEhyCNqHtQWr0wnymp+G/YTU00rfa8vzquie/gtKOpmonVXBOD+FNXcL3of6QJk5p2TX7Fha56NUbGmdjgfjTc60ThefeqzqfaJ5p2/BUvSixYKGZFM9Pd7mlbagTxNRfWntScWylJIb60bs7u2BSu9dAMmDHrVP0t27GKqZCeQaSjQOVkbXUNrE9iDXf0s3rVR0s9iK6dAPWqpCTYcOk7uGg+8VfoOmqD52/lRli2AJkk8g/8AajbelXb5myeI7Vk5s0UEL7+lH2Z/P91BPpj3FNrlsAwZHv61y2EGSZpqQnFMT29NuIULJJgAUd+gSAS4j5GaYprgkAQP9Zr13qAjIJ/d91POQlCP0Gt9DsRkt+NUfoq0D9oiiU1IP1MY+4ffVVx27mhOXoNR8JnTWguFE/vqjUEAgDiouxNVMpq0iWxR/wDELWMNIotsVVn23IIyNshSCJOc/dTz4c6i7aWy11ibhRSxJBJ9yR6jtWc+Mgfo4HEuP+k/zpl0YE6eyc/s07ewpYoFJo0hubsxIqhV5waDTePssPuNF2N/IBn8KmqKTstt4z5oo63f4oW1ddvsn8KtKmM4+dQ9lrQd4sCh9VqFIIOaoIHrQ5RTNJRG5FN+8O1UfSmHarTZjjFC3FPpNbKjFsI+mMewrg1frFCS3YVUyNTxROTDzqVnMfxrja4DAAil3htXfo7/ACp4oM2G/pEjiqX1ZNUjSN7UVY0sc0UkCcmVDNEJaBxFXLbHpU2s45iobLUSP0e2BJ5oNlHaimsk4Jqv6N70kNgLaf3rtu2Bzmr7liPWKoYCrJObqibleMVzcPSigsrdjVW40SWFQ3UAMRcEYNXWtS3pPsKoSysc15EXuYqKRdtFmq1ZIiINAknsDRm1PWfvqxtvYimnRL2Brp3ImPzonS6YEebHvNdCwe5++K5duL6R99N2xUkMjqVCQIoK+R60EXFdDD0oUKBzstKnkV5agrek1Lwz8qokR/GjjwreftjsTk4HBHrTn4Xb/wCWteyAfhSL40VgibSsyMsu4eaQBHucffTX4YaNOgbkCDAgcA4H31LLRpFWYMVagFBBgBz+deOrrPFmloY7f7UUHdt5+tNU+PUvFoUWhOSZMWfevG2BXPHFUXb08RT2JtEjznihrkGuNJrq2WiYqyGwh9OoURz60JcT3qbBwKqa2x9aaQmzyIZ5FTkVAaZj/wC9eOnI70aErLBUgaHz61zdRQ7DVep75oAXKt+kiMRUuJSkEVBnihm1FUtep4icwm480Oy1A3a4bw9adCyR1rVVOgHeuPfqprtOmJtEqjvFVtd96pa5TxJyC0LDuasDGrUq2aVjopDVYGq9R8q7FFjor3n0rviGphasC+wpBTB91SRoq6B6V2iwosTUCAP3CprcjnM1RvNRZqmi8jP/ABg4MD5Sds9j/P8AOi/hBp04wMR2jtH+WknxXrE3Mu5d4IgEt6EAmB6rxNMPhC+NhClWG0GVJbIgkZA7XPzFOgbNMbf9kVzw/wCwK6jD5V3fSFo4LQ/o14oK8WroNGwOFK5935VKa5NMCO8+n5Vw3T6VOa4DQKzioTUvoh9astvU9wqW2WkgXwiOCakLBPJNWE+tRZadhRVcsgVSEBq4oa5tpiKGtelQNuiK41OyWgN0zioMtGFagy07JxAGqsmjmtg9qrNgVWROLAWqsmjX04qlrFVaJxYK1QNFGyK9sosKY0VasVKxo6vqRnxmxGCqk+4Mr/qPaujq+rj9sf6U/q+I7jbj76jFl5I2oSpBKxyda1cL+tBxP1EJI94GfnVn6Z1n9JfT9mmfy57RRix5I2ASphaxtv4g1U82/vRP4H5/hVi9f1fPkj/hrH86WDDNGuIrkVlT1/UiCQn/APP/AL1IfEt/+jb/APKw7fOjBhmjTlahcgAkzABOOcDt70hX4jvjm0hByI3DGR6nuOfavanr5ZCGslc5O/8Ao5M4+X40YtApJsxGv6de87O6sBAZnw26N0cROePei/guzct6i3u8ZVOCAPIQwIG7HEx7mPaiLmqD6UeRg1x3uEzP1iNoz6AR91WjVIEsPDibZE5Am2xGPkp/H7qFfQ3Rv1X2qwClGm+IbLKpk7mjEfa4IEnOaJPXLAwXP4flIxNKmGSGG2u7KXDrljHmOSe3pPofap/pvT/1n/pNKmPKIcErwSg16xYP+9UfMMP4VYvUrP8AWp+J/lSpjtBO0VEpVR19rvcUfj/KufpC1/Wr+f8AKimFotg10Gh/p9r+tX8/5V36ba/rF/GnTFaL2rhJqn6Za/rU/Gu/Srf9Yn4x+ZpUO0TrlR+kW/6xP/Ov86811By6f+Zf50BaONXCaru620s7r1oRz51kTngGeKHt9X07YF+3niTt9vtAU6C0E1BqG1PV7CCTdU+gSWJ7wIxx70MnxBpj/vCv95G/KJp0JtBxFRpPqPii2B5Lbt8yFH4iZoS78UPB22lB9SxMfdj99OibQ/YVBlrJXfiHUx9ZRk8LbKxAAHeMgnJ71S3xFqQMkHvPhpn8BFVTJtGuZKgVrKt8TaggAKhMjIWSfYrMfh6VAfE+oP2LfplSPy3U6YWi3SBDcRNmCjMLqnJ2bi20cQFHtXGYW3Cpd5B+yR6BiQCTMGc0DoL1ryhgSxZp9SAw9Ss4nEAmZmi9RbTcEsuFO0bhdCypJEnyyTgFe0TNZZG+KCtLeXeiKWbeYuSIJWfMCy8QwMR7UVYVGnaXXaGY+eSQCcKIwYj5UCga2zFWtCCSGk7TMmCNwJ/AUHrerhYJa2BCiURzMySMHntn5UW30xUl2htYt7vMLj4wQWgndIUY5yO9Wr06wXYHxTCllAZYxEkkiRE8QaX2uootu3dCq3jPsBQAb2BLAHz+UASJj7UxirdXrrasNzgFjJIYO47hIQBSvHBnOaiU2vpShGjvULIs7Z77olvNieccAyJ7xRWqtW7dpHh4cI3PYg9oz+NLdV1pFuSqNdLMDuJTBkkNgTIkgcjiRzXtNqfEDXEuNJBBDoAZkbly2RmZHahTv6Jxj4Ok0bAlN0MrthnAA4I4njmDxPvSrqPS77FpvAgySu6FgnHDwZxijdFpXJZFvo91myIwQIEgASeSeZgVG4nhOj3EueJ54uOriym0+Xw1AO2UUTPBGKMn6Uox8M/c6EyBoIETIBjtuz5vYfhSNbJ5k9jz/wB6+kdP0x6iSfCUqmWuNccEMw4G1cngngUq1Wj0+nueFsVl3KGYNuQRtOA6zjdPerjP0iXGvgB8L69g5S5cIUjBMeUrmTyTj0zgUyXqCsTsEkMR5gIP1irFgIGApInmnOn+B7Y3EqCT9UKzeGOTMFZmTODEYqfUui/R7TP9W2is5UNGQZBHlM9sSJmJFJ8iGuPRnn1e18wP1pGAQZZXmfxAgVC51AiGFoPMggK2DAI4OR/3pJ9JsjPhgQtsjJJB3EtcOYM7QI9xWvPSN7kbrqK/nH6vbbEDC75MEg8DvNClXYnHwV/pV0GDsDQDAJmckfL0n5VRpesuHWNpMgTtGWJgY7jOeKMPTQA28XGTcQ0kQChOAQRwwYye0cVf0/oGnIEW7rAEQ+51LSN0wFI7EYIz+FNzQsGXa/XkJsACMpZC5CAyoLeU7jElTmO/rSvT9eUTNq27EjzEBjySYAMKTI49BTDqPw9pwq3GNwSTPiXBBBknzJJkGBkCd3elzdCsqSwa4TLwtsNDQBsC3GTBmTmZiMd5i1RUouyX6dBORgehI9e/zP5D0owddtFllGCx5gpGY7gnjil1rpNtmU+BqlthkRgzKC87gWBCkyY3QBA2xwaM1nwzaIBtXygKSVuyWJgkYEZORECPem2hVLwrt69brLaDku7xMYhio2xO2ff3q69aKqqoboK3WTLDYGUBgQu6CSDkgdhQw6NYsy6u5dBuW5tgBhG0wGIZQZ7+nvVVxngqt83LhfeLfhDDbQGJMyx9gBxS/oaXqOafrBDBmBfEfVWYzIEfj91ct6y3tDXm2liwIhNwiCJQNIBJP3VFOgu4At5b1W0wHGSXLwM4iKpf4cIeFuCRJkqJ7cqG5MmPWDVZIlxfh231DcZVkwO4jEmJI++jtN1G3cJVILMCCoEboBI2k/fM8ia907pdhVgfrLrsom4h8MgCSAqklSJA3d4NWfoq2twFLL+IEJDI6IJKlc2mJAHmGe09zTyDAA1fURtDeGNu4LIWDJE4z6dqZWdIjuiWyGZ5C7WVjgDcI3SmZyYmKotdFW23hLdJYz5Qm8YiWY/Zx957RBoC/wBOsBiHeHViCdjKZzAXaZYzmZFLL9hh+hjrOmtbcK6+Y7ZBKjcWPAzk+w5qnSBDLKwTzAjepJgTt8q7oeTxx71JdQ0b2cO0RuYktK+VRxBYQPXjiudTTc7KxVZ/ZAHLDBLFo8ogz3mCIxRk/rBxXxFWtW3uneigeWIM47mR65++h9yQ360g52nY0EweYEKJgfz7r9TodQJQWy0jfysELMsJMng8elDN1K2bZDNbEZK7vPE9xGCar/Sf8Grtahv1sz9QnfiM5Cg8jueKu06qyiXQEYO0l5juxkZ/hFLdJ1Bb6kKUW5EmcqVxicEEe09qo+glsqbZ9STcWT8g4/0KNgq8NFrutCwSNSrq2WCI4uiCSIJQeUwB6R2irtH1u3qBbD6eTe3eECSVOxmA3OACvaRPIIqnoujVEVbywoUC2LN1WmGLBmg+XDDHHI7GvL0nwLtm5obNtlAyLjuR4jSPJAKkwZn1Fc6Z00NrfQdzKLuktAFWLgF8EBgoDgk+hgGTik2u6bespNjwwcK4g/Wgk2x4klTEeXdifatbperW1AQhlYfWAQxuJJOI5Jz6mguu3Gu22VFNthcRiwJ3HyFYgLJbawEe0dqVsKM34d57BfYlt9ha4V2KFAYKP1nAmRgf0j34zuv6XddhbtwzjdKSu4EBGMSZMbgAPf51otNoL6gKysVgjG4NDCCyGRxKnEx7VZb+HB4pNy2+5bY8zl7kDBKsGBxDzMR5uZkU6V2TKLaB+l9FNnS79QqWnuOwU3FD3bZWPOjQduG+6Jq23qbbHxDeG3zpt86kqykyHBIY7gBn50Rpunm6i2pUpaDwGuWECktuYwxUmW+4AVbY+GVLKjpbKAqpEnaC4YAF0B2xzPqAJM1NyvQ8aA7T6VGddyNjebhtXRcDBgEWSYH1jkDPHpUtD1G3dZVA7v4izqI25hlm5Bwcr8uKs0/QN/7BrYeTtsu5S9AlS20gEg8TPE1Vr+jfR2F27aDAxuKlXO4wBt2sTnInsRV2CiEgX9qCzbFsAK+0XtgEzu+uZPrPEn2qtuj6htyvZbw/shm3J3PkMLuJ9ATVq6uz4YWXS65Gw7QBBC7YZzFvvkxxmJqen1H0j/7ibq7lZLtl7hHILlhIACkZHqQMGknobWyVvTapRCpdYQNv7cAEQoHMDBHrAqWts3M2rguMHTYCTdKSVgSCeAeS2IWrbHRbIt+Cvijz+Kdq7BJBAIQmSuZwCRTHR3b1htwvvyWUEXCHLAg71Hocwe+alrwpPRlx0ZGc7yWhYJQqgYqRnd3EZrQPdV0XdqQyhh5blwXEwrRumAGwBE0LcuLbG9DqS5ZXCXDdYFl4O15BUTM8SB3FNbPU9VsJ1eqti1c8RQJBJCwGYEKIywEe3eadslJCBkCFmBVyuGG62WJdQ1vwwoZbiQYc/wBIGJImrdNoLZdA+4sxukk3Ni4VW3SFhWI3Y98Qasu37Soy3Lt1iSUfaWEhFm2SrYckqoBExun1pdrev21ktobfhgxO/DbmBDskzwqEj39qakGNDsXX2MrPvtpad2ndLAMIIZlgk8AY+qYPapaLrg8K3m5Zmdik5AJOzAOCeY96TarqtlbpQMChDZD31VpgklLy7ZgESpOfvhxoOn6cAEXXTdg55wAIMT9UxFP+xf0H6TqNx2VRcub5G2TM5VSwJEfa9fWlq/EVm4V3FrpYnYrKkuUwQSw8pB2+nE9jXfiXpVu0v1nK7lUztYICGjB4Aj171m01tsMG2MzgfX8IMfNAef7UBTPODS0Mb3r1vexNrYXYkqCrsJAIZgkxO7k+9Qu6m0VP6otMEsJR0AEGSIZRPvxJo/puttGCblsnJRUVluTBKswJMQcEbfeaoW/ccvvbxIZkuKyKVndukRJIAIz+VOwoV3uvaVAVs2WfBWfFdVIyAQZM9swMUo0evVLou+GywceGwPHZtw2v3kkZmi9Z1GxvDWU8TzByUTcMkE7gcgcCIirbOit6zVtujToZJhkAA5hROJ82SMdxTM6C7GussCE+mOZO4L4O4EnjiRzgcUde1Vu1cCvacuVBe091bZzAC3FtpLHzLOftCl2n6RZtX0PjpsD4urfUsEOVYImd07cCKfPcsO1wHURbfzbhYuSzliWLgfWWY4JnNJspIGTqVs3VRNMlu6CwUi8d6kgbgQUnMgQezYigdH19WUXWWxp929G3Xbik7YBTaijGB34JNS1WqQNbZbzlrRBCvacEkD6m7lVmCAeI7Uiuolxyb6+TDhFJIBYksA3oJ5p/Ni66HV/r2lULtuaV9sttu+JdQN/ZG5R6crV2l6wNQywmkyYW5bsXFKtHKutwRmJ/Cszb01iGIsjcQ20bSQJPGf7JPFW2LiWiPBtqkEtO23JONhk8Rn/RotDph3T9T+rRijeLthuWNsi4QAGO6AS2Cw7gekMbWqvs7KLl13VpZUVbvmMcgCY9gKzj3H3O1t7gZyCfDuG23GZKkY3QY9a71Dql9rni+bxVVQGLAsSgHm3lvTvFJghxrNBbu3V3LtvqvhjyG05PE3BiDtOf4UQnQkAhrAJ7k6m8T/6TFesaJb2y54v0iFC73QOFPdIOVgyIOTAPFM10sCASBWc+THo0hxqQm+Ftw0q2xqlS4SrKRvdlVfrWwu2W8jMJnmjPiXX6drZOlXUkuNhuG4VRNhBVGBEEkE8GQRFQ0FvqO5Tc1jLt2EjxklhgOFHhzzIwe44p/aR5M3rzIY2ozYB+0xj6xPv91GqFR851r65l8yuQirw5kBZAZiBkweTJ96N0Ou17qwLFWLWmVnZlI2ERuTJOBOJJM0z6p1qwjurBBcTco8RiS20yFYZIHaI4rt+6uosWr2ktIbk+YW1BZG2gMCJBMMcfcapRaJVBn0k27TW0uXA4ELIQLtYRvz5go8w2YnHyria2wVjUI11ioXxLd5gWWQYKgqQBAgCeBWW1fSr+ouFLkvcQBSsqt5RtDS6yC6yVySYJ9qrtvd0ynTpffTuXlouLAO2GBCOx5g5iPvqsf2Dl+jap0q27k+JcUMWMrpzKhgYEs/m9N2c570P1jWPpLSqzC6rFgCttg6kEbASDuBK7juAA3YzM1kG6ldJLNeBJUWdrFyGHkVroLMdtwwTPMtiKFNwHaBdLsgi0QqoXBHm8Vok/WmTJ+VCSByZu+h66zqGOzeriCN4AYjO6CCTAPI9/emPUen+IuxQpfywzGAPacnvXzjpo1FtxdV7Hlbdk7hOdwIiIIBB++tvq/iFy9hNPYZATuufq0CKssGAbHmBUMD3D/gpJXoIt1sV67pl0oDcVfDVIPnncODgDmDxS4aK7bbT3LVvYqb2YzEwiDdn65ndz2p71DV2XUW7gIIY7JJEqykx5G43Rz6Um1Nh7g2kzbVoClCywfKCCGDHG7k1OcY9il2aFdeyXLDG1eYJLuw8MJcDKY2nfjJ4IHFEdU+Iptm4qi0VMqjujXGgyPIPKTjBDSPyrHa4rsljbA8suUO4GGA3kk4gRxBNUWbaSsOQ0qW8qhR3HC+UcZpLkh2N2w3rXV717wgiuRblUXZJWQZ+q2VwMnia7Y6vq1CqdIbhUllZtwPptKjEDFWWNaiXGWOFjczttb6hYAjgjHuZ9KKfrFsZiOMS052kd+f5VS5IsimvoOnXNcpZn0u5WxtZjAn+i07u3B+4il+p1d24HR9OLaPMgMSTP1oLMCOAY9c0fqepDnyhIQxiRAO4gx34+6g7Gq8RtpO3MITndInzfLH50fkV6Q/8ARfduuPDTdtgAQV3dyJbsR/Cnv6WuWwFmwXAkMx2EZJB27iBHY/yrOa18+IWbLIBBGJLYH4flTW3qSSNywc8PiO3Ye9Wtk3Q31nXr1y06XWtlTbQkW0yQZOG7wAx/CkWi1f1j5SRtERI7ZJ7Y4959aL1F1tjYMsIWJk5BweOJz7Vmr90oyxADrldxJQA4Ej3j8Kxknlodj/ql8JBUndt7Erkg7SpXtyD7036N8RKyXX2+c7LjDdtUXCAtwKNpJG5Swg8H2rK9QvbkB4ODGYiMR8ppb0e8Q7LIgpc5BIkAmTnAiciiH7BypjOxeFyUFwqRJ+qoGP7sfv7H2p30npdpCbhvEDzqQF54jcZzlpjjFZDprwzehIwJj1wTmfxp5Y1Z2sOTP1Y8xJ7xz3z91TyOfxjg14aJrXmBBUiQJg88SRPfmqbZJUgDzFVAAGMlzJxj6oz8/as+eoOLiqjj9YUYGJkEAbTPBxn5+0Veti7afYwJUMTChyRggw0HuZ7xAq+Ntqn2OxvqLg3N5l+uw5HrjApabgBiAQXC9+CwBP76sTpzAbkfJ3Fot3AWYldu7y8wDj3oRujOcO6qSQwAt3NxCtz97Aj7iO1a6Fs59N8s7VBieP8A8gX/AKa5f18TECBc9MRcXb/6JNePw6YKl1DHcRIYEhjMKpHmzB+70xVi/DbSZLT5jlHCtuiVK7e8czRaCmVanUMPNbEiI4aMsYjbWn1GmGot6dC0t4VtMiAuPNtPJyO9KrXT7oRbYchBJAFm43JnMmDE+lVaa1dB2pqEMEnaqDyzyTkkZ7YHoBUzpoqOns+kdE6DZ01vapE9zRbaZfUV8vv67UIJN47ZA3EGMmB2qQ6pqIBW7vBnKgxgkEccyK5GvtnQppdGtGohSYJicAiSVBJAkwTAOJqxNeAQWe2bbEgFSGYSispaDAH15+aV810wYrCbdhYmdhUloidvp70dduop2KFHOYCqTiZX7jBE81rySx6Mch/1jRaV7z3WRbhIQuQxKllG1gqAycBaFPUUZTbU7VVYARpiTjaT7/uPyrNJ1OWbO3lWIyT/AHce45zzQfULu1jDSGEwpBAM5k+uCYxBNZPOXbJyNZfu2F1P0i4uy7KecEbSApVgBwAQTIMVV8TQ7i/baA6w/wCrRjuUohbfMkFNmfUP6msbqOpHygkEgwIwDySSYyZxRPSdUxlZkSSAWIAx93H4Zrbjyj2yW70H6e2pXdtO4OrtMLywY+X0lT3ruldU2ZHlG31mVAMY+VVayy4UMxWN6SM/0udxOIz90elBglQGydoyAM4AUn5eU1vVg3TNFpbgATbajeSRuVlACkguJ5Pmj8atfUrJVecSVnYRM5gy3YSM0BotQWtyrES4yQrttyFVhPl+q3z3Ckl7WXCxtohgiDsB3Etmcd8x/OuOfHJyropyHr6kMzLuLOBLEiF5jBPsSPuoe7rA67h2yxkYBIiSDJGM+npSkau9aCpeSBcUMNw8zLnafXmOfShvHjg4zPv8/atIwIbNHcgyXcAv9YBfLcEeaQMkR3PqPlXPpaKhVDECZk/kx+VKtNana7F3YEnygBYG4DzlgRgA4Hc17Q2mYndP1WYcESVJB+WRjtmmuMNjW31Dc68ZVjAkQeTn1gT+FBdV1X20aIPqZzPIP3VXoNPcDTABQPnkEx27RmvavThm2lsYJIyxnk54jH41SgT8K7uoaAJLA8+nvmi+j27j7QqEwxJA7lhCySed0Y/dXNTetooW0i94ZzBWTIJM/ZkZ9qLXW77ZW3b379rsFV7ilgVCydvI3MarEaiUWdDce4VJW0FKSzttQGZyRJBP1QY9aY6LpNsS1+7pHVWLbhq1JgMYUpIMncYOMDMYrl5r11P2BDbD5iqgc+UwTMDHagr42WFTUKpcz512cBsAkfaiRx29qrrQ6rYVq307eIi7ipd2hHt7YYkjaM1V1hLN66rrvUBESCUAhFgY24yP9TQFzrLIP1SoS2TKlyMAAQABgCq06zqG+s1wcf7uFHtCiY++ikFBN7RsAF3oQe4JIInAfaCBHvxUE6Y6XNy3dNO0/wC8wC0jHl82CD99STUOTI1It+x3/wCZTTnpmqYHz9QQj0Kqf8opKPgrS7EWg6adrS9kEtyzlSIPpHB/jV1/ppZf22nJxM3GPY8QpjtWl6t1cWrKXVe3cJu+HIIVYFtmeRBI5WCKz2j+LblzUIj3Et22YBtgHGZlzJ++KKY9UDX+nMf97YEZUKz7hnt5PyB9aqvdId8PqLDAZjxLgUznEpmvo1i/pYBF+2f+cH8zRj37AUEm2cnkr7fj/wB6rBkfkR830uguWyGtanS24/oXI+5gwUN6Gaf9Tu3HZbumFo3PAtW2bx7WGVf1hVS5EbjI+dab6PafIRI9ZWq26Mm6Lb2mLDaQQJElfqzG484x2zScX2UuSL0Y3Q/DVkqGvtda6wHiHdafJ584Yk9szR6ak6Y+FpVuXBmRcJ8vAO3fheeByY+dB6pgeLEDe4kJ5mVWOzA9RE/fS9NejB2csx4+tECJjHA9o7Vx83LKWkaqNMbaWwLlpbd63ds2wWARnjd/fMk7ZMBePwobrl22tspZsW08gJZFAPrIZche0n04oO715xbXw3CWy20Q2eJ86xDcGeJ3DND29ba8RjBBK5IwWmJUJJC7cQRgxWK/Jbcnrwt1VAn6WKAWyp2KVO2ecmZnPcRie9OG16MZD3AOANzjaB2xSm90pnko26CBuaBg5SST3A/KtH8P/D1l0c32lxcIG246CNqkY2GeTmtfxqRCsT6F2VQd4Cjwxljw0YG7MwSPaKH6jqUIDwdxYQCWwq8gHuSSCfl2rR/CvwidQgu+NbWEQlWLC4jEEAt5ZGBgTkGKr6z8GoEN1b9oLKiHchg7Bg25pOFwSvea1k432GLMT1O/JAEADICkkKW8zDdycmojUArzyeIwCeY9q1vTvglGQsLqXSyMLW1iqs8kTPpgx8vwV3umhUZjaW3CW3gs287jsO3bIJDTgkH2xWkaa0S0ILrmP9flRGkuAESdu0Y9TJHJ9BTDS9OAPnSYe2vmY2w0kTyJiPbvWl040rm6yaezbZFh4AuWQScbAwgkANyPSqEkAdHsi94nimV8O48tG0BQ7EmWECNo3epGDWoTo+mtJfe0ttElEVmTxkuKZ/VySCWMAGACD7TWH1vU2tgOjESXgso3Eq+CpK4MifbHtXbXUrpUF9u3yMCzLvkHBZwpLTAwT86uxn0h109i3hUt7grz+r3IbhceVwCQFjAwQDWS1qOx+k2XPjvLeUudjKwVg5jaBKhhuwc+lIm6hdJi2yks0kIq7iSQckLmT60V9KvBmtuqAozby920qEsZMAsA2TBIJqEtlXYNrLm5/wBcyHYCABhAD5icSc8emcRilwe0wm2j7vNiZH9gT6cyTTTU3kY5ti4+dwthDaWR9lkJV+0ZMUd0zTXrllms29Oq7vDKbCGZlBYfUG4MADBbB3n7qJoJ6f8AD2oNveltcR5me0AsrAA3HaD7TMVzqGgZjcueFatqp2r+sQqeBtHhk47yPX3rlix1GMae+yA5PhOV74MCBz3A4+VB6jqVwzaYKPsGUKqpAJ4KqQ0rjEyI70lYaC7ejUJLbioEkWWUPBgEk3uVwMqCRS0jSKQ7i9sJ4NxQ5wJ2sBtH4U3sdE1Lwyaayo7b9yckEkIzFhOeY5NGnSdQQqRprJAHmCFXLe5zuHYYB++nsltCvTracn6C1xQDwSlz/m37ARntTlelalxua8xiSQCTgDMjjtPFVJ8QaseU2bYYbfLIX7QmAx5j1jn2qV3r+puq9pdNbO5mTbvLP5G8w2IQ0BlGTAMd5oFv0c9K6IQoD3LjqOAZG2e3Yjvg0Z+gNP8A1a8z6/upRodf1G3bFsaAbFJKqrKFBMkkIbhyTn8aq1PxZqkIFzSFD3kETx9XEHvx6iqTS+ENN/TQHpqKICgfdUBoF7j8qQ/+KtQxG3SO3eILyR9SCsjEZHNRtfFd/j6IxYYITewB74RCR99GvBYy9NCen2j9mfw/0Kk3R7XJRRNILnxZdH19HfA9StwD8Snv+VVJ8aKWC/R2MxgEs31RuIWOJ3fdFP8Aj4GMh+3SrUfVWOapPRLR+x+R9qU2fjJI8unucTgTkROflntROj+Mbdxtot3JPpyCNsSCR6nzcCKX8AxmGH4cs/0R+X8an+grYUDyiJIEKYBAjt7GqtN1C9cYy2nsAHl7ltiwBIBAUtyB3Ne1vSBcMrrbVxiFzvCKCGHYdtu7PrtouI1CYP1GzZtiblz2AMD/ANh70ItrTKGa29gNtEAuhLEsuBEQe+4+ld1fw2wtuLbszREWygRpPE7piPUduKQ634Z1JkeBdIYEAgWruSSADsMyYPv6VLp6LUWgm5aXeFu31Tglg4XiYEIG/ojI9fWl17patba8GG3xFRh5QY2sQew2nyrEhjPGK2XxBqr96zdtjTazdct2kAbT3iJtsWJMLSLQ6ULbVb+ivswPlIGotGCokbfDM5zPHNR+KK6LybMVo9bsbeEBI+zvnJ5ZVjiFjkxg1d0rpj37zeGoI85JLbUWf7RIjLARntX0j9OD6raJyBCmbZzt4lmsiT++vf8AiGx9rQnB/q1X5R5Bu4/KanAvQB08W7di74ljxbe+wFt7LYJJW4FuDxVYEDY0DAO6jdXY8527SMETatA5E5AWut8TabcSdIwkgv5dswGCyZG4iT8pNU3PibRzjSM3uWg/m1GLKsymj6yqq0yD5ZIJDjYynIPlYSoEe2PSmPT9dccN4Fy2bks83Sr3GBgbURRtXJMnmsF9JO4xwcEeo5+fan3QNZ4L7hGyIYmfrBTtIPp5iMGuaXGqtGYz6v1LU2bi+IrKkqrbT+raPNKjBnEx7Ckl0NuP1rrm6GJQMSysQwcSJBbJzmTR/U+oF7RX6xMgYBbtkHJj1Apn0jUq9y1btW9QBMlnZQFVPOdxVPMMEAGDkZrbh1HYCk9G1Qi41k4cvLsAQAABuHI7nNGdLTwUbc0lzugJccNH2AwCoJzLbsTXm666a0m1cdEIhzJYlcltwbnOIPrV2t051hAs3F2KWEBGXwyYIDwYeYIHYQa0hNvsJJALASBf099l3MYnbhpJIIEYngntzV3V9NbAtPY2WSEDbLlwG6e6nYxK89hk0wX4P1RUL9MxERDwAewFFr8CMQBc1DkYkBRmPnJrbEzc0JdNpdc5VzvUNG4tFpSMYAtiYIq3U6JrSXmu27VxW2qiL4kqJYnaGXk4mK3FnphSBvcgAKATiBx/qasuaERn9/epafwFMwPSejLqAd9z6OgaAgtsJkDzS0D8qar02zpzC+Hc2ncPFBIIEDDKRx6RWqXReh4/jPp8jXfog7/wJp2yclZmbXxDevMRscABUX9ddCALEhAk4IAH8RSd9baXqNtlAUpqLSAm6xhdwViAZ5Ekk+9fQRajgkfeKF/RYmQMzMxSp+DzvsNS+p4u2h/zg/uqzcp/3qEexoQ6Qn5/IV1NGV4JH4fypmYB8YFBp2Y5MqodT5xPeP4UL8AwdMwViIuNO4qXMwfMZE8zmj9d0lbgO/axmSSAZ9ORVOn6PbQQEQf8q/hAFL6Vaqh3sWMsDn17V1rSEGJK9xyMj75paLKWwWKrEj66qF9sgD99U37qL5iUPsDzgHAmO4P31WSJxbMz8RhLWsRELooayHtwNrbmSROPLDEVvktBPKECqDAGYjtgCBWH1GksXLvilQTz5mYwRwNm/wCX404XVsQ8KAQSMNcDD0OWgZkduIrNciN5cMmkPLtu4bTCyVLOAFJZrYAY5ZWzmJjEVhdV0hluC3dsG2GW1suneSFJJZyrbQ8MzHdI+qI7UxvWn8pt3T9Y+ZmIAhUYhQW4IbbPtTjpZ0V57dooviH6qLevFgqPFwksxG2ZgdwOBUvkfhpHjxXYh+DdKwe6l5967h4YYttZpYEmW8zAKOZ9pp7qOq2Sm1LttdzlZ37BCFS+DBIyB6Gafa34WtyxRTv5HiMWtZL7pWCJ4IMT7jNK1+G7qiyPG0xgMUDWpA3wTu8uAO244qozdEy41ldmZ6nqL5G2y8+YOqm4m1lQiQR/RJnucH8CnvA7lRQLqqp2DbImIyfKRyDBNWdd01/6TZNzSWGd5tKuwm1cgwrC3uifNP4VZa/Uu7nptrKDcFN3amMOoG4WxkkBYjFKysUL7X0o3Lha2xUhfDbbaJEbNy8xBG4/d70s6tZv3Wi4FRBfVxdYW0C20EFfQkESBma2fSfoTIrXtPa+kXFmbamWB8qk+aS05MCOPes/1X4cXT2lLhfFVru79pncP1c7xAKgNxEhh6U29Ao77HVrRrAbwkk5wE4ORGQsxHtPtRmgsAAfqwDGYnvMAEGsIt/wjdvFUIO0qp4WA2ARHIUfMye9VaXqG3UaktDKiqyqCVUTds2/KBIn9ZMQeO1NbRL0z6RZd4yLqc4D3OATEw0ScYqHjMVVj4ykjKm7cJHzO7BjMe8VhtLr/wBbqXIYrb8IqBccAB8Hy8T34qOj6s4uaosbm21cCqq3HgBnKAZbPHrRQWbmxqmIkvfUAGVa4xbHsT35jmoW7+4Bt10T2ZjP31jbPXH8S+WuXglt7aqoeRDozCZM/YPel2p+JL4dx49wAMQAdjY7Z2UqQ7MUv8/3GmWh+ov/ADfxr1erFgx58OftB/e/hWz+xc/4Z/eK5Xqh/wDLJXZ8167/ALTd+Z/6a0n/AMOeNR/et/5q9Xq1h0gl0z6Ra4X7q4/I+Zr1erqXRy/SNztVlr+Ner1IZV/vP/1r+966f4V2vU0BIcV5uV+Z/dXK9Wnwllj8UHd5HzH72r1erNgW3/rfjXE5PzrtepMZlOt/7Rc/ur/00Lo/2Q+6vV6sJdnXx9INHKfK3/1GmPwf+31H/Cv/AMa9XqmJrIG+HubX98/4Vut837Bvncr1eqo9kxI6v/L/AJ1rH/Gn/wBLu/K3/nr1ep/BMJ+Of9p0n99v8Na2L/WtfJP+mvV6khfTE3Psf8bWf4wqnr3+y2vlb/whXq9VPoEZTV/sm/u/5HrNj9rqv7o/xrNcr1EOhy7DbXGp/uWaut/X13/Ft/47V6vUySy59XV/8TSf4GopVrv2j/3jXq9SfQ4n/9k="
+    },
+    {
+      "type": "image",
+      "url": "https://upload.wikimedia.org/wikipedia/commons/thumb/4/4b/La_Tour_Eiffel_vue_de_la_Tour_Saint-Jacques%2C_Paris_ao%C3%BBt_2014_%282%29.jpg/2880px-La_Tour_Eiffel_vue_de_la_Tour_Saint-Jacques%2C_Paris_ao%C3%BBt_2014_%282%29.jpg"
+    },
+    {
+      "type": "image",
+      "url": "https://upload.wikimedia.org/wikipedia/commons/thumb/a/a8/Tour_Eiffel_Wikimedia_Commons.jpg/1024px-Tour_Eiffel_Wikimedia_Commons.jpg",
+      "text": "The Eiffel Tower is one of the most famous landmarks in Paris."
+    },
+    {
+      "type": "image",
+      "url": "https://upload.wikimedia.org/wikipedia/commons/6/64/Dall-e_3_%28jan_%2724%29_artificial_intelligence_icon.png",
+      "text": "i love machine learning"
+    }
+  ],
+  "compute_scores": true
+}'
+```
+
+## Configuration highlights
+
+- Base image: `fredbaseten/flash-attn-truss:v0`
+- Predict concurrency: **16**
diff --git a/nemotron/llama-nemoretriever-colembed-3b-v1/base_image/Dockerfile b/llm/nemotron/llama-nemoretriever-colembed-3b-v1/base_image/Dockerfile
similarity index 100%
rename from nemotron/llama-nemoretriever-colembed-3b-v1/base_image/Dockerfile
rename to llm/nemotron/llama-nemoretriever-colembed-3b-v1/base_image/Dockerfile
diff --git a/nemotron/llama-nemoretriever-colembed-3b-v1/config.yaml b/llm/nemotron/llama-nemoretriever-colembed-3b-v1/config.yaml
similarity index 100%
rename from nemotron/llama-nemoretriever-colembed-3b-v1/config.yaml
rename to llm/nemotron/llama-nemoretriever-colembed-3b-v1/config.yaml
diff --git a/nemotron/llama-nemoretriever-colembed-3b-v1/model/model.py b/llm/nemotron/llama-nemoretriever-colembed-3b-v1/model/model.py
similarity index 100%
rename from nemotron/llama-nemoretriever-colembed-3b-v1/model/model.py
rename to llm/nemotron/llama-nemoretriever-colembed-3b-v1/model/model.py
diff --git a/llm/nemotron/nemotron-3-nano-nvfp4/README.md b/llm/nemotron/nemotron-3-nano-nvfp4/README.md
new file mode 100644
index 000000000..6841b4c6a
--- /dev/null
+++ b/llm/nemotron/nemotron-3-nano-nvfp4/README.md
@@ -0,0 +1,59 @@
+# Nemotron 3 Nano
+
+Deploy [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | B200 |
+| Quantization | FP4 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.12.0`
+- Predict concurrency: **8**
+- System packages: `wget`
+- Streaming: **enabled**
diff --git a/llm/nemotron/nemotron-3-nano-nvfp4/config.yaml b/llm/nemotron/nemotron-3-nano-nvfp4/config.yaml
new file mode 100644
index 000000000..97dc549e8
--- /dev/null
+++ b/llm/nemotron/nemotron-3-nano-nvfp4/config.yaml
@@ -0,0 +1,34 @@
+description: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.12.0
+model_metadata:
+  example_model_input: {
+    messages: [
+      {
+        role: "user",
+        content: "Write me a short story about a cat."
+      }
+    ],
+    stream: true,
+    max_tokens: 512,
+    temperature: 0.6
+  }
+  repo_id: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: sh -c "wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py -O /app/nano_v3_reasoning_parser.py && HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 --max-num-seqs 8 --tensor-parallel-size 1 --max-model-len 262144 --port 8000 --trust-remote-code --tool-call-parser qwen3_coder --reasoning-parser-plugin /app/nano_v3_reasoning_parser.py --reasoning-parser nano_v3"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: B200
+  use_gpu: true
+runtime:
+  predict_concurrency : 8
+model_name: Nemotron 3 Nano
+environment_variables:
+  hf_access_token: null
+system_packages:
+  - wget
diff --git a/llm/nemotron/nemotron-3-nano/README.md b/llm/nemotron/nemotron-3-nano/README.md
new file mode 100644
index 000000000..eb83f9e55
--- /dev/null
+++ b/llm/nemotron/nemotron-3-nano/README.md
@@ -0,0 +1,58 @@
+# Nemotron 3 Nano
+
+Deploy [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.12.0`
+- Predict concurrency: **8**
+- System packages: `wget`
+- Streaming: **enabled**
diff --git a/llm/nemotron/nemotron-3-nano/config.yaml b/llm/nemotron/nemotron-3-nano/config.yaml
new file mode 100644
index 000000000..46643d44b
--- /dev/null
+++ b/llm/nemotron/nemotron-3-nano/config.yaml
@@ -0,0 +1,34 @@
+description: "nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.12.0
+model_metadata:
+  example_model_input: {
+    messages: [
+      {
+        role: "user",
+        content: "Write me a short story about a cat."
+      }
+    ],
+    stream: true,
+    max_tokens: 512,
+    temperature: 0.6
+  }
+  repo_id: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: sh -c "wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py -O /app/nano_v3_reasoning_parser.py && HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 --max-num-seqs 8 --tensor-parallel-size 1 --max-model-len 262144 --port 8000 --trust-remote-code --tool-call-parser qwen3_coder --reasoning-parser-plugin /app/nano_v3_reasoning_parser.py --reasoning-parser nano_v3"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100
+  use_gpu: true
+runtime:
+  predict_concurrency : 8
+model_name: Nemotron 3 Nano
+environment_variables:
+  hf_access_token: null
+system_packages:
+  - wget
diff --git a/llm/nemotron/nemotron-nano-12b-v2-vl-bf16/README.md b/llm/nemotron/nemotron-nano-12b-v2-vl-bf16/README.md
new file mode 100644
index 000000000..26fae6c31
--- /dev/null
+++ b/llm/nemotron/nemotron-nano-12b-v2-vl-bf16/README.md
@@ -0,0 +1,57 @@
+# NVIDIA Nemotron Nano 12B v2 VL BF16
+
+Deploy [nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16](https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+| Python | py312 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.11.0`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **16**
+- Streaming: **enabled**
diff --git a/llm/nemotron/nemotron-nano-12b-v2-vl-bf16/config.yaml b/llm/nemotron/nemotron-nano-12b-v2-vl-bf16/config.yaml
new file mode 100644
index 000000000..af53d2bcc
--- /dev/null
+++ b/llm/nemotron/nemotron-nano-12b-v2-vl-bf16/config.yaml
@@ -0,0 +1,45 @@
+description: "Nemotron Nano 12B v2 VL for vision-language tasks"
+base_image:
+  image: vllm/vllm-openai:v0.11.0
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    model: ""
+    messages:
+      - role: user
+        content:
+          - type: image_url
+            image_url:
+              url: "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+          - type: text
+            text: "Describe this image in detail."
+    stream: true
+  tags:
+    - openai-compatible
+model_name: NVIDIA Nemotron Nano 12B v2 VL BF16
+requirements:
+  - transformers>=4.55.0
+  - accelerate==1.2.1
+  - timm==1.0.12
+  - einops==0.8.0
+  - open-clip-torch==2.29.0
+  - pillow==10.4.0
+python_version: py312
+model_cache:
+  - repo_id: nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16
+    revision: main
+    use_volume: true
+    volume_folder: "nvidia-nemotron-nano-12b-v2-vl-bf16"
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+docker_server:
+  start_command: vllm serve nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 --tensor-parallel-size 1 --served-model-name nvidia-nemotron-nano-12b-v2-vl-bf16 --trust-remote-code --max-model-len 16384 --gpu-memory-utilization 0.9
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 16
diff --git a/llm/nemotron/nemotron-ultra-253b/README.md b/llm/nemotron/nemotron-ultra-253b/README.md
new file mode 100644
index 000000000..cbcde2375
--- /dev/null
+++ b/llm/nemotron/nemotron-ultra-253b/README.md
@@ -0,0 +1,58 @@
+# Briton-nemotron-253b-tp8-fp8
+
+Deploy [michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm](https://huggingface.co/michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm](https://huggingface.co/michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:8 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Tensor parallelism: **8** GPUs
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **65,536**
+- Streaming: **enabled**
diff --git a/llm/nemotron/nemotron-ultra-253b/config.yaml b/llm/nemotron/nemotron-ultra-253b/config.yaml
new file mode 100644
index 000000000..5b1a5dd81
--- /dev/null
+++ b/llm/nemotron/nemotron-ultra-253b/config.yaml
@@ -0,0 +1,36 @@
+description: "michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm for text generation"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+      - content: detailed thinking on
+        role: system
+      - content: Tell me everything you know about optimized inference.
+        role: user
+    stream: true
+    temperature: 0.5
+  tags:
+    - openai-compatible
+model_name: Briton-nemotron-253b-tp8-fp8
+resources:
+  accelerator: H100:8
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      # pre-quanitzed checkpoint in plain FP8
+      repo: michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm
+      source: HF
+    max_batch_size: 64
+    max_seq_len: 65536
+    quantization_type: fp8
+    tensor_parallel_count: 8
+    speculator:
+      lookahead_ngram_size: 5
+      lookahead_verification_set_size: 5
+      lookahead_windows_size: 7
+      num_draft_tokens: 47
+      speculative_decoding_mode: LOOKAHEAD_DECODING
diff --git a/llm/nsql/README.md b/llm/nsql/README.md
new file mode 100644
index 000000000..67614f44e
--- /dev/null
+++ b/llm/nsql/README.md
@@ -0,0 +1,29 @@
+# NSQL 350M
+
+NSQL is an open-source text-to-SQL AI model developed by Numbers Station.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/nsql/config.yaml b/llm/nsql/config.yaml
new file mode 100644
index 000000000..85a08f956
--- /dev/null
+++ b/llm/nsql/config.yaml
@@ -0,0 +1,22 @@
+description: NSQL is an open-source text-to-SQL AI model developed by Numbers Station.
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "NumbersStation/nsql-350M"
+  avatar_url: https://aeiljuispo.cloudimg.io/v7/https://cdn-uploads.huggingface.co/production/uploads/649c7ee8f97bd6fd710a9eb5/nBg1Fyo22RrqRJrkz9IYB.png
+  cover_image_url: https://global-uploads.webflow.com/6348b2d49808811e3f7a0fff/640690727b722a05771960ec_graphic-data-p-800.png
+  tags:
+  - code-generation
+  example_model_input: {"schema": "CREATE TABLE stadium (stadium_id number, location text, name text, capacity number)", "query": "What is the total capacity of all stadiums?"}
+model_name: NSQL 350M
+python_version: py39
+requirements:
+- torch==2.1.0
+- transformers>=4.29.0
+resources:
+  accelerator: A10G
+  cpu: '8'
+  memory: 30Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/templates/trt-llm/model/__init__.py b/llm/nsql/model/__init__.py
similarity index 100%
rename from templates/trt-llm/model/__init__.py
rename to llm/nsql/model/__init__.py
diff --git a/nsql/model/model.py b/llm/nsql/model/model.py
similarity index 100%
rename from nsql/model/model.py
rename to llm/nsql/model/model.py
diff --git a/llm/openai/gpt-oss-120b/README.md b/llm/openai/gpt-oss-120b/README.md
new file mode 100644
index 000000000..8de0d681b
--- /dev/null
+++ b/llm/openai/gpt-oss-120b/README.md
@@ -0,0 +1,54 @@
+# GPT OSS 120B
+
+Deploy [openai/gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [openai/gpt-oss-120b](https://huggingface.co/openai/gpt-oss-120b) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="openai/gpt-oss-120b",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "openai/gpt-oss-120b", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Model cache: **volume-mounted** for fast cold starts
+- Streaming: **enabled**
diff --git a/llm/openai/gpt-oss-120b/config.yaml b/llm/openai/gpt-oss-120b/config.yaml
new file mode 100644
index 000000000..b77dfc192
--- /dev/null
+++ b/llm/openai/gpt-oss-120b/config.yaml
@@ -0,0 +1,70 @@
+description: "openai/gpt-oss-120b for text generation"
+model_name: GPT OSS 120B
+build_commands:
+  - python -c 'from openai_harmony import load_harmony_encoding; load_harmony_encoding("HarmonyGptOss")'
+model_metadata:
+  repo_id: openai/gpt-oss-120b
+  example_model_input: {
+    "model": "openai/gpt-oss-120b",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    "stream": true,
+    "max_tokens": 2048,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+model_cache:
+  - repo_id: openai/gpt-oss-120b
+    revision: refs/pr/35
+    use_volume: true
+    volume_folder: trt_model
+trt_llm:
+  build:
+    checkpoint_repository:
+      repo: michaelfeil/empty-model
+      revision: main
+      source: HF
+  inference_stack: v2
+  runtime:
+    enable_chunked_prefill: true
+    max_batch_size: 64
+    max_num_tokens: 8192
+    max_seq_len: 131072
+    patch_kwargs:
+      model_path: /app/model_cache/trt_model
+      chat_processor: harmony
+      moe_expert_parallel_size: 1
+      backend: pytorch
+      cuda_graph_config:
+        enable_padding: true
+      disable_overlap_scheduler: 1
+      enable_autotuner: 0
+      enable_iter_perf_stats: 0
+      enable_trtllm_sampler: 1
+      guided_decoding_backend: xgrammar
+      kv_cache_config:
+        enable_block_reuse: true
+        free_gpu_memory_fraction: 0.8
+        event_buffer_max_size: 1024
+      max_beam_width: 1
+      max_input_len: 131072
+      model_level_stop_words:
+      - "<|call|>"
+      tokenizer_limit_length: 131072
+      trust_remote_code: 1
+      moe_config:
+        backend: TRTLLM
+    served_model_name: openai/gpt-oss-120b
+    tensor_parallel_size: 1
+  version_overrides:
+    v2_llm_version: null
diff --git a/llm/openai/gpt-oss-20b/README.md b/llm/openai/gpt-oss-20b/README.md
new file mode 100644
index 000000000..88ea7ab1c
--- /dev/null
+++ b/llm/openai/gpt-oss-20b/README.md
@@ -0,0 +1,54 @@
+# GPT OSS 20B
+
+Deploy [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [openai/gpt-oss-20b](https://huggingface.co/openai/gpt-oss-20b) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="openai/gpt-oss-20b",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "openai/gpt-oss-20b", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Model cache: **volume-mounted** for fast cold starts
+- Streaming: **enabled**
diff --git a/llm/openai/gpt-oss-20b/config.yaml b/llm/openai/gpt-oss-20b/config.yaml
new file mode 100644
index 000000000..cc06a36ad
--- /dev/null
+++ b/llm/openai/gpt-oss-20b/config.yaml
@@ -0,0 +1,70 @@
+description: "openai/gpt-oss-20b for text generation"
+model_name: GPT OSS 20B
+build_commands:
+  - python -c 'from openai_harmony import load_harmony_encoding; load_harmony_encoding("HarmonyGptOss")'
+model_metadata:
+  repo_id: openai/gpt-oss-20b
+  example_model_input: {
+    "model": "openai/gpt-oss-20b",
+    "messages": [
+      {
+      "role": "user",
+      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
+      }
+    ],
+    "stream": true,
+    "max_tokens": 2048,
+    "temperature": 0.5
+  }
+  tags:
+  - openai-compatible
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+model_cache:
+  - repo_id: openai/gpt-oss-20b
+    revision: refs/pr/36
+    use_volume: true
+    volume_folder: trt_model
+trt_llm:
+  build:
+    checkpoint_repository:
+      repo: michaelfeil/empty-model
+      revision: main
+      source: HF
+  inference_stack: v2
+  runtime:
+    enable_chunked_prefill: true
+    max_batch_size: 64
+    max_num_tokens: 8192
+    max_seq_len: 131072
+    patch_kwargs:
+      model_path: /app/model_cache/trt_model
+      chat_processor: harmony
+      moe_expert_parallel_size: 1
+      backend: pytorch
+      cuda_graph_config:
+        enable_padding: true
+      disable_overlap_scheduler: 1
+      enable_autotuner: 0
+      enable_iter_perf_stats: 0
+      enable_trtllm_sampler: 1
+      guided_decoding_backend: xgrammar
+      kv_cache_config:
+        enable_block_reuse: true
+        free_gpu_memory_fraction: 0.8
+        event_buffer_max_size: 1024
+      max_beam_width: 1
+      max_input_len: 131072
+      model_level_stop_words:
+      - "<|call|>"
+      tokenizer_limit_length: 131072
+      trust_remote_code: 1
+      moe_config:
+        backend: CUTLASS
+    served_model_name: openai/gpt-oss-20b
+    tensor_parallel_size: 1
+  version_overrides:
+    v2_llm_version: null
diff --git a/personaplex-7b-v1/Dockerfile b/llm/personaplex-7b-v1/Dockerfile
similarity index 100%
rename from personaplex-7b-v1/Dockerfile
rename to llm/personaplex-7b-v1/Dockerfile
diff --git a/llm/personaplex-7b-v1/README.md b/llm/personaplex-7b-v1/README.md
new file mode 100644
index 000000000..4f335ca9b
--- /dev/null
+++ b/llm/personaplex-7b-v1/README.md
@@ -0,0 +1,31 @@
+# Personaplex 7V H100
+
+Speech-to-speech model powered by NVIDIA Personaplex
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/personaplex-7b-v1](https://huggingface.co/nvidia/personaplex-7b-v1) |
+| Task | Text generation |
+| Engine | Docker Server |
+| GPU | H100 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/api/chat \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `basetenservice/personaplex-7v:fork`
diff --git a/personaplex-7b-v1/client.py b/llm/personaplex-7b-v1/client.py
similarity index 100%
rename from personaplex-7b-v1/client.py
rename to llm/personaplex-7b-v1/client.py
diff --git a/llm/personaplex-7b-v1/config.yaml b/llm/personaplex-7b-v1/config.yaml
new file mode 100644
index 000000000..544001ed0
--- /dev/null
+++ b/llm/personaplex-7b-v1/config.yaml
@@ -0,0 +1,23 @@
+description: Speech-to-speech model powered by NVIDIA Personaplex
+base_image:
+  image: basetenservice/personaplex-7v:fork
+model_metadata:
+  repo_id: nvidia/personaplex-7b-v1
+  tags:
+    - speech-to-speech
+  example_model_input: {"message": "Hello, how are you today?"}
+docker_server:
+  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m moshi.server --host 0.0.0.0 --port 8998"
+  readiness_endpoint: /
+  liveness_endpoint: /
+  predict_endpoint: /api/chat
+  server_port: 8998
+resources:
+  accelerator: H100
+  use_gpu: true
+model_name: Personaplex 7V H100
+secrets:
+  hf_access_token: null
+runtime:
+  transport:
+    kind: websocket
diff --git a/personaplex-7b-v1/requirements-client.txt b/llm/personaplex-7b-v1/requirements-client.txt
similarity index 100%
rename from personaplex-7b-v1/requirements-client.txt
rename to llm/personaplex-7b-v1/requirements-client.txt
diff --git a/llm/phi/phi-3-mini-128k-instruct/README.md b/llm/phi/phi-3-mini-128k-instruct/README.md
new file mode 100644
index 000000000..f0ca30149
--- /dev/null
+++ b/llm/phi/phi-3-mini-128k-instruct/README.md
@@ -0,0 +1,29 @@
+# Phi-3-Mini-128K-Instruct
+
+Deploy Phi-3-Mini-128K-Instruct for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/phi/phi-3-mini-128k-instruct/config.yaml b/llm/phi/phi-3-mini-128k-instruct/config.yaml
new file mode 100644
index 000000000..918ffab65
--- /dev/null
+++ b/llm/phi/phi-3-mini-128k-instruct/config.yaml
@@ -0,0 +1,18 @@
+description: "Phi-3-Mini-128K-Instruct for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "microsoft/Phi-3-mini-128k-instruct"
+  example_model_input: {"messages": [{"role": "user", "content": "What is the meaning of life?"}]}
+model_name: Phi-3-Mini-128K-Instruct
+python_version: py39
+requirements:
+  - accelerate==0.30.1
+  - einops==0.8.0
+  - transformers==4.40.1
+  - torch==2.3.0
+resources:
+  accelerator: T4
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/model/__init__.py b/llm/phi/phi-3-mini-128k-instruct/model/__init__.py
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/high_throughput/model/__init__.py
rename to llm/phi/phi-3-mini-128k-instruct/model/__init__.py
diff --git a/phi/phi-3-mini-128k-instruct/model/model.py b/llm/phi/phi-3-mini-128k-instruct/model/model.py
similarity index 100%
rename from phi/phi-3-mini-128k-instruct/model/model.py
rename to llm/phi/phi-3-mini-128k-instruct/model/model.py
diff --git a/llm/phi/phi-3-mini-4k-instruct/README.md b/llm/phi/phi-3-mini-4k-instruct/README.md
new file mode 100644
index 000000000..b40a43542
--- /dev/null
+++ b/llm/phi/phi-3-mini-4k-instruct/README.md
@@ -0,0 +1,29 @@
+# Phi-3-Mini-4K-Instruct
+
+Deploy Phi-3-Mini-4K-Instruct for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | T4 |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/phi/phi-3-mini-4k-instruct/config.yaml b/llm/phi/phi-3-mini-4k-instruct/config.yaml
new file mode 100644
index 000000000..9bb07a3fd
--- /dev/null
+++ b/llm/phi/phi-3-mini-4k-instruct/config.yaml
@@ -0,0 +1,18 @@
+description: "Phi-3-Mini-4K-Instruct for text generation"
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "microsoft/Phi-3-mini-4k-instruct"
+  example_model_input: {"messages": [{"role": "user", "content": "What is the meaning of life?"}]}
+model_name: Phi-3-Mini-4K-Instruct
+python_version: py39
+requirements:
+  - accelerate==0.30.1
+  - einops==0.8.0
+  - transformers==4.40.1
+  - torch==2.3.0
+resources:
+  accelerator: T4
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/model/__init__.py b/llm/phi/phi-3-mini-4k-instruct/model/__init__.py
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/large_context/model/__init__.py
rename to llm/phi/phi-3-mini-4k-instruct/model/__init__.py
diff --git a/phi/phi-3-mini-4k-instruct/model/model.py b/llm/phi/phi-3-mini-4k-instruct/model/model.py
similarity index 100%
rename from phi/phi-3-mini-4k-instruct/model/model.py
rename to llm/phi/phi-3-mini-4k-instruct/model/model.py
diff --git a/llm/phi/phi-3.5-mini/README.md b/llm/phi/phi-3.5-mini/README.md
new file mode 100644
index 000000000..2f6bbb085
--- /dev/null
+++ b/llm/phi/phi-3.5-mini/README.md
@@ -0,0 +1,39 @@
+# Phi 3.5 Mini Instruct VLLM openai compatible
+
+Deploy [microsoft/Phi-3.5-mini-instruct](https://huggingface.co/microsoft/Phi-3.5-mini-instruct) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [microsoft/Phi-3.5-mini-instruct](https://huggingface.co/microsoft/Phi-3.5-mini-instruct) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "messages": [
+    {
+      "role": "user",
+      "content": "what is the meaning of life"
+    }
+  ]
+}'
+```
+
+## Configuration highlights
+
+- Predict concurrency: **128**
diff --git a/llm/phi/phi-3.5-mini/config.yaml b/llm/phi/phi-3.5-mini/config.yaml
new file mode 100644
index 000000000..68f2b7062
--- /dev/null
+++ b/llm/phi/phi-3.5-mini/config.yaml
@@ -0,0 +1,19 @@
+description: "microsoft/Phi-3.5-mini-instruct for text generation"
+model_name: "Phi 3.5 Mini Instruct VLLM openai compatible"
+python_version: py311
+model_metadata:
+  example_model_input: {"messages": [{"role": "user", "content": "what is the meaning of life"}]}
+  repo_id: microsoft/Phi-3.5-mini-instruct
+  openai_compatible: true
+  vllm_config:
+    tensor_parallel_size: 1
+    max_model_len: 10000
+requirements:
+  - vllm==0.5.4
+resources:
+  accelerator: A10G
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+secrets:
+  hf_access_token: null
diff --git a/trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/model/__init__.py b/llm/phi/phi-3.5-mini/model/__init__.py
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-70b-instruct/low_ttft/model/__init__.py
rename to llm/phi/phi-3.5-mini/model/__init__.py
diff --git a/phi/phi-3.5-mini/model/helper.py b/llm/phi/phi-3.5-mini/model/helper.py
similarity index 100%
rename from phi/phi-3.5-mini/model/helper.py
rename to llm/phi/phi-3.5-mini/model/helper.py
diff --git a/phi/phi-3.5-mini/model/model.py b/llm/phi/phi-3.5-mini/model/model.py
similarity index 100%
rename from phi/phi-3.5-mini/model/model.py
rename to llm/phi/phi-3.5-mini/model/model.py
diff --git a/qwen/qwen-7b-chat/README.md b/llm/qwen/_archive/qwen-7b-chat/README.md
similarity index 100%
rename from qwen/qwen-7b-chat/README.md
rename to llm/qwen/_archive/qwen-7b-chat/README.md
diff --git a/qwen/qwen-7b-chat/config.yaml b/llm/qwen/_archive/qwen-7b-chat/config.yaml
similarity index 100%
rename from qwen/qwen-7b-chat/config.yaml
rename to llm/qwen/_archive/qwen-7b-chat/config.yaml
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/model/__init__.py b/llm/qwen/_archive/qwen-7b-chat/model/__init__.py
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/high_throughput/model/__init__.py
rename to llm/qwen/_archive/qwen-7b-chat/model/__init__.py
diff --git a/qwen/qwen-7b-chat/model/model.py b/llm/qwen/_archive/qwen-7b-chat/model/model.py
similarity index 100%
rename from qwen/qwen-7b-chat/model/model.py
rename to llm/qwen/_archive/qwen-7b-chat/model/model.py
diff --git a/llm/qwen/engine-qwen-2-5-14b-coder-instruct/README.md b/llm/qwen/engine-qwen-2-5-14b-coder-instruct/README.md
new file mode 100644
index 000000000..bae332e8a
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-14b-coder-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen Coder 2.5 14B Instruct
+
+Deploy [Qwen/Qwen2.5-Coder-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Coder-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-14B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Coder-14B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Coder-14B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-14b-coder-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-14b-coder-instruct/config.yaml
new file mode 100644
index 000000000..09ffacb03
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-14b-coder-instruct/config.yaml
@@ -0,0 +1,50 @@
+description: "Qwen 2.5 Coder 14B Instruct for code generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+          },
+          { role: "user", content: "Write a Python script for fizzbuzz." },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-Coder-14B-Instruct
+model_name: Qwen Coder 2.5 14B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Coder-14B-Instruct
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-14b-instruct/README.md b/llm/qwen/engine-qwen-2-5-14b-instruct/README.md
new file mode 100644
index 000000000..cc997098a
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-14b-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen 2.5 14B Instruct
+
+Deploy [Qwen/Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-14B-Instruct](https://huggingface.co/Qwen/Qwen2.5-14B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-14B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-14B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-14b-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-14b-instruct/config.yaml
new file mode 100644
index 000000000..fab818e2d
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-14b-instruct/config.yaml
@@ -0,0 +1,45 @@
+description: "Qwen/Qwen2.5-14B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    max_tokens: 512
+    messages:
+      - content: You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
+        role: system
+      - content: What does Tongyi Qianwen mean?
+        role: user
+    stream: true
+    temperature: 0.9
+  repo_id: Qwen/Qwen2.5-14B-Instruct
+model_name: Qwen 2.5 14B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-14B-Instruct
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-32b-coder-instruct/README.md b/llm/qwen/engine-qwen-2-5-32b-coder-instruct/README.md
new file mode 100644
index 000000000..343b1ab34
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-32b-coder-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen Coder 2.5 32B Instruct
+
+Deploy [Qwen/Qwen2.5-Coder-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Coder-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-32B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Coder-32B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Coder-32B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-32b-coder-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-32b-coder-instruct/config.yaml
new file mode 100644
index 000000000..408ae3d02
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-32b-coder-instruct/config.yaml
@@ -0,0 +1,50 @@
+description: "Qwen 2.5 Coder 32B Instruct for code generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+          },
+          { role: "user", content: "Write a Python script for fizzbuzz." },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-Coder-32B-Instruct
+model_name: Qwen Coder 2.5 32B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Coder-32B-Instruct
+      source: HF
+    num_builder_gpus: 2
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-32b-instruct/README.md b/llm/qwen/engine-qwen-2-5-32b-instruct/README.md
new file mode 100644
index 000000000..10a592d7c
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-32b-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen 2.5 32B Instruct
+
+Deploy [Qwen/Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-32B-Instruct](https://huggingface.co/Qwen/Qwen2.5-32B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-32B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-32B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-32b-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-32b-instruct/config.yaml
new file mode 100644
index 000000000..ff926bcb9
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-32b-instruct/config.yaml
@@ -0,0 +1,50 @@
+description: "Qwen/Qwen2.5-32B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+          },
+          { role: "user", content: "What does Tongyi Qianwen mean?" },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-32B-Instruct
+model_name: Qwen 2.5 32B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-32B-Instruct
+      source: HF
+    num_builder_gpus: 2
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-3b-instruct/README.md b/llm/qwen/engine-qwen-2-5-3b-instruct/README.md
new file mode 100644
index 000000000..dde39acc9
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-3b-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen 2.5 3B Instruct
+
+Deploy [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-3B-Instruct](https://huggingface.co/Qwen/Qwen2.5-3B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | A10G |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-3b-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-3b-instruct/config.yaml
new file mode 100644
index 000000000..c5a0ffc2a
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-3b-instruct/config.yaml
@@ -0,0 +1,50 @@
+description: "Qwen/Qwen2.5-3B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+          },
+          { role: "user", content: "What does Tongyi Qianwen mean?" },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-3B-Instruct
+model_name: Qwen 2.5 3B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: A10G
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-3B-Instruct
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-72b-instruct/README.md b/llm/qwen/engine-qwen-2-5-72b-instruct/README.md
new file mode 100644
index 000000000..de37f4f18
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-72b-instruct/README.md
@@ -0,0 +1,62 @@
+# Qwen 2.5 72B Instruct
+
+Deploy [Qwen/Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-72B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-72B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-72b-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-72b-instruct/config.yaml
new file mode 100644
index 000000000..bec27fb19
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-72b-instruct/config.yaml
@@ -0,0 +1,50 @@
+description: "Qwen/Qwen2.5-72B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+          },
+          { role: "user", content: "What does Tongyi Qianwen mean?" },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-72B-Instruct
+model_name: Qwen 2.5 72B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:2
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-72B-Instruct
+      source: HF
+    num_builder_gpus: 4
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 2
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-72b-math-instruct/README.md b/llm/qwen/engine-qwen-2-5-72b-math-instruct/README.md
new file mode 100644
index 000000000..873bd2455
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-72b-math-instruct/README.md
@@ -0,0 +1,62 @@
+# Qwen Math 2.5 72B Instruct
+
+Deploy [Qwen/Qwen2.5-Math-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Math-72B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Math-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Math-72B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Math-72B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Math-72B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-72b-math-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-72b-math-instruct/config.yaml
new file mode 100644
index 000000000..0f5dc2130
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-72b-math-instruct/config.yaml
@@ -0,0 +1,53 @@
+description: "Qwen 2.5 Math 72B Instruct for mathematical reasoning"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "Please reason step by step, and put your final answer within \\boxed{}.",
+          },
+          {
+            role: "user",
+            content: "Find the value of $x$ that satisfies the equation $4x+5 = 6x+7$.",
+          },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-Math-72B-Instruct
+model_name: Qwen Math 2.5 72B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100:2
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Math-72B-Instruct
+      source: HF
+    num_builder_gpus: 4
+    quantization_type: fp8
+    max_seq_len: 32768
+    tensor_parallel_count: 2
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-7b-coder-instruct/README.md b/llm/qwen/engine-qwen-2-5-7b-coder-instruct/README.md
new file mode 100644
index 000000000..b42074182
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-7b-coder-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen Coder 2.5 7B Instruct
+
+Deploy [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Coder-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Coder-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-7b-coder-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-7b-coder-instruct/config.yaml
new file mode 100644
index 000000000..4b59cd31d
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-7b-coder-instruct/config.yaml
@@ -0,0 +1,50 @@
+description: "Qwen 2.5 Coder 7B Instruct for code generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+          },
+          { role: "user", content: "Write a Python script for fizzbuzz." },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-Coder-7B-Instruct
+model_name: Qwen Coder 2.5 7B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Coder-7B-Instruct
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-7b-instruct/README.md b/llm/qwen/engine-qwen-2-5-7b-instruct/README.md
new file mode 100644
index 000000000..95718dcde
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-7b-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen 2.5 7B Instruct
+
+Deploy [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-7b-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-7b-instruct/config.yaml
new file mode 100644
index 000000000..c465bff47
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-7b-instruct/config.yaml
@@ -0,0 +1,50 @@
+description: "Qwen/Qwen2.5-7B-Instruct for text generation"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
+          },
+          { role: "user", content: "What does Tongyi Qianwen mean?" },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-7B-Instruct
+model_name: Qwen 2.5 7B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-7B-Instruct
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-2-5-7b-math-instruct/README.md b/llm/qwen/engine-qwen-2-5-7b-math-instruct/README.md
new file mode 100644
index 000000000..abd86e80c
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-7b-math-instruct/README.md
@@ -0,0 +1,61 @@
+# Qwen Math 2.5 7B Instruct
+
+Deploy [Qwen/Qwen2.5-Math-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Math-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Math-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Math-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Math-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_paged_context_fmha**
+- Plugin: **paged_kv_cache**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-2-5-7b-math-instruct/config.yaml b/llm/qwen/engine-qwen-2-5-7b-math-instruct/config.yaml
new file mode 100644
index 000000000..697d2362d
--- /dev/null
+++ b/llm/qwen/engine-qwen-2-5-7b-math-instruct/config.yaml
@@ -0,0 +1,53 @@
+description: "Qwen 2.5 Math 7B Instruct for mathematical reasoning"
+build_commands: []
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  tags:
+    - openai-compatible
+  example_model_input:
+    {
+      messages:
+        [
+          {
+            role: "system",
+            content: "Please reason step by step, and put your final answer within \\boxed{}.",
+          },
+          {
+            role: "user",
+            content: "Find the value of $x$ that satisfies the equation $4x+5 = 6x+7$.",
+          },
+        ],
+      stream: true,
+      max_tokens: 512,
+      temperature: 0.9,
+    }
+  repo_id: Qwen/Qwen2.5-Math-7B-Instruct
+model_name: Qwen Math 2.5 7B Instruct
+python_version: py39
+requirements: []
+resources:
+  accelerator: H100_40GB
+  cpu: "1"
+  memory: 24Gi
+  use_gpu: true
+secrets: {}
+system_packages: []
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Math-7B-Instruct
+      source: HF
+    num_builder_gpus: 1
+    quantization_type: no_quant
+    max_seq_len: 32768
+    tensor_parallel_count: 1
+    plugin_configuration:
+      use_paged_context_fmha: true
+      use_fp8_context_fmha: false
+      paged_kv_cache: true
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
+    request_default_max_tokens: 32768
diff --git a/llm/qwen/engine-qwen-3-06b/README.md b/llm/qwen/engine-qwen-3-06b/README.md
new file mode 100644
index 000000000..d7e0d5a34
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-06b/README.md
@@ -0,0 +1,59 @@
+# library-model-qwen3-06b-engine
+
+Deploy [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-0.6B](https://huggingface.co/Qwen/Qwen3-0.6B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-0.6B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-0.6B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-3-06b/config.yaml b/llm/qwen/engine-qwen-3-06b/config.yaml
new file mode 100644
index 000000000..8da3fe294
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-06b/config.yaml
@@ -0,0 +1,41 @@
+description: "Qwen/Qwen3-0.6B for text generation"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+      - content: Tell me everything you know about optimized inference.
+        role: user
+    stream: true
+    temperature: 0.5
+    chat_template_kwargs:
+      enable_thinking: true
+  tags:
+    - openai-compatible
+model_name: library-model-qwen3-06b-engine
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    checkpoint_repository:
+      repo: Qwen/Qwen3-0.6B
+      revision: main
+      source: HF
+    max_batch_size: 64
+    num_builder_gpus: 1
+    max_seq_len: 40960
+    # plugin_configuration:
+    #   use_fp8_context_fmha: true
+    quantization_type: fp8
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 16
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/llm/qwen/engine-qwen-3-30b-a3b-instruct-2507/README.md b/llm/qwen/engine-qwen-3-30b-a3b-instruct-2507/README.md
new file mode 100644
index 000000000..49a09fb3e
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-30b-a3b-instruct-2507/README.md
@@ -0,0 +1,62 @@
+# qwen3-30b-a3b-instruct-2507-fp8_kv
+
+Deploy [Qwen/Qwen3-30B-A3B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-30B-A3B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-30B-A3B-Instruct-2507",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-30B-A3B-Instruct-2507", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-3-30b-a3b-instruct-2507/config.yaml b/llm/qwen/engine-qwen-3-30b-a3b-instruct-2507/config.yaml
new file mode 100644
index 000000000..ece2a67c2
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-30b-a3b-instruct-2507/config.yaml
@@ -0,0 +1,38 @@
+description: "Qwen/Qwen3-30B-A3B-Instruct-2507 for text generation"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+      - content: Tell me everything you know about optimized inference.
+        role: user
+    stream: true
+    temperature: 0.5
+    chat_template_kwargs:
+      enable_thinking: false
+  tags:
+    - openai-compatible
+model_name: qwen3-30b-a3b-instruct-2507-fp8_kv
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+secrets:
+  hf_access_token: null
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-30B-A3B-Instruct-2507
+      revision: main
+      source: HF
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/llm/qwen/engine-qwen-3-32b/README.md b/llm/qwen/engine-qwen-3-32b/README.md
new file mode 100644
index 000000000..3f9eea508
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-32b/README.md
@@ -0,0 +1,59 @@
+# library-model-qwen3-32B-engine
+
+Deploy [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-3-32b/config.yaml b/llm/qwen/engine-qwen-3-32b/config.yaml
new file mode 100644
index 000000000..150c0bc5c
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-32b/config.yaml
@@ -0,0 +1,41 @@
+description: "Qwen/Qwen3-32B for text generation"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+      - content: Tell me everything you know about optimized inference.
+        role: user
+    stream: true
+    temperature: 0.5
+    chat_template_kwargs:
+      enable_thinking: true
+  tags:
+    - openai-compatible
+model_name: library-model-qwen3-32B-engine
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    checkpoint_repository:
+      repo: Qwen/Qwen3-32B
+      revision: main
+      source: HF
+    max_batch_size: 64
+    num_builder_gpus: 1
+    max_seq_len: 40960
+    # plugin_configuration:
+    #   use_fp8_context_fmha: true
+    quantization_type: fp8
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 16
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/llm/qwen/engine-qwen-3-4b/README.md b/llm/qwen/engine-qwen-3-4b/README.md
new file mode 100644
index 000000000..9994df61a
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-4b/README.md
@@ -0,0 +1,59 @@
+# library-model-qwen3-4b-engine
+
+Deploy [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-4B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-4B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/llm/qwen/engine-qwen-3-4b/config.yaml b/llm/qwen/engine-qwen-3-4b/config.yaml
new file mode 100644
index 000000000..cc2d3dfd1
--- /dev/null
+++ b/llm/qwen/engine-qwen-3-4b/config.yaml
@@ -0,0 +1,41 @@
+description: "Qwen/Qwen3-4B for text generation"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+      - content: Tell me everything you know about optimized inference.
+        role: user
+    stream: true
+    temperature: 0.5
+    chat_template_kwargs:
+      enable_thinking: true
+  tags:
+    - openai-compatible
+model_name: library-model-qwen3-4b-engine
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    checkpoint_repository:
+      repo: Qwen/Qwen3-4B
+      revision: main
+      source: HF
+    max_batch_size: 64
+    num_builder_gpus: 1
+    max_seq_len: 40960
+    # plugin_configuration:
+    #   use_fp8_context_fmha: true
+    quantization_type: fp8
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 16
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/qwen/model_auto.py b/llm/qwen/model_auto.py
similarity index 100%
rename from qwen/model_auto.py
rename to llm/qwen/model_auto.py
diff --git a/llm/qwen/qwen-3-235B-A22B-instruct-2507-trt/README.md b/llm/qwen/qwen-3-235B-A22B-instruct-2507-trt/README.md
new file mode 100644
index 000000000..5b4c57cb7
--- /dev/null
+++ b/llm/qwen/qwen-3-235B-A22B-instruct-2507-trt/README.md
@@ -0,0 +1,53 @@
+# Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
+
+Deploy [Qwen/Qwen3-235B-A22B-Instruct-2507-FP8](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-235B-A22B-Instruct-2507-FP8](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507-FP8) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-235B-A22B-Instruct-2507-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Model cache: **volume-mounted** for fast cold starts
diff --git a/llm/qwen/qwen-3-235B-A22B-instruct-2507-trt/config.yaml b/llm/qwen/qwen-3-235B-A22B-instruct-2507-trt/config.yaml
new file mode 100644
index 000000000..9cae82ce2
--- /dev/null
+++ b/llm/qwen/qwen-3-235B-A22B-instruct-2507-trt/config.yaml
@@ -0,0 +1,55 @@
+description: "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8 for text generation"
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What does Tongyi Qianwen mean?"
+    stream: false
+    model: "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8"
+    max_tokens: 512
+    temperature: 0.6
+  tags:
+    - openai-compatible
+  repo_id: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
+model_name: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
+model_cache:
+  - repo_id: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
+    use_volume: true
+    revision: main
+    volume_folder: trt_model
+resources:
+  accelerator: H100:8
+  cpu: "1"
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    checkpoint_repository:
+      repo: michaelfeil/empty-model
+      revision: main
+      source: HF
+  inference_stack: v2
+  runtime:
+    enable_chunked_prefill: true
+    max_batch_size: 256
+    max_num_tokens: 8192
+    max_seq_len: 262144
+    served_model_name: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
+    tensor_parallel_size: 8
+    patch_kwargs:
+      disable_overlap_scheduler: True
+      model_path: /app/model_cache/trt_model
+      moe_expert_parallel_size: 8
+      cuda_graph_config:
+        enable_padding: true
+        max_batch_size: 256
+      enable_autotune: false
+      guided_decoding_backend: "xgrammar"
+      enable_iter_perf_stats: 0
+      kv_cache_config:
+        enable_block_reuse: true
+        free_gpu_memory_fraction: 0.8
+  version_overrides:
+    v2_llm_version: null
diff --git a/llm/qwen/qwen-3-235B-sglang/README.md b/llm/qwen/qwen-3-235B-sglang/README.md
new file mode 100644
index 000000000..881b1d58a
--- /dev/null
+++ b/llm/qwen/qwen-3-235B-sglang/README.md
@@ -0,0 +1,56 @@
+# Qwen 3 235B SGLang
+
+Deploy [Qwen/Qwen3-235B-A22B-FP8](https://huggingface.co/Qwen/Qwen3-235B-A22B-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-235B-A22B-FP8](https://huggingface.co/Qwen/Qwen3-235B-A22B-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:4 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-235B-A22B-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-235B-A22B-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.6.post1-cu124`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/qwen/qwen-3-235B-sglang/config.yaml b/llm/qwen/qwen-3-235B-sglang/config.yaml
new file mode 100644
index 000000000..2a8825c22
--- /dev/null
+++ b/llm/qwen/qwen-3-235B-sglang/config.yaml
@@ -0,0 +1,36 @@
+description: "Qwen/Qwen3-235B-A22B-FP8 for text generation"
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What does Tongyi Qianwen mean?"
+    stream: true
+    model: "Qwen/Qwen3-235B-A22B"
+    max_tokens: 32768
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: Qwen 3 235B SGLang
+base_image:
+  image: lmsysorg/sglang:v0.4.6.post1-cu124
+model_cache:
+  - repo_id: Qwen/Qwen3-235B-A22B-FP8
+    revision: 57c8978fa7d601431cfd6750dd7355b5cdfa5a18
+    use_volume: true
+    volume_folder: "qwen3"
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+docker_server:
+  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3 --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-235B-A22B --tp 4 --reasoning-parser qwen3"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:4
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/qwen/qwen-3-30B-A3-coder/README.md b/llm/qwen/qwen-3-30B-A3-coder/README.md
new file mode 100644
index 000000000..bfe7202b5
--- /dev/null
+++ b/llm/qwen/qwen-3-30B-A3-coder/README.md
@@ -0,0 +1,56 @@
+# Qwen 3 Coder
+
+Deploy [Qwen/Qwen3-Coder-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-Coder-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-Coder-30B-A3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-Coder-30B-A3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.10.post2-cu126`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/qwen/qwen-3-30B-A3-coder/config.yaml b/llm/qwen/qwen-3-30B-A3-coder/config.yaml
new file mode 100644
index 000000000..3cc4f7305
--- /dev/null
+++ b/llm/qwen/qwen-3-30B-A3-coder/config.yaml
@@ -0,0 +1,36 @@
+description: "Qwen3 Coder 30B A3B Instruct for code generation"
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write a quick sort algorithm."
+    stream: true
+    model: "Qwen/Qwen3-Coder-30B-A3B-Instruct"
+    max_tokens: 1024
+    temperature: 0.7
+  tags:
+    - openai-compatible
+model_name: Qwen 3 Coder
+base_image:
+  image: lmsysorg/sglang:v0.4.10.post2-cu126
+model_cache:
+  - repo_id: Qwen/Qwen3-Coder-30B-A3B-Instruct
+    revision: main
+    use_volume: true
+    volume_folder: "qwen3-coder"
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+docker_server:
+  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3-coder --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-Coder-30B-A3B-Instruct --tp 1 --reasoning-parser qwen3"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/qwen/qwen-3-30B-A3-sglang/README.md b/llm/qwen/qwen-3-30B-A3-sglang/README.md
new file mode 100644
index 000000000..eae0813fa
--- /dev/null
+++ b/llm/qwen/qwen-3-30B-A3-sglang/README.md
@@ -0,0 +1,57 @@
+# Qwen 3 30B-A3 SGLang
+
+Deploy [Qwen/Qwen3-30B-A3B-FP8](https://huggingface.co/Qwen/Qwen3-30B-A3B-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-30B-A3B-FP8](https://huggingface.co/Qwen/Qwen3-30B-A3B-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-30B-A3B-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-30B-A3B-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.6.post1-cu124`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
diff --git a/llm/qwen/qwen-3-30B-A3-sglang/config.yaml b/llm/qwen/qwen-3-30B-A3-sglang/config.yaml
new file mode 100644
index 000000000..70438bac2
--- /dev/null
+++ b/llm/qwen/qwen-3-30B-A3-sglang/config.yaml
@@ -0,0 +1,38 @@
+description: "Qwen/Qwen3-30B-A3B-FP8 for text generation"
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What does Tongyi Qianwen mean?"
+    stream: false
+    model: "Qwen/Qwen3-32B"
+    max_tokens: 512
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: Qwen 3 30B-A3 SGLang
+environment_variables:
+  hf_access_token: null
+base_image:
+  image: lmsysorg/sglang:v0.4.6.post1-cu124
+model_cache:
+  - repo_id: Qwen/Qwen3-30B-A3B-FP8
+    revision: 2daf1706ac267bae18c90a217a060817c0cebb66
+    use_volume: true
+    volume_folder: "qwen3"
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+docker_server:
+  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3 --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-30B-A3B --tp 1 --reasoning-parser qwen3"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/qwen/qwen-3-30B-A3-vllm/README.md b/llm/qwen/qwen-3-30B-A3-vllm/README.md
new file mode 100644
index 000000000..b5b637cf8
--- /dev/null
+++ b/llm/qwen/qwen-3-30B-A3-vllm/README.md
@@ -0,0 +1,55 @@
+# Qwen 3 30B-A3B vLLM
+
+Deploy [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-30B-A3B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-30B-A3B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.8.5`
+- Predict concurrency: **32**
+- Environment variables: `VLLM_LOGGING_LEVEL`
diff --git a/llm/qwen/qwen-3-30B-A3-vllm/config.yaml b/llm/qwen/qwen-3-30B-A3-vllm/config.yaml
new file mode 100644
index 000000000..c0e5db27d
--- /dev/null
+++ b/llm/qwen/qwen-3-30B-A3-vllm/config.yaml
@@ -0,0 +1,31 @@
+description: "Qwen/Qwen3-30B-A3B for text generation"
+base_image:
+  image: vllm/vllm-openai:v0.8.5
+docker_server:
+  start_command: sh -c "vllm serve Qwen/Qwen3-30B-A3B --enable-reasoning --reasoning-parser deepseek_r1 --served-model-name qwen30b --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+model_metadata:
+  repo_id: Qwen/Qwen3-30B-A3B
+  example_model_input:
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What does Tongyi Qianwen mean?"
+    stream: false
+    model: "qwen30b"
+    max_tokens: 512
+    temperature: 0.7
+  tags:
+    - openai-compatible
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
+model_name: Qwen 3 30B-A3B vLLM
+environment_variables:
+  VLLM_LOGGING_LEVEL: WARNING
diff --git a/llm/qwen/qwen-3-32B-sglang/README.md b/llm/qwen/qwen-3-32B-sglang/README.md
new file mode 100644
index 000000000..5b5e59aa6
--- /dev/null
+++ b/llm/qwen/qwen-3-32B-sglang/README.md
@@ -0,0 +1,56 @@
+# Qwen 3 32B SGLang
+
+Deploy [Qwen/Qwen3-32B-FP8](https://huggingface.co/Qwen/Qwen3-32B-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-32B-FP8](https://huggingface.co/Qwen/Qwen3-32B-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-32B-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-32B-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.6.post1-cu124`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/qwen/qwen-3-32B-sglang/config.yaml b/llm/qwen/qwen-3-32B-sglang/config.yaml
new file mode 100644
index 000000000..a4d88d4be
--- /dev/null
+++ b/llm/qwen/qwen-3-32B-sglang/config.yaml
@@ -0,0 +1,36 @@
+description: "Qwen/Qwen3-32B-FP8 for text generation"
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What does Tongyi Qianwen mean?"
+    stream: true
+    model: "Qwen/Qwen3-32B"
+    max_tokens: 32768
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: Qwen 3 32B SGLang
+base_image:
+  image: lmsysorg/sglang:v0.4.6.post1-cu124
+model_cache:
+  - repo_id: Qwen/Qwen3-32B-FP8
+    revision: 37f3f67a7a82b002377985796c57f4321b85fb9a
+    use_volume: true
+    volume_folder: "qwen3"
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+docker_server:
+  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3 --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-32B --tp 1 --reasoning-parser qwen3"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/qwen/qwen-3-next-80B-A3-instruct-sglang/README.md b/llm/qwen/qwen-3-next-80B-A3-instruct-sglang/README.md
new file mode 100644
index 000000000..657509ec5
--- /dev/null
+++ b/llm/qwen/qwen-3-next-80B-A3-instruct-sglang/README.md
@@ -0,0 +1,56 @@
+# Qwen3-Next-80B-A3B-Instruct
+
+Deploy [Qwen/Qwen3-Next-80B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-Next-80B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Instruct) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-Next-80B-A3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-Next-80B-A3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.5.3rc1-cu126`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **128**
+- Streaming: **enabled**
diff --git a/llm/qwen/qwen-3-next-80B-A3-instruct-sglang/config.yaml b/llm/qwen/qwen-3-next-80B-A3-instruct-sglang/config.yaml
new file mode 100644
index 000000000..4e1a6bcc7
--- /dev/null
+++ b/llm/qwen/qwen-3-next-80B-A3-instruct-sglang/config.yaml
@@ -0,0 +1,39 @@
+description: "Qwen/Qwen3-Next-80B-A3B-Instruct for text generation"
+base_image:
+  #image: lmsysorg/sglang@sha256:c977d3c5cf66029c8c37436a777381cb5bc861527da1b405c90ae2360417eedb
+  image: lmsysorg/sglang:v0.5.3rc1-cu126
+# build_commands:
+#   - pip install --upgrade pip
+#   - pip uninstall -y sglang
+#   - git clone https://github.com/sgl-project/sglang.git && cd sglang && pip install -e "python[all]"
+model_metadata:
+  repo_id: Qwen/Qwen3-Next-80B-A3B-Instruct
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write FizzBuzz in Python"
+    stream: true
+    model: "Qwen/Qwen3-Next-80B-A3B-Instruct"
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: sh -c "SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python3 -m sglang.launch_server --model-path Qwen/Qwen3-Next-80B-A3B-Instruct-FP8 --revision c5f5f263bdd5cc134092897864e8905d8fe7b928 --tp-size 2 --context-length 262144 --mem-fraction-static 0.8 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --served-model-name Qwen/Qwen3-Next-80B-A3B-Instruct --tool-call-parser qwen25 --host 0.0.0.0 --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+model_cache:
+  - repo_id: Qwen/Qwen3-Next-80B-A3B-Instruct-FP8
+    revision: c5f5f263bdd5cc134092897864e8905d8fe7b928
+    use_volume: true
+    volume_folder: qwen
+model_name: Qwen3-Next-80B-A3B-Instruct
diff --git a/llm/qwen/qwen-3-next-80B-A3-thinking-sglang/README.md b/llm/qwen/qwen-3-next-80B-A3-thinking-sglang/README.md
new file mode 100644
index 000000000..4ecc99334
--- /dev/null
+++ b/llm/qwen/qwen-3-next-80B-A3-thinking-sglang/README.md
@@ -0,0 +1,56 @@
+# Qwen3-Next-80B-A3B-Thinking
+
+Deploy [Qwen/Qwen3-Next-80B-A3B-Thinking-FP8](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-Next-80B-A3B-Thinking-FP8](https://huggingface.co/Qwen/Qwen3-Next-80B-A3B-Thinking-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-Next-80B-A3B-Thinking-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-Next-80B-A3B-Thinking-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.5.3rc1-cu126`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **128**
+- Streaming: **enabled**
diff --git a/llm/qwen/qwen-3-next-80B-A3-thinking-sglang/config.yaml b/llm/qwen/qwen-3-next-80B-A3-thinking-sglang/config.yaml
new file mode 100644
index 000000000..83015b226
--- /dev/null
+++ b/llm/qwen/qwen-3-next-80B-A3-thinking-sglang/config.yaml
@@ -0,0 +1,39 @@
+description: "Qwen/Qwen3-Next-80B-A3B-Thinking-FP8 for text generation"
+base_image:
+  #image: lmsysorg/sglang@sha256:c977d3c5cf66029c8c37436a777381cb5bc861527da1b405c90ae2360417eedb
+  image: lmsysorg/sglang:v0.5.3rc1-cu126
+# build_commands:
+#   - pip install --upgrade pip
+#   - pip uninstall -y sglang
+#   - git clone https://github.com/sgl-project/sglang.git && cd sglang && pip install -e "python[all]"
+model_metadata:
+  repo_id: Qwen/Qwen3-Next-80B-A3B-Thinking-FP8
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write FizzBuzz in Python"
+    stream: true
+    model: "Qwen/Qwen3-Next-80B-A3B-Thinking"
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: sh -c "SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python3 -m sglang.launch_server --model-path Qwen/Qwen3-Next-80B-A3B-Thinking-FP8 --revision 1a28d48a94e799860201879be67616b9e21c4bd2 --reasoning-parser qwen3-thinking --tp-size 2 --context-length 262144 --mem-fraction-static 0.8 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --served-model-name Qwen/Qwen3-Next-80B-A3B-Thinking --tool-call-parser qwen25 --host 0.0.0.0 --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+model_cache:
+  - repo_id: Qwen/Qwen3-Next-80B-A3B-Thinking-FP8
+    revision: 1a28d48a94e799860201879be67616b9e21c4bd2
+    use_volume: true
+    volume_folder: qwen
+model_name: Qwen3-Next-80B-A3B-Thinking
diff --git a/llm/qwen/qwen-3-vl-30b-a3b-instruct/README.md b/llm/qwen/qwen-3-vl-30b-a3b-instruct/README.md
new file mode 100644
index 000000000..47ec50bda
--- /dev/null
+++ b/llm/qwen/qwen-3-vl-30b-a3b-instruct/README.md
@@ -0,0 +1,54 @@
+# Qwen3-VL-30B-A3B-Instruct-FP8
+
+Deploy [Qwen/Qwen3-VL-30B-A3B-Instruct-FP8](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-VL-30B-A3B-Instruct-FP8](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Instruct-FP8) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-VL-30B-A3B-Instruct-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:2f7dbc9b42c51ba192e3dded515e4e07cdfdabea`
+- Predict concurrency: **32**
diff --git a/llm/qwen/qwen-3-vl-30b-a3b-instruct/config.yaml b/llm/qwen/qwen-3-vl-30b-a3b-instruct/config.yaml
new file mode 100644
index 000000000..dcc2adf3b
--- /dev/null
+++ b/llm/qwen/qwen-3-vl-30b-a3b-instruct/config.yaml
@@ -0,0 +1,36 @@
+description: "Qwen3 VL 30B A3B Instruct for vision-language tasks"
+model_metadata:
+  repo_id: "Qwen/Qwen3-VL-30B-A3B-Instruct-FP8"
+  example_model_input: # Loads sample request into Baseten playground
+    model: "Qwen/Qwen3-VL-30B-A3B-Thinking"
+    stream: false
+    max_tokens: 4096
+    messages:
+      - role: user
+        content:
+          - type: text
+            text: "What's in this image?"
+          - type: image_url
+            image_url:
+              url: "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true"
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: Qwen3-VL-30B-A3B-Instruct-FP8
+base_image:
+  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:2f7dbc9b42c51ba192e3dded515e4e07cdfdabea
+build_commands:
+  - pip install --pre --upgrade transformers
+  - pip uninstall -y vllm
+  - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git@d3d649efec8161b62e8db576f8d1d02a77d22897
+docker_server:
+  start_command: python3 -m vllm.entrypoints.openai.api_server --model Qwen/Qwen3-VL-30B-A3B-Instruct-FP8 --tool-call-parser hermes --reasoning-parser qwen3 --served-model-name Qwen/Qwen3-VL-30B-A3B-Instruct --enable-expert-parallel --enable-auto-tool-choice --tensor-parallel-size 2 --host 0.0.0.0 --port 8000
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/qwen/qwen-3-vl-30b-a3b-thinking/README.md b/llm/qwen/qwen-3-vl-30b-a3b-thinking/README.md
new file mode 100644
index 000000000..b81c4b27c
--- /dev/null
+++ b/llm/qwen/qwen-3-vl-30b-a3b-thinking/README.md
@@ -0,0 +1,54 @@
+# Qwen3-VL-30B-A3B-Thinking-FP8
+
+Deploy [Qwen/Qwen3-VL-30B-A3B-Thinking-FP8](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Thinking-FP8) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-VL-30B-A3B-Thinking-FP8](https://huggingface.co/Qwen/Qwen3-VL-30B-A3B-Thinking-FP8) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-VL-30B-A3B-Thinking-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-VL-30B-A3B-Thinking-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:2f7dbc9b42c51ba192e3dded515e4e07cdfdabea`
+- Predict concurrency: **32**
diff --git a/llm/qwen/qwen-3-vl-30b-a3b-thinking/config.yaml b/llm/qwen/qwen-3-vl-30b-a3b-thinking/config.yaml
new file mode 100644
index 000000000..347f77999
--- /dev/null
+++ b/llm/qwen/qwen-3-vl-30b-a3b-thinking/config.yaml
@@ -0,0 +1,36 @@
+description: "Qwen3 VL 30B A3B Thinking for vision-language tasks"
+model_metadata:
+  repo_id: "Qwen/Qwen3-VL-30B-A3B-Thinking-FP8"
+  example_model_input: # Loads sample request into Baseten playground
+    model: "Qwen/Qwen3-VL-30B-A3B-Thinking"
+    stream: false
+    max_tokens: 4096
+    messages:
+      - role: user
+        content:
+          - type: text
+            text: "What's in this image?"
+          - type: image_url
+            image_url:
+              url: "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true"
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: Qwen3-VL-30B-A3B-Thinking-FP8
+base_image:
+  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:2f7dbc9b42c51ba192e3dded515e4e07cdfdabea
+build_commands:
+  - pip install --pre --upgrade transformers
+  - pip uninstall -y vllm
+  - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git@d3d649efec8161b62e8db576f8d1d02a77d22897
+docker_server:
+  start_command: python3 -m vllm.entrypoints.openai.api_server --model Qwen/Qwen3-VL-30B-A3B-Thinking-FP8 --tool-call-parser hermes --reasoning-parser qwen3 --served-model-name Qwen/Qwen3-VL-30B-A3B-Thinking --enable-expert-parallel --enable-auto-tool-choice --tensor-parallel-size 2 --host 0.0.0.0 --port 8000
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/qwen/qwen-3-vl-32b/README.md b/llm/qwen/qwen-3-vl-32b/README.md
new file mode 100644
index 000000000..07125e980
--- /dev/null
+++ b/llm/qwen/qwen-3-vl-32b/README.md
@@ -0,0 +1,57 @@
+# Qwen 3 VL 32B
+
+Deploy [Qwen/Qwen3-VL-32B-Instruct-FP8](https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct-FP8) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-VL-32B-Instruct-FP8](https://huggingface.co/Qwen/Qwen3-VL-32B-Instruct-FP8) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:1 |
+| OpenAI compatible | Yes |
+| Python | py312 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-VL-32B-Instruct-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-VL-32B-Instruct-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `vllm/vllm-openai:v0.11.0`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **16**
+- Streaming: **enabled**
diff --git a/llm/qwen/qwen-3-vl-32b/config.yaml b/llm/qwen/qwen-3-vl-32b/config.yaml
new file mode 100644
index 000000000..2d0b6553c
--- /dev/null
+++ b/llm/qwen/qwen-3-vl-32b/config.yaml
@@ -0,0 +1,45 @@
+description: "Qwen3 VL 32B Instruct for vision-language tasks"
+base_image:
+  image: vllm/vllm-openai:v0.11.0
+model_metadata:
+  example_model_input: # Loads sample request into Baseten playground
+    model: ""
+    messages:
+      - role: user
+        content:
+          - type: image_url
+            image_url:
+              url: "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
+          - type: text
+            text: "Describe this image in detail."
+    stream: true
+  tags:
+    - openai-compatible
+model_name: Qwen 3 VL 32B
+requirements:
+  - transformers>=4.55.0
+  - accelerate==1.2.1
+  - timm==1.0.12
+  - einops==0.8.0
+  - open-clip-torch==2.29.0
+  - pillow==10.4.0
+python_version: py312
+model_cache:
+  - repo_id: Qwen/Qwen3-VL-32B-Instruct-FP8
+    revision: main
+    use_volume: true
+    volume_folder: "qwen-3-vl-32b"
+    ignore_patterns:
+      - "original/*"
+      - "*.pth"
+docker_server:
+  start_command: vllm serve Qwen/Qwen3-VL-32B-Instruct-FP8 --tensor-parallel-size 1 --served-model-name qwen-3-vl-32b --trust-remote-code --max-model-len 16384 --gpu-memory-utilization 0.9
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:1
+  use_gpu: true
+runtime:
+  predict_concurrency: 16
diff --git a/llm/qwen/qwen-coder-next/README.md b/llm/qwen/qwen-coder-next/README.md
new file mode 100644
index 000000000..10535d7ed
--- /dev/null
+++ b/llm/qwen/qwen-coder-next/README.md
@@ -0,0 +1,56 @@
+# Qwen3-Coder-Next
+
+Deploy [Qwen/Qwen3-Coder-Next-FP8](https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-Coder-Next-FP8](https://huggingface.co/Qwen/Qwen3-Coder-Next-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-Coder-Next-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-Coder-Next-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:nightly-dev-20260202-9227d4f7`
+- Predict concurrency: **64**
+- System packages: `tmux, htop, nload`
+- Streaming: **enabled**
diff --git a/llm/qwen/qwen-coder-next/config.yaml b/llm/qwen/qwen-coder-next/config.yaml
new file mode 100644
index 000000000..56482c0ab
--- /dev/null
+++ b/llm/qwen/qwen-coder-next/config.yaml
@@ -0,0 +1,41 @@
+description: "Qwen3 Coder Next for code generation"
+model_metadata:
+  repo_id: "Qwen/Qwen3-Coder-Next-FP8"
+  example_model_input:
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What is the meaning of life?"
+    stream: true
+    model: Qwen/Qwen3-Coder-Next
+    max_tokens: 32768
+    temperature: 0.7
+  tags:
+    - openai-compatible
+
+base_image:
+  image: lmsysorg/sglang:nightly-dev-20260202-9227d4f7
+
+
+build_commands:
+  - pip uninstall -y sglang
+  - git clone https://github.com/sgl-project/sglang.git && cd sglang && pip install --upgrade pip && pip install -e "python" && pip install nvidia-cudnn-cu12>=9.16.0.29
+
+docker_server:
+  start_command: sh -c 'python -m sglang.launch_server --model Qwen/Qwen3-Coder-Next-FP8 --port 30000 --tp-size 2 --tool-call-parser qwen3_coder'
+  readiness_endpoint: /health_generate
+  liveness_endpoint: /health_generate
+  predict_endpoint: /v1/chat/completions
+  server_port: 30000
+requirements: []
+system_packages:
+- tmux
+- htop
+- nload
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency : 64
+model_name: Qwen3-Coder-Next
diff --git a/llm/qwen/qwen-image/README.md b/llm/qwen/qwen-image/README.md
new file mode 100644
index 000000000..7ad9803ed
--- /dev/null
+++ b/llm/qwen/qwen-image/README.md
@@ -0,0 +1,37 @@
+# Qwen Image
+
+Deploy [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen-Image](https://huggingface.co/Qwen/Qwen-Image) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | H100 |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "A beautiful sunset over a mountain landscape with golden clouds, Ultra HD, 4K, cinematic composition",
+  "width": 1024,
+  "height": 1024,
+  "num_inference_steps": 50,
+  "true_cfg_scale": 4.0,
+  "seed": 42
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/qwen/qwen-image/assets/generated_image.jpg b/llm/qwen/qwen-image/assets/generated_image.jpg
similarity index 100%
rename from qwen/qwen-image/assets/generated_image.jpg
rename to llm/qwen/qwen-image/assets/generated_image.jpg
diff --git a/llm/qwen/qwen-image/config.yaml b/llm/qwen/qwen-image/config.yaml
new file mode 100644
index 000000000..514c3bb13
--- /dev/null
+++ b/llm/qwen/qwen-image/config.yaml
@@ -0,0 +1,38 @@
+description: "Qwen Image for vision-language tasks"
+external_package_dirs: []
+model_cache:
+  - repo_id: Qwen/Qwen-Image
+    use_volume: false
+    allow_patterns:
+      - "*.json"
+      - "*.safetensors"
+      - "*.bin"
+      - "*.txt"
+      - "*.md"
+model_metadata:
+  example_model_input: {
+    "prompt": "A beautiful sunset over a mountain landscape with golden clouds, Ultra HD, 4K, cinematic composition",
+    "width": 1024,
+    "height": 1024,
+    "num_inference_steps": 50,
+    "true_cfg_scale": 4.0,
+    "seed": 42
+  }
+model_name: Qwen Image
+python_version: py311
+requirements:
+  - git+https://github.com/huggingface/diffusers
+  - torch>=2.5.1
+  - transformers>=4.48.2
+  - accelerate==1.2.1
+  - safetensors==0.4.1
+  - pillow==10.4.0
+  - numpy==1.26.0
+resources:
+  accelerator: H100
+  use_gpu: true
+secrets: {}
+system_packages:
+  - ffmpeg
+  - libsm6
+  - libxext6
diff --git a/qwen/qwen-image/model/model.py b/llm/qwen/qwen-image/model/model.py
similarity index 100%
rename from qwen/qwen-image/model/model.py
rename to llm/qwen/qwen-image/model/model.py
diff --git a/llm/qwen/qwen-vl/README.md b/llm/qwen/qwen-vl/README.md
new file mode 100644
index 000000000..d178adc65
--- /dev/null
+++ b/llm/qwen/qwen-vl/README.md
@@ -0,0 +1,30 @@
+# Qwen VL
+
+Deploy [Qwen/Qwen-VL](https://huggingface.co/Qwen/Qwen-VL) for text generation using a Custom (Truss) engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen-VL](https://huggingface.co/Qwen/Qwen-VL) |
+| Task | Text generation |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"prompt": "What is machine learning?", "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/llm/qwen/qwen-vl/config.yaml b/llm/qwen/qwen-vl/config.yaml
new file mode 100644
index 000000000..622219b40
--- /dev/null
+++ b/llm/qwen/qwen-vl/config.yaml
@@ -0,0 +1,30 @@
+description: "Qwen VL for vision-language tasks"
+environment_variables: {}
+external_package_dirs: []
+model_cache:
+- allow_patterns:
+  - '*.json'
+  - '*.fp16.safetensors'
+  - '*.bin'
+  - '*.tiktoken'
+  - '*.py'
+  repo_id: Qwen/Qwen-VL
+  use_volume: false
+model_metadata:
+  example_model_input: {"image": "https://upload.wikimedia.org/wikipedia/commons/thumb/3/3a/Cat03.jpg/1200px-Cat03.jpg", "prompt": "Describe this image in detail"}
+model_name: Qwen VL
+python_version: py310
+requirements:
+- torch==2.0.1
+- accelerate==0.24.0
+- transformers==4.35.0
+- einops==0.7.0
+- torchvision==0.15.2
+- matplotlib==3.8.2
+- tiktoken==0.5.2
+- transformers_stream_generator==0.0.4
+resources:
+  accelerator: A10G
+  use_gpu: true
+secrets: {}
+system_packages: []
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/model/__init__.py b/llm/qwen/qwen-vl/model/__init__.py
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/large_context/model/__init__.py
rename to llm/qwen/qwen-vl/model/__init__.py
diff --git a/qwen/qwen-vl/model/model.py b/llm/qwen/qwen-vl/model/model.py
similarity index 100%
rename from qwen/qwen-vl/model/model.py
rename to llm/qwen/qwen-vl/model/model.py
diff --git a/llm/seed/seed-llm/README.md b/llm/seed/seed-llm/README.md
new file mode 100644
index 000000000..0d2cd8ffb
--- /dev/null
+++ b/llm/seed/seed-llm/README.md
@@ -0,0 +1,56 @@
+# Seed-OSS-36B-Instruct
+
+Deploy [ByteDance-Seed/Seed-OSS-36B-Instruct](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct) for text generation using a vLLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [ByteDance-Seed/Seed-OSS-36B-Instruct](https://huggingface.co/ByteDance-Seed/Seed-OSS-36B-Instruct) |
+| Task | Text generation |
+| Engine | vLLM |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="ByteDance-Seed/Seed-OSS-36B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "ByteDance-Seed/Seed-OSS-36B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:78336a0c3ee4eb9dba6e37959d926160e91623fd`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **128**
+- Streaming: **enabled**
diff --git a/llm/seed/seed-llm/config.yaml b/llm/seed/seed-llm/config.yaml
new file mode 100644
index 000000000..445c2dd58
--- /dev/null
+++ b/llm/seed/seed-llm/config.yaml
@@ -0,0 +1,38 @@
+description: "ByteDance-Seed/Seed-OSS-36B-Instruct for text generation"
+base_image:
+  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:78336a0c3ee4eb9dba6e37959d926160e91623fd
+build_commands:
+  - pip install --pre --upgrade transformers
+  - pip uninstall -y vllm
+  - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git@78336a0c3ee4eb9dba6e37959d926160e91623fd
+model_metadata:
+  repo_id: ByteDance-Seed/Seed-OSS-36B-Instruct
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write FizzBuzz in Python"
+    stream: true
+    model: "ByteDance-Seed/Seed-OSS-36B-Instruct"
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+docker_server:
+  start_command: python3 -m vllm.entrypoints.openai.api_server --model  ByteDance-Seed/Seed-OSS-36B-Instruct -O3 --tensor-parallel-size 2 --tool-call-parser seed_oss --served-model-name ByteDance-Seed/Seed-OSS-36B-Instruct --enable-auto-tool-choice --max-model-len 65536 --gpu-memory-utilization=0.95 --host 0.0.0.0 --port 8000
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency: 128
+model_cache:
+  - repo_id: ByteDance-Seed/Seed-OSS-36B-Instruct
+    revision: 497f1dca95ebdec98e41d517b9f060ee753c902f
+    use_volume: true
+    volume_folder: glm
+model_name: Seed-OSS-36B-Instruct
diff --git a/llm/z-ai/glm-4-5-air-fp8/README.md b/llm/z-ai/glm-4-5-air-fp8/README.md
new file mode 100644
index 000000000..30ed43ef8
--- /dev/null
+++ b/llm/z-ai/glm-4-5-air-fp8/README.md
@@ -0,0 +1,56 @@
+# GLM 4.5 Air FP8
+
+Deploy [zai-org/GLM-4.5-Air-FP8](https://huggingface.co/zai-org/GLM-4.5-Air-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [zai-org/GLM-4.5-Air-FP8](https://huggingface.co/zai-org/GLM-4.5-Air-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:4 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="zai-org/GLM-4.5-Air-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "zai-org/GLM-4.5-Air-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.9.post6-cu126`
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/z-ai/glm-4-5-air-fp8/config.yaml b/llm/z-ai/glm-4-5-air-fp8/config.yaml
new file mode 100644
index 000000000..d669acd48
--- /dev/null
+++ b/llm/z-ai/glm-4-5-air-fp8/config.yaml
@@ -0,0 +1,29 @@
+description: "zai-org/GLM-4.5-Air-FP8 for text generation"
+model_metadata:
+  repo_id: "zai-org/GLM-4.5-Air-FP8"
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write FizzBuzz in Python"
+    stream: true
+    model: "baseten-sglang"
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: GLM 4.5 Air FP8
+base_image:
+  image: lmsysorg/sglang:v0.4.9.post6-cu126
+docker_server:
+  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air-FP8 --tp-size 4   --tool-call-parser glm45    --reasoning-parser glm45    --speculative-algorithm EAGLE   --speculative-num-steps 3    --speculative-eagle-topk 1    --speculative-num-draft-tokens 4   --mem-fraction-static 0.7   --disable-shared-experts-fusion   --served-model-name glm-4.5-air-fp8   --host 0.0.0.0   --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:4
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/z-ai/glm-4-5-fp8/README.md b/llm/z-ai/glm-4-5-fp8/README.md
new file mode 100644
index 000000000..cfe5e3308
--- /dev/null
+++ b/llm/z-ai/glm-4-5-fp8/README.md
@@ -0,0 +1,56 @@
+# GLM 4.5 FP8
+
+Deploy [zai-org/GLM-4.5-FP8](https://huggingface.co/zai-org/GLM-4.5-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [zai-org/GLM-4.5-FP8](https://huggingface.co/zai-org/GLM-4.5-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:8 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="zai-org/GLM-4.5-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "zai-org/GLM-4.5-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.4.9.post6-cu126`
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/z-ai/glm-4-5-fp8/config.yaml b/llm/z-ai/glm-4-5-fp8/config.yaml
new file mode 100644
index 000000000..df10ead3c
--- /dev/null
+++ b/llm/z-ai/glm-4-5-fp8/config.yaml
@@ -0,0 +1,29 @@
+description: "zai-org/GLM-4.5-FP8 for text generation"
+model_metadata:
+  repo_id: "zai-org/GLM-4.5-FP8"
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write FizzBuzz in Python"
+    stream: true
+    model: "baseten-sglang"
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: GLM 4.5 FP8
+base_image:
+  image: lmsysorg/sglang:v0.4.9.post6-cu126
+docker_server:
+  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-FP8 --tp-size 4   --tool-call-parser glm45    --reasoning-parser glm45    --speculative-algorithm EAGLE   --speculative-num-steps 3    --speculative-eagle-topk 1    --speculative-num-draft-tokens 4   --mem-fraction-static 0.7   --disable-shared-experts-fusion   --served-model-name glm-4.5-fp8   --host 0.0.0.0   --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:8
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/z-ai/glm-4-5-v/README.md b/llm/z-ai/glm-4-5-v/README.md
new file mode 100644
index 000000000..5e66e3a3a
--- /dev/null
+++ b/llm/z-ai/glm-4-5-v/README.md
@@ -0,0 +1,54 @@
+# GLM-4.5V-FP8
+
+Deploy [zai-org/GLM-4.5V-FP8](https://huggingface.co/zai-org/GLM-4.5V-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [zai-org/GLM-4.5V-FP8](https://huggingface.co/zai-org/GLM-4.5V-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="zai-org/GLM-4.5V-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "zai-org/GLM-4.5V-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.5.4.post1-cu129-amd64`
+- Predict concurrency: **32**
diff --git a/llm/z-ai/glm-4-5-v/config.yaml b/llm/z-ai/glm-4-5-v/config.yaml
new file mode 100644
index 000000000..3a10ddd00
--- /dev/null
+++ b/llm/z-ai/glm-4-5-v/config.yaml
@@ -0,0 +1,30 @@
+description: "GLM-4.5V for vision-language tasks"
+model_metadata:
+  repo_id: "zai-org/GLM-4.5V-FP8"
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write FizzBuzz in Python"
+    stream: false
+    model: "zai-org/GLM-4.5V-FP8"
+    top_p: 0.95
+    extra_body: { "top_k": 40 }
+    max_tokens: 2048
+  tags:
+    - openai-compatible
+model_name: GLM-4.5V-FP8
+base_image:
+  image: lmsysorg/sglang:v0.5.4.post1-cu129-amd64
+docker_server:
+  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.5V-FP8 --tp-size 8 --tool-call-parser glm45 --reasoning-parser glm45 --mem-fraction-static 0.9 --served-model-name zai-org/GLM-4.5V --host 0.0.0.0 --port 8000 --enable-cache-report"
+  readiness_endpoint: /health_generate
+  liveness_endpoint: /health_generate
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:8
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/z-ai/glm-4-6-fp8/README.md b/llm/z-ai/glm-4-6-fp8/README.md
new file mode 100644
index 000000000..5ddcaa9b8
--- /dev/null
+++ b/llm/z-ai/glm-4-6-fp8/README.md
@@ -0,0 +1,56 @@
+# GLM-4.6-FP8
+
+Deploy [zai-org/GLM-4.6-FP8](https://huggingface.co/zai-org/GLM-4.6-FP8) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [zai-org/GLM-4.6-FP8](https://huggingface.co/zai-org/GLM-4.6-FP8) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:8 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="zai-org/GLM-4.6-FP8",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "zai-org/GLM-4.6-FP8", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:v0.5.3rc1-cu126`
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/z-ai/glm-4-6-fp8/config.yaml b/llm/z-ai/glm-4-6-fp8/config.yaml
new file mode 100644
index 000000000..89214233d
--- /dev/null
+++ b/llm/z-ai/glm-4-6-fp8/config.yaml
@@ -0,0 +1,33 @@
+description: "zai-org/GLM-4.6-FP8 for text generation"
+model_metadata:
+  repo_id: "zai-org/GLM-4.6-FP8"
+  example_model_input: # Loads sample request into Baseten playground
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "Write FizzBuzz in Python"
+    stream: true
+    model: "zai-org/GLM-4.6"
+    max_tokens: 4096
+    temperature: 0.6
+  tags:
+    - openai-compatible
+model_name: GLM-4.6-FP8
+base_image:
+  image: lmsysorg/sglang:v0.5.3rc1-cu126
+# build_commands:
+#   - pip install --upgrade pip
+#   - pip uninstall -y sglang
+#   - git clone https://github.com/sgl-project/sglang.git && cd sglang && git checkout 229d2b95f19573ece9c1c5d6b357df9874e04f59 && pip install -e "python[all]"
+docker_server:
+  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.6-FP8 --tp-size 8 --tool-call-parser glm45 --reasoning-parser glm45 --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --mem-fraction-static 0.9 --disable-shared-experts-fusion --served-model-name zai-org/GLM-4.6 --host 0.0.0.0 --port 8000"
+  readiness_endpoint: /health
+  liveness_endpoint: /health
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:8
+  use_gpu: true
+runtime:
+  predict_concurrency: 32
diff --git a/llm/z-ai/glm-4-7-flash/README.md b/llm/z-ai/glm-4-7-flash/README.md
new file mode 100644
index 000000000..88978ae85
--- /dev/null
+++ b/llm/z-ai/glm-4-7-flash/README.md
@@ -0,0 +1,55 @@
+# GLM 4.7 Flash
+
+Deploy [zai-org/GLM-4.7-Flash](https://huggingface.co/zai-org/GLM-4.7-Flash) for text generation using a SGLang engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [zai-org/GLM-4.7-Flash](https://huggingface.co/zai-org/GLM-4.7-Flash) |
+| Task | Text generation |
+| Engine | SGLang |
+| GPU | H100:2 |
+| OpenAI compatible | Yes |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="zai-org/GLM-4.7-Flash",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "zai-org/GLM-4.7-Flash", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Base image: `lmsysorg/sglang:nightly-dev-20260122-e6ccb294`
+- Predict concurrency: **32**
+- Streaming: **enabled**
diff --git a/llm/z-ai/glm-4-7-flash/config.yaml b/llm/z-ai/glm-4-7-flash/config.yaml
new file mode 100644
index 000000000..e58b9f2ca
--- /dev/null
+++ b/llm/z-ai/glm-4-7-flash/config.yaml
@@ -0,0 +1,35 @@
+description: "zai-org/GLM-4.7-Flash for text generation"
+model_metadata:
+  repo_id: "zai-org/GLM-4.7-Flash"
+  example_model_input:
+    messages:
+      - role: system
+        content: "You are a helpful assistant."
+      - role: user
+        content: "What is the meaning of life?"
+    stream: true
+    model: zai-org/GLM-4.7-Flash
+    max_tokens: 32768
+    temperature: 0.7
+  tags:
+    - openai-compatible
+base_image:
+  image: lmsysorg/sglang:nightly-dev-20260122-e6ccb294
+
+build_commands:
+  - pip uninstall -y transformers
+  - pip install git+https://github.com/huggingface/transformers.git@76732b4e7120808ff989edbd16401f61fa6a0afa
+
+docker_server:
+  start_command: python3 -m sglang.launch_server --model-path zai-org/GLM-4.7-Flash --tp-size 2 --tool-call-parser glm47 --reasoning-parser glm45 --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --mem-fraction-static 0.8 --served-model-name zai-org/GLM-4.7-Flash --host 0.0.0.0 --port 8000
+  readiness_endpoint: /health_generate
+  liveness_endpoint: /health_generate
+  predict_endpoint: /v1/chat/completions
+  server_port: 8000
+resources:
+  accelerator: H100:2
+  use_gpu: true
+runtime:
+  predict_concurrency : 32
+
+model_name: GLM 4.7 Flash
diff --git a/lora/engine-lora/README.md b/lora/engine-lora/README.md
deleted file mode 100644
index 9f22e950b..000000000
--- a/lora/engine-lora/README.md
+++ /dev/null
@@ -1,63 +0,0 @@
-# IMPORTANT NOTICE
-
-Currently we're seeing an issue where using LoRAs with TensorRT-LLM shows a significant drop in performance. It is strongly recommended to go with the [vLLM example](../vllm-lora)
-
-# Mistral 7B Instruct LoRA
-
-This is an example of a truss that supports **dynamic swapping of LoRA adapters**—allowing you to serve multiple fine-tuned variants efficiently from a single GPU. In this example, we deploy a **Mistral 7B Instruct** on [TensorRT-LLM Engine Builder](https://docs.baseten.co/performance/examples/mistral-trt). This model will be an expert in finance, medicine and law.
-
-- 📄 **TensorRT-LLM Details:** [Performance Example (Baseten Docs)](https://docs.baseten.co/performance/examples/mistral-trt)
-- 💡 **LoRA Swapping Overview:** [Baseten Blog: Serving 10,000 Fine-Tuned LLMs from One GPU](https://www.baseten.co/blog/how-to-serve-10-000-fine-tuned-llms-from-a-single-gpu/)
-
----
-
-## 🛠️ Implementing LoRA Swapping
-
-Extending a base TensorRT-LLM deployment to support LoRA swapping requires three config changes:
-
-### 1. Configure `lora_adapters`
-
-List each LoRA adapter’s name and its download source. Supported sources are:
-- `HF` for HuggingFace
-- `GCS` for Google Cloud Storage
-- `REMOTE_URL` for any direct download
-
-**Example (`config.yaml`):**
-```yaml
-lora_adapters:
-  legal:
-    source: HF
-    repo: Aretoss/Lexgen
-  finance:
-    source: HF
-    repo: vaibhav1/lora-mistral-finance
-  medical:
-    source: HF
-    repo: Imsachinsingh00/Fine_tuned_LoRA_Mistral_MTSDialog_Summarization
-```
-
-### 2. Set `served_model_name` (Optional)
-
-Set this parameter if you wish to allow requests to the base model (without any LoRA applied).
-
----
-
-### 3. Select Adapter or Base Model at Request Time
-
-Specify the desired adapter or base model using the `model` field in your request payload.
-
-**Example request body**
-
-```json
-{
-  "model": "finance", // # Or legal, medical, or mistral
-  "stream": true,
-  "messages": [
-    {"role": "user", "content": "What would you choose in 2008?"}
-  ],
-  "max_tokens": 1024,
-  "temperature": 0.9
-}
-```
-
-### For full details, see [documentation](https://docs.baseten.co/development/model/performance/engine-builder-config)
diff --git a/lora/engine-lora/config.yaml b/lora/engine-lora/config.yaml
deleted file mode 100644
index 55dd7908d..000000000
--- a/lora/engine-lora/config.yaml
+++ /dev/null
@@ -1,56 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      model: "finance",
-      messages:
-        [{ role: "user", content: "How would you choose back in 2008?" }],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: mistralai/Mistral-7B-Instruct-v0.3
-model_name: Mistral 7B Instruct Engine Lora
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mistral-7B-Instruct-v0.3
-      source: HF
-    lora_adapters:
-      legal:
-        source: HF
-        repo: Aretoss/Lexgen
-      finance:
-        source: HF
-        repo: vaibhav1/lora-mistral-finance
-      medical:
-        source: HF
-        repo: Imsachinsingh00/Fine_tuned_LoRA_Mistral_MTSDialog_Summarization
-    max_seq_len: 32768
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
-    served_model_name: mistral
diff --git a/lora/sglang-lora/README.md b/lora/sglang-lora/README.md
deleted file mode 100644
index 58a383f40..000000000
--- a/lora/sglang-lora/README.md
+++ /dev/null
@@ -1,47 +0,0 @@
-# IMPORTANT NOTICE
-
-Currently, [support for openai compatible server with lora is not yet implemented in SGLang](https://github.com/sgl-project/sglang/issues/2929). It is strongly recommended to go with the [vLLM example](../vllm-lora).
-
-# Mistral 7B Instruct LoRA
-
-This is an example of a truss that supports **dynamic swapping of LoRA adapters**—allowing you to serve multiple fine-tuned variants efficiently from a single GPU. In this example, we deploy a **Mistral 7B Instruct** with SGLang. This model will be an expert in finance, medicine and law.
-
-- 💡 **LoRA Swapping Overview:** [Baseten Blog: Serving 10,000 Fine-Tuned LLMs from One GPU](https://www.baseten.co/blog/how-to-serve-10-000-fine-tuned-llms-from-a-single-gpu/)
-
----
-
-## 🛠️ Implementing LoRA Swapping
-
-Extending a base SGLang custom server deployment to support LoRA swapping requires two config changes:
-
-### 1. Configure `lora_paths`
-
-List each LoRA adapter’s name and its huggingface repo.
-
-**Example (`start_command` in `config.yaml`):**
-```
---enable-lora --lora-paths legal=Aretoss/Lexgen finance=vaibhav1/lora-mistral-finance medical=Imsachinsingh00/Fine_tuned_LoRA_Mistral_MTSDialog_Summarization --disable-radix-cache
-```
-
-Note that the `--disable-radix-cache` flag is necessary because lora with radix attention is not yet implemented in SGLang.
-
----
-
-### 3. Select Adapter or Base Model at Request Time
-
-This API is not openai compatible. Follow the example model input in the `README.md`.
-
-**Example request body**
-
-```json
-{
-"text": [
-    "What would you choose in 2008?",
-    "What would you choose in 2008?",
-],
-"sampling_params": {"max_new_tokens": 1000, "temperature": 1.0},
-"lora_path": ["legal", "finance"],
-}
-```
-
-### For full details, see SGLang's [documentation](https://docs.sglang.ai/backend/lora.html)
diff --git a/lora/sglang-lora/config.yaml b/lora/sglang-lora/config.yaml
deleted file mode 100644
index 9d64c5fd9..000000000
--- a/lora/sglang-lora/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-base_image:
-  image: lmsysorg/sglang:v0.4.9.post6-cu126
-model_metadata:
-  example_model_input: {
-    "text": [
-        "What would you choose in 2008?",
-        "What would you choose in 2008?",
-    ],
-    "sampling_params": {"max_new_tokens": 1000, "temperature": 1.0},
-    "lora_path": ["legal", "finance"],
-  }
-  repo_id: mistralai/Mistral-7B-Instruct-v0.3
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m sglang.launch_server --model-path mistralai/Mistral-7B-Instruct-v0.3 --port 8000 --trust-remote-code --enable-lora --lora-paths legal=Aretoss/Lexgen finance=vaibhav1/lora-mistral-finance medical=Imsachinsingh00/Fine_tuned_LoRA_Mistral_MTSDialog_Summarization --disable-radix-cache"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /generate
-  server_port: 8000
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 24Gi
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: Mistral-7B-Instruct SGLang Lora
-environment_variables:
-  hf_access_token: null
-requirements:
-  - protobuf
diff --git a/lora/vllm-lora/README.md b/lora/vllm-lora/README.md
deleted file mode 100644
index 15b75b7f0..000000000
--- a/lora/vllm-lora/README.md
+++ /dev/null
@@ -1,46 +0,0 @@
-# Mistral 7B Instruct LoRA
-
-This is an example of a truss that supports **dynamic swapping of LoRA adapters**—allowing you to serve multiple fine-tuned variants efficiently from a single GPU. In this example, we deploy a **Mistral 7B Instruct** on vLLM's server. This model will be an expert in finance and law.
-
-- 💡 **LoRA Swapping Overview:** [Baseten Blog: Serving 10,000 Fine-Tuned LLMs from One GPU](https://www.baseten.co/blog/how-to-serve-10-000-fine-tuned-llms-from-a-single-gpu/)
-
----
-
-## 🛠️ Implementing LoRA Swapping
-
-Extending a base vLLM deployment to support LoRA swapping requires three config changes:
-
-### 1. Configure `lora_modules`
-
-List each LoRA adapter’s name and its huggingface repo.
-
-**Example (`config.yaml`):**
-```
---enable-lora --lora-modules finance=vaibhav1/lora-mistral-finance legal=Aretoss/Lexgen
-```
-
-### 2. Set `served_model_name` (Optional)
-
-Set this parameter if you wish to allow requests to the base model (without any LoRA applied).
-
----
-
-### 3. Select Adapter or Base Model at Request Time
-
-Specify the desired adapter or base model using the `model` field in your request payload.
-
-**Example request body**
-
-```json
-{
-  "model": "finance", // Or "legal" or "mistral"
-  "stream": true,
-  "messages": [
-    {"role": "user", "content": "What would you choose in 2008?"}
-  ],
-  "max_tokens": 1024,
-  "temperature": 0.9
-}
-```
-
-### For full details, see [documentation](https://docs.vllm.ai/en/v0.9.1/features/lora.html#lora-model-lineage-in-model-card)
diff --git a/lora/vllm-lora/config.yaml b/lora/vllm-lora/config.yaml
deleted file mode 100644
index a6e9f84d6..000000000
--- a/lora/vllm-lora/config.yaml
+++ /dev/null
@@ -1,32 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.9.2
-model_metadata:
-  example_model_input: {
-    model: "finance",
-    messages: [
-      {
-        role: "user",
-        content: "How would you choose back in 2008?"
-      }
-    ],
-    stream: true,
-    max_tokens: 512,
-    temperature: 0.9
-  }
-  repo_id: mistralai/Mistral-7B-Instruct-v0.3
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Mistral-7B-Instruct-v0.3 --tokenizer_mode mistral --config_format mistral --load_format mistral --served-model-name mistral --max-model-len 16384 --port 8000 --gpu-memory-utilization 0.90 --disable-custom-all-reduce --trust-remote-code --enable-lora --lora-modules finance=vaibhav1/lora-mistral-finance legal=Aretoss/Lexgen"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 24Gi
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-model_name: Mistral-7B-Instruct VLLM Lora
-environment_variables:
-  hf_access_token: null
diff --git a/magic-animate/README.md b/magic-animate/README.md
deleted file mode 100644
index a07290037..000000000
--- a/magic-animate/README.md
+++ /dev/null
@@ -1,89 +0,0 @@
-# Magic Animate Truss
-
-This repository packages [Magic Animate](https://github.com/magic-research/magic-animate) as a [Truss](https://truss.baseten.co/).
-
-Magic Animate allows you to create human image animation using a diffusion model. The model combines two inputs, a picture of a person and a densepose motion sequence, so that the human in the picture gets animated based on the provided motion sequence. Here is an example:
-
-
-https://github.com/htrivedi99/truss-examples/assets/15642666/914d7b50-c0e3-40fc-a146-c53743c82cbd
-
-
-## Deploying Magic Animate
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd magic-animate
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `magic-animate` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking the model
-
-Here are the following inputs for the model:
-1. `reference_image` (required): The image of the person you'd like to animate as a base64 string. Square images work better since each image gets resized to 512 x 512.
-2. `motion_sequence` (required): A densepose motion sequence as a base64 string. You can use something like [detectron2](https://github.com/facebookresearch/detectron2/tree/main/projects/DensePose) to create custom densepose sequences.
-3. `seed` (optional): A random seed for the model.
-4. `steps` (optional): The number of iterations the model runs through.
-5. `guidance_scale` (optional): Used to determine how closely the image generation follows the prompt.
-6. `grid` (optional): A boolean value which controls if the output comes back with all the clips or just the individual output.
-
-If `grid` is passed in as `True` the model will respond with a JSON object with two keys: `output` and `grid_clip`. If `grid` is not passed in or is set to `False` the JSON object returned only has one key `output`. Both `output` and `grid_clip` are MP4 files returned as a base64 string.
-
-Here is an example of how to invoke this model:
-
-```python
-from PIL import Image
-import base64
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-def mp4_to_base64(file_path: str):
-    with open(file_path, "rb") as mp4_file:
-        binary_data = mp4_file.read()
-        base64_data = base64.b64encode(binary_data)
-        base64_string = base64_data.decode("utf-8")
-
-    return base64_string
-
-def base64_to_mp4(base64_string, output_file_path):
-    binary_data = base64.b64decode(base64_string)
-    with open(output_file_path, "wb") as output_file:
-        output_file.write(binary_data)
-
-img = Image.open("/path/to/image/monalisa.png")
-input_img = pil_to_b64(img)
-motion_sequence = mp4_to_base64("/path/to/densepose/sequence/demo4.mp4")
-data = {"reference_image": input_img, "motion_sequence": motion_sequence, "steps": 10, "grid": True}
-res = requests.post("https://model-<model-id>.api.baseten.co/development/predict", headers=headers, json=data)
-res = res.json()
-base64_to_mp4(res.get("output"), "magic-animate.mp4")
-base64_to_mp4(res.get("grid_clip"), "grid.mp4")
-```
-
-Here is the example `monalisa.png` image:
-
-![monalisa](https://github.com/htrivedi99/truss-examples/assets/15642666/9e9f4e40-6c55-415b-b37c-3271572ffb77)
-
-
-Here is the example densepose sequence `demo4.mp4`:
-
-https://github.com/htrivedi99/truss-examples/assets/15642666/c20a9761-1279-4c0b-9de1-7fd73cb43fd7
diff --git a/magic-animate/config.yaml b/magic-animate/config.yaml
deleted file mode 100644
index f2c414ef1..000000000
--- a/magic-animate/config.yaml
+++ /dev/null
@@ -1,36 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    guidance_scale: 7.5
-    motion_sequence: <BASE64 MP4 FILE>
-    reference_image: <BASE64 IMAGE>
-    seed: 1
-    steps: 10
-model_name: Magic Animate
-python_version: py310
-requirements:
-- torch==2.0.1
-- torchvision==0.15.2
-- xformers==0.0.22
-- diffusers==0.21.4
-- pillow==9.5.0
-- numpy==1.24.4
-- omegaconf==2.3.0
-- transformers==4.32.0
-- einops==0.6.1
-- imageio==2.9.0
-- imageio-ffmpeg==0.4.3
-- tqdm==4.66.1
-- websockets==11.0.3
-- accelerate==0.22.0
-- huggingface-hub==0.16.4
-- av==11.0.0
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 15Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/metavoice-1b/README.md b/metavoice-1b/README.md
deleted file mode 100644
index b58075bbe..000000000
--- a/metavoice-1b/README.md
+++ /dev/null
@@ -1,59 +0,0 @@
-# MetaVoice 1B Truss
-[MetaVoice-1B](https://github.com/metavoiceio/metavoice-src) is an Apache-licensed, 1.2B parameter base model trained on 100K hours of speech for TTS (text-to-speech).
-
-This model is packaged using [Truss](https://trussml.com), the simplest way to serve AI/ML models in production.
-
-## Deploy MetaVoice 1B
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd metavoice-1b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `metavoice-1b` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using MetaVoice through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-
-## Invoking MetaVoice
-
-To use MetaVoice 1B, follow this command pattern, keeping in mind the 220-character limit for input text:
-
-```sh
-truss predict -d '{"text": "Your input text here"}' | python process.py
-```
-
-### Understanding process.py
-The process.py script is essential for handling the output from MetaVoice 1B. It reads the base64 encoded audio from the standard input, decodes it, and saves it as a WAV file. Here's a brief overview of how it works:
-
-```python
-import base64
-import sys
-
-b64_audio = sys.stdin.read()
-b64_audio = b64_audio.split('"')[1]  # Extracting the base64 string
-
-wav_file = open("output.wav", "wb")
-decode_string = base64.b64decode(b64_audio)
-wav_file.write(decode_string)
-```
-
-## Notes
-- `flash_attn` requires installation with `--no-build-isolation`. As this isn't supported, installing the wheel directly seems to work (see `config.yaml`).
-
-## To Dos
-- Add support for passing additional reference voices to use
diff --git a/metavoice-1b/config.yaml b/metavoice-1b/config.yaml
deleted file mode 100644
index e37e4c6ec..000000000
--- a/metavoice-1b/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-model_name: MetaVoice 1B
-description: MetaVoice is a transformer-based model for TTS
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input: '"text to speech models are cool"'
-python_version: py311
-data_dir: data
-model_cache:
-  - repo_id: metavoiceio/metavoice-1B-v0.1
-    use_volume: false
-    allow_patterns:
-      - "*.pt"
-  - repo_id: facebook/multiband-diffusion
-    allow_patterns:
-      - mbd_comp_8.pt
-  - repo_id: facebook/encodec_24khz
-    allow_patterns:
-      - "*.safetensors"
-
-requirements_file: ./requirements.txt
-resources:
-  accelerator: "A10G"
-  use_gpu: true
-secrets:
-  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
-system_packages:
-- ffmpeg
diff --git a/metrics/datadog/config.yaml b/metrics/datadog/config.yaml
deleted file mode 100644
index f74727ab9..000000000
--- a/metrics/datadog/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-base_image:
-  image: chriswirick/truss_fastapi_datadog_vllm:v0.11.0h
-docker_server:
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  readiness_endpoint: /health
-  server_port: 8000
-  start_command: sh -c "export DD_API_KEY=$(cat /secrets/dd_api_key | tr -d '\n\r' | xargs) && mkdir -p /tmp/datadog-agent /var/log/datadog && /opt/datadog-agent/bin/agent/agent run 2>&1 & sleep 3 && HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve Qwen/Qwen3-30B-A3B --reasoning-parser deepseek_r1 --served-model-name qwen30b --port 8000"
-environment_variables:
-  DD_SITE: "us5.datadoghq.com"
-  DD_HOSTNAME: "truss-vllm-server"
-  DD_SERVICE: "truss-vllm"
-  DD_ENV: "production"
-  DD_RUN_PATH: "/tmp/datadog-agent"
-  DD_AUTH_TOKEN_FILE_PATH: "/tmp/datadog-agent/auth_token"
-  DD_INVENTORIES_CHECKS_ENABLED: "false"
-  DD_OTLP_CONFIG_RECEIVER_PROTOCOLS_GRPC_ENDPOINT: ""
-  DD_CLOUD_PROVIDER_METADATA: "[]"
-  VLLM_LOGGING_LEVEL: WARNING
-model_metadata:
-  repo_id: Qwen/Qwen3-30B-A3B
-  example_model_input:
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What does Tongyi Qianwen mean?"
-    stream: false
-    model: "qwen30b"
-    max_tokens: 512
-    temperature: 0.7
-  tags:
-    - openai-compatible
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
-model_name: truss_fastapi_datadog
-secrets:
-  dd_api_key: null
-  hf_access_token: null
diff --git a/midnight/README.md b/midnight/README.md
deleted file mode 100644
index d25492e59..000000000
--- a/midnight/README.md
+++ /dev/null
@@ -1,254 +0,0 @@
-# Kaiko Midnight - Pathology Foundation Model
-
-A Baseten deployment of the [Kaiko Midnight](https://huggingface.co/kaiko-ai/midnight) pathology foundation model for medical image analysis and classification.
-
-## Overview
-
-Kaiko Midnight is a 1.14B parameter pathology foundation model based on DINOv2, optimized for medical image analysis. It provides both classification and segmentation embeddings for pathology images.
-
-## Deploy Kiako Midnight
-First, clone this repository:
-
-```bash
-git clone https://github.com/basetenlabs/truss-examples/
-cd midnight
-```
-
-Before deployment:
-
-Make sure you have a Baseten account and API key.
-Install the latest version of Truss: `pip install --upgrade truss`
-With midnight as your working directory, you can deploy the model with:
-
-```bash
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see Truss documentation.
-
-Once your Truss is deployed, you can start using Midnight through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-Note: If you run into the following error during the build phase, downgrade truss with: `pip install truss==0.9.111`
-
-```
-error: failed to solve: process "/bin/sh -c uv pip install --python $(which python3) -r base_server_requirements.txt --no-cache-dir" did not complete successfully: exit code: 2
-```
-
-## GPU Requirements
-
-- **Minimum**: T4 (16GB VRAM)
-
-## API Usage
-
-#### Single Image Processing
-```json
-{
-  "image_url": "https://upload.wikimedia.org/wikipedia/commons/8/80/Breast_DCIS_histopathology_%281%29.jpg",
-  "task": "classification",
-  "batch_size": 1
-}
-```
-
-#### Single Image with Base64
-```json
-{
-  "image_base64": "iVBORw0KGgoAAAANSUhEUgAA...",
-  "task": "classification",
-  "batch_size": 1
-}
-```
-
-#### Batch Processing (Multiple Images)
-```json
-{
-  "image_urls": [
-    "https://example.com/image1.jpg",
-    "https://example.com/image2.jpg",
-    "https://example.com/image3.jpg",
-    "https://example.com/image4.jpg"
-  ],
-  "task": "classification",
-  "batch_size": 4
-}
-```
-
-#### True Batch Processing with Base64
-```json
-{
-  "image_base64_list": [
-    "iVBORw0KGgoAAAANSUhEUgAA...",
-    "iVBORw0KGgoAAAANSUhEUgAA...",
-    "iVBORw0KGgoAAAANSUhEUgAA...",
-    "iVBORw0KGgoAAAANSUhEUgAA..."
-  ],
-  "task": "classification",
-  "batch_size": 4
-}
-```
-
-### Response Format
-```json
-{
-  "embeddings": [[0.123, 0.456, ...], [0.789, 0.012, ...]],
-  "embedding_shape": [2, 3072],
-  "task": "classification",
-  "model_id": "kaiko-ai/midnight",
-  "input_size": 224,
-  "actual_batch_size": 2,
-  "requested_batch_size": 4,
-  "optimal_batch_size": 8,
-  "gpu_memory_gb": 16.0
-}
-```
-
-## Example Usage
-
-### Python Client - Single Image
-```python
-import requests
-import base64
-
-# Baseten endpoint URL
-endpoint_url = "https://your-baseten-endpoint.baseten.co"
-
-# Example request for classification
-request_data = {
-    "image_url": "https://upload.wikimedia.org/wikipedia/commons/8/80/Breast_DCIS_histopathology_%281%29.jpg",
-    "task": "classification",
-    "batch_size": 1
-}
-
-# Make prediction
-response = requests.post(
-    endpoint_url,
-    headers={"Authorization": "Api-Key YOUR_API_KEY"},
-    json=request_data
-)
-result = response.json()
-
-print(f"Task: {result['task']}")
-print(f"Embedding shape: {result['embedding_shape']}")
-print(f"Embedding (first 10 values): {result['embeddings'][0][:10]}")
-```
-
-### Python Client - True Batch Processing
-```python
-import requests
-import base64
-
-def encode_image_to_base64(image_path):
-    """Convert local image to base64 string"""
-    with open(image_path, "rb") as image_file:
-        return base64.b64encode(image_file.read()).decode('utf-8')
-
-# Baseten endpoint URL
-endpoint_url = "https://your-baseten-endpoint.baseten.co"
-
-# Encode multiple local images
-image_paths = ["path/to/image1.jpg", "path/to/image2.jpg", "path/to/image3.jpg", "path/to/image4.jpg"]
-image_base64_list = [encode_image_to_base64(path) for path in image_paths]
-
-# Example request for batch classification
-request_data = {
-    "image_base64_list": image_base64_list,
-    "task": "classification",
-    "batch_size": 4
-}
-
-# Make prediction
-response = requests.post(
-    endpoint_url,
-    headers={"Authorization": "Api-Key YOUR_API_KEY"},
-    json=request_data
-)
-result = response.json()
-
-print(f"Task: {result['task']}")
-print(f"Embedding shape: {result['embedding_shape']}")
-print(f"Actual batch size: {result['actual_batch_size']}")
-print(f"Number of embeddings: {len(result['embeddings'])}")
-```
-
-### cURL Example - Single Image
-```bash
-curl -X POST "https://your-baseten-endpoint.baseten.co" \
-  -H "Authorization: Api-Key YOUR_API_KEY" \
-  -H "Content-Type: application/json" \
-  -d '{
-    "image_url": "https://upload.wikimedia.org/wikipedia/commons/8/80/Breast_DCIS_histopathology_%281%29.jpg",
-    "task": "classification",
-    "batch_size": 1
-  }'
-```
-
-### cURL Example - True Batch Processing
-```bash
-# First, encode your images to base64
-IMAGE1_BASE64=$(base64 -i path/to/image1.jpg)
-IMAGE2_BASE64=$(base64 -i path/to/image2.jpg)
-IMAGE3_BASE64=$(base64 -i path/to/image3.jpg)
-IMAGE4_BASE64=$(base64 -i path/to/image4.jpg)
-
-# Then make the batch request
-curl -X POST "https://your-baseten-endpoint.baseten.co" \
-  -H "Authorization: Api-Key YOUR_API_KEY" \
-  -H "Content-Type: application/json" \
-  -d "{
-    \"image_base64_list\": [\"$IMAGE1_BASE64\", \"$IMAGE2_BASE64\", \"$IMAGE3_BASE64\", \"$IMAGE4_BASE64\"],
-    \"task\": \"classification\",
-    \"batch_size\": 4
-  }"
-```
-
-## Input Requirements
-
-### Image Requirements
-- **Format**: JPEG, PNG, TIFF, or any format supported by PIL
-- **Size**: Any size (automatically resized to 224x224)
-- **Channels**: RGB (automatically converted if needed)
-- **Max Size**: Subject to Baseten's request size limits (typically 10MB)
-
-### Base64 Requirements
-- **Encoding**: Standard base64 encoding
-- **Data URL support**: Both raw base64 and data URLs (data:image/...) are supported
-- **Size limit**: Subject to Baseten's request size limits (typically 10MB)
-- **Format**: Any image format supported by PIL
-
-### Batch Processing Requirements
-- **True Batch**: Process multiple different images in a single request
-- **Batch Size**: Automatically optimized based on GPU memory
-- **Image Count**: Should match or exceed requested batch_size for optimal performance
-- **Memory Efficiency**: Better GPU utilization than single image processing
-
-## Task Types
-
-### Classification
-- **Purpose**: Global image embeddings for classification tasks
-- **Output**: Concatenated CLS token + mean patch embeddings
-- **Shape**: `(batch_size, 3072)` - 1536 + 1536 dimensions
-- **Use Cases**: Image classification, similarity search, feature extraction
-
-### Segmentation
-- **Purpose**: Spatial embeddings for segmentation tasks
-- **Output**: Patch embeddings reshaped to spatial dimensions
-- **Shape**: `(batch_size, 1536, 16, 16)` - 16x16 spatial grid
-- **Use Cases**: Semantic segmentation, object detection, spatial analysis
-
-
-
-## Model Architecture
-
-### Base Model
-- **Architecture**: DINOv2 Vision Transformer
-- **Parameters**: 1.14B parameters
-- **Precision**: F32 (float32)
-- **Input Size**: 224x224 pixels
-- **Patch Size**: 14x14 pixels
-- **Hidden Size**: 1536 dimensions
-
-### Embedding Extraction
-- **Classification**: CLS token + mean patch embeddings (3072 dimensions)
-- **Segmentation**: Patch embeddings reshaped to spatial grid (1536x16x16)
-- **Normalization**: Mean=(0.5,0.5,0.5), Std=(0.5,0.5,0.5)
diff --git a/midnight/config.yaml b/midnight/config.yaml
deleted file mode 100644
index fdfbdaf0b..000000000
--- a/midnight/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-model_name: Kaiko Midnight
-description: Pathology foundation model for medical image analysis and classification
-python_version: py39
-base_image:
-  image: nvcr.io/nvidia/pytorch:25.06-py3
-requirements_file: ./requirements.txt
-resources:
-  accelerator: T4
-  use_gpu: true
-  memory: 1Gi
-  cpu: '1'
-runtime:
-  predict_concurrency: 32
-model_metadata:
-  example_model_input:
-    image_url: "https://upload.wikimedia.org/wikipedia/commons/8/80/Breast_DCIS_histopathology_%281%29.jpg"
-    task: "classification"  # or "segmentation"
-    batch_size: 1
-  tags:
-    - medical-imaging
-    - pathology
-    - computer-vision
-    - embeddings
diff --git a/minimax/minimax_m2_1/config.yaml b/minimax/minimax_m2_1/config.yaml
deleted file mode 100644
index b241509ba..000000000
--- a/minimax/minimax_m2_1/config.yaml
+++ /dev/null
@@ -1,51 +0,0 @@
-base_image:
-  image: lmsysorg/sglang:nightly-dev-20260126-48f4340b
-docker_server:
-  liveness_endpoint: /health_generate
-  predict_endpoint: /v1/chat/completions
-  readiness_endpoint: /health_generate
-  server_port: 8000
-  start_command: sh -c "truss-transfer-cli && find /app/model_cache/checkpoint -type f -print0 | xargs -0 -P 0 -I {} dd if={} of=/dev/null bs=4M && python3 -m sglang.launch_server --model-path /app/model_cache/checkpoint --tp-size 8 --ep-size 8 --tool-call-parser minimax-m2 --trust-remote-code --host 0.0.0.0 --reasoning-parser minimax --port 8000 --mem-fraction-static 0.85"
-#environment_variables:
-#  SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN: 1
-model_cache:
-- allow_patterns:
-  - '*.json'
-  - '*.safetensors'
-  - '*.txt'
-  - '*.model'
-  - '*.py'
-  - '*.jinja'
-  repo_id: MiniMaxAI/MiniMax-M2.1
-  revision: 927ea2b64008fe4a1e31a4e107a6b75916b3b44a
-  use_volume: true
-  volume_folder: checkpoint
-model_metadata:
-  example_model_input:
-    max_tokens: 4096
-    messages:
-    - content: You are a helpful assistant.
-      role: system
-    - content: Who won the world series in 2020?
-      role: user
-    model: MiniMaxAI/MiniMax-M2
-    stream: true
-    temperature: 0.6
-  model_name: MiniMax-M2
-  tags:
-  - openai-compatible
-model_name: minimax
-resources:
-  accelerator: H100:8
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-runtime:
-  health_checks:
-    restart_check_delay_seconds: 1200
-    restart_threshold_seconds: 600
-    stop_traffic_threshold_seconds: 1800
-  is_websocket_endpoint: false
-  predict_concurrency: 32
-  transport:
-    kind: http
diff --git a/mistral/engine-devstral/config.yaml b/mistral/engine-devstral/config.yaml
deleted file mode 100644
index 76a59c3d4..000000000
--- a/mistral/engine-devstral/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-model_metadata:
-  example_model_input: {
-    messages: [
-      {
-        role: "system",
-        content: "<Load system prompt from Hugging Face>"
-      },
-      {
-        role: "user",
-        content: "<your-command>"
-      }
-    ],
-    stream: true,
-    max_tokens: 512,
-    temperature: 0.15,
-    top_p: 1.0,
-    top_k: 40,
-    frequency_penalty: 1
-  }
-  tags:
-    - openai-compatible
-model_name: Devstral Small 2505
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: mistralai/Devstral-Small-2505
-      revision: "refs/pr/8"
-      source: HF
-    num_builder_gpus: 2
-    max_batch_size: 64
-    max_seq_len: 131072
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-    speculator: # optional: use speculative decoding
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 8
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-  runtime:
-    enable_chunked_context: true
diff --git a/mistral/engine-mistral-small-3/README.md b/mistral/engine-mistral-small-3/README.md
deleted file mode 100644
index 4cbb6ba55..000000000
--- a/mistral/engine-mistral-small-3/README.md
+++ /dev/null
@@ -1 +0,0 @@
-# Mistral Small 3 (2501)
diff --git a/mistral/engine-mistral-small-3/config.yaml b/mistral/engine-mistral-small-3/config.yaml
deleted file mode 100644
index 0032b2815..000000000
--- a/mistral/engine-mistral-small-3/config.yaml
+++ /dev/null
@@ -1,53 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are a knowledgable, engaging, meteorology teacher.",
-          },
-          {
-            role: "user",
-            content: "What is the impact of the Mistral wind on the French climate?",
-          },
-        ],
-      stream: true,
-      max_tokens: 1024,
-      temperature: 0.15,
-    }
-  repo_id: mistralai/Mistral-Small-24B-Instruct-2501
-model_name: Mistral Small 3 Instruct FP8
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mistral-Small-24B-Instruct-2501
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: fp8_kv
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: true
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/mistral/engine-mixtral-8x22b-instruct/README.md b/mistral/engine-mixtral-8x22b-instruct/README.md
deleted file mode 100644
index e41b6265c..000000000
--- a/mistral/engine-mixtral-8x22b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Mistral 8x22B Instruct
-
-This deployment of Mistral 8x22B Instruct uses the TensorRT-LLM Engine Builder.
-
-For details, see: https://docs.baseten.co/performance/examples/mistral-trt
diff --git a/mistral/engine-mixtral-8x22b-instruct/config.yaml b/mistral/engine-mixtral-8x22b-instruct/config.yaml
deleted file mode 100644
index f5574031b..000000000
--- a/mistral/engine-mixtral-8x22b-instruct/config.yaml
+++ /dev/null
@@ -1,45 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are a knowledgable, engaging, geology teacher.",
-          },
-          {
-            role: "user",
-            content: "What is the impact of the Mistral wind on the French climate?",
-          },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: mistralai/Mixtral-8x22B-Instruct-v0.1
-model_name: Mistral 8x22B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:2
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mixtral-8x22B-Instruct-v0.1
-      source: HF
-    max_seq_len: 8192
-    num_builder_gpus: 4
-    quantization_type: fp8_kv
-    tensor_parallel_count: 2
diff --git a/mistral/engine-mixtral-8x7b-instruct/README.md b/mistral/engine-mixtral-8x7b-instruct/README.md
deleted file mode 100644
index 8f08fc2d6..000000000
--- a/mistral/engine-mixtral-8x7b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Mistral 8x7B Instruct
-
-This deployment of Mistral 8x7B Instruct uses the TensorRT-LLM Engine Builder.
-
-For details, see: https://docs.baseten.co/performance/examples/mistral-trt
diff --git a/mistral/engine-mixtral-8x7b-instruct/config.yaml b/mistral/engine-mixtral-8x7b-instruct/config.yaml
deleted file mode 100644
index c1f992e5c..000000000
--- a/mistral/engine-mixtral-8x7b-instruct/config.yaml
+++ /dev/null
@@ -1,53 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are a knowledgable, engaging, meteorology teacher.",
-          },
-          {
-            role: "user",
-            content: "What is the impact of the Mistral wind on the French climate?",
-          },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: mistralai/Mixtral-8x7B-Instruct-v0.1
-model_name: Mistral 8x7B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets:
-  hf_access_token: set token in baseten workspace
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: mistralai/Mixtral-8x7B-Instruct-v0.1
-      source: HF
-    num_builder_gpus: 2
-    quantization_type: fp8_kv
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: true
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/mistral/mistral-7b/README.md b/mistral/mistral-7b/README.md
deleted file mode 100644
index 608985a26..000000000
--- a/mistral/mistral-7b/README.md
+++ /dev/null
@@ -1,70 +0,0 @@
-# Mistral 7B Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mistral 7B. Mistral 7B parameter language model released by [Mistral](https://mistral.ai/) that outperforms other models in the 7B model class. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Mistral 7B.
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp)). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mistral-7b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mistral-7b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-This seven billion parameter model is running in `float16` so that it fits on an A10G.
-
-## Mistral 7B API documentation
-
-This section provides an overview of the Mistral 7B API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __stream__: (optional, default=False) A boolean if the model should stream a response back.
-- __max_new_tokens__ (optional, default=512): The maximum number of tokens to return, counting input tokens. Maximum of 4096.
-- __temperature__ (optional, default=0.1): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results.
-- __top_p__ (optional, default=0.75): The cumulative probability threshold for token sampling. The model will only consider tokens whose cumulative probability is below this threshold.
-- __top_k__ (optional, default=40): The number of top tokens to consider when sampling. The model will only consider the top_k highest-probability tokens.
-
-The API also supports passing any parameter supported by HuggingFace's `Transformers.generate`.
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "What is the meaning of life?", "max_new_tokens": 4096}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "What's the meaning of life?",
-           "max_new_tokens": 4096
-         }'
-```
diff --git a/mistral/mistral-7b/config.yaml b/mistral/mistral-7b/config.yaml
deleted file mode 100644
index ce64c041b..000000000
--- a/mistral/mistral-7b/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  example_model_input:
-    prompt: What is the Mistral wind?
-  pretty_name: Mistral 7B
-  tags:
-  - text-generation
-model_name: mistral-7b
-python_version: py311
-requirements:
-- transformers==4.42.3
-- sentencepiece
-- accelerate
-- torch==2.0.1
-- numpy==1.26.4
-resources:
-  accelerator: A10G
-  memory: 25Gi
-  use_gpu: true
-secrets:
-  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
-system_packages: []
diff --git a/mistral/mistral-small-3.1/config.yaml b/mistral/mistral-small-3.1/config.yaml
deleted file mode 100644
index 159319966..000000000
--- a/mistral/mistral-small-3.1/config.yaml
+++ /dev/null
@@ -1,56 +0,0 @@
-#vllm serve mistralai/Mistral-Small-3.1-24B-Instruct-2503 --tokenizer_mode mistral --config_format mistral --load_format mistral --tool-call-parser mistral --enable-auto-tool-choice --limit_mm_per_prompt 'image=10' --tensor-parallel-size 2
-base_image:
-  image: vllm/vllm-openai:v0.10.1.1
-model_metadata:
-  repo_id: mistralai/Mistral-Small-3.1-24B-Instruct-2503
-  example_model_input: {
-    "model": "mistral",
-    "messages": [
-      {
-      "role": "user",
-      "content": [
-        {
-        "type": "text",
-        "text": "Describe this image in one sentence."
-        },
-        {
-        "type": "image_url",
-        "image_url": {
-          "url": "https://picsum.photos/id/237/200/300"
-        }
-        }
-      ]
-      }
-    ],
-    "stream": true,
-    "max_tokens": 512,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-docker_server:
-  start_command: "sh -c \"b10-compile-cache & VLLM_USE_V1=1 HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve mistralai/Mistral-Small-3.1-24B-Instruct-2503 --tokenizer_mode mistral --config_format mistral --load_format mistral --tool-call-parser mistral --enable-auto-tool-choice --served-model-name mistral --max-num-seqs 8 --max-model-len 16384 --tensor-parallel-size 1 --gpu-memory-utilization 0.95\""
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
-  hf_access_token: null
-requirements:
-- huggingface_hub
-- hf_transfer
-- datasets
-- b10-transfer
-resources:
-  accelerator: H100:1
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  health_checks:
-    restart_check_delay_seconds: 300      # Waits 5 minutes after deployment before starting health checks
-    restart_threshold_seconds: 300       # Triggers a restart if health checks fail for 5 minutes
-    stop_traffic_threshold_seconds: 120  # Stops traffic if health checks fail for 2 minutes
-  predict_concurrency : 8
-model_name: Mistral Small 3.1
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/README.md b/mistral/mixtral-8x22b-trt-int8-weights-only/README.md
deleted file mode 100644
index ee9591770..000000000
--- a/mistral/mixtral-8x22b-trt-int8-weights-only/README.md
+++ /dev/null
@@ -1,52 +0,0 @@
-# Mixtral 8x22B Instruct Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mixtral 8x22B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Mistral 7B Instruct.
-
-**Warning: This example is only intended for usage on four A100 GPUs, changing your resource type for this deployment will result in unsupported behavior**
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten. Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-## Deploying Mixtral 8x22B Instruct
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mixtral-8x22b-trt-int8-weights-only
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x22b-trt-int8-weights-only` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Mixtral 8x22B Instruct API documentation
-This section provides an overview of the Mistral 8x22B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided instruction.
-
-### API route: `predict`
-
-We expect requests will the following information:
-
-- ```messages``` (str): The prompt you'd like to complete
-- ```max_tokens``` (int, default: 50): The max token count. This includes the number of tokens in your prompt so if this value is less than your prompt, you'll just recieve a truncated version of the prompt.
-- ```temperature``` (float, default: 0.7): Determines the creativity of the model output.
-- ```top_p``` (float, default: 0.7): Used to control model generation in conjunction with temperature.
-- ```top_k``` (int, default: 50): Chooses the token based on a distribution of the top K tokens.
-- ```beam_width``` (int, default:50): The number of beams to compute. This must be 1 for this version of TRT-LLM. Inflight-batching does not support beams > 1.
-- ```bad_words_list``` (list, default:[]): A list of words to not include in generated output.
-- ```stop_words_list``` (list, default:[]): A list of words to stop generation upon encountering.
-- ```repetition_penalty``` (float, defualt: 1.0): A repetition penalty to incentivize not repeating tokens.
-
-This Truss will stream responses back. Responses will be buffered chunks of text.
diff --git a/mistral/mixtral-8x22b-trt-int8-weights-only/config.yaml b/mistral/mixtral-8x22b-trt-int8-weights-only/config.yaml
deleted file mode 100644
index c6b5e6a8b..000000000
--- a/mistral/mixtral-8x22b-trt-int8-weights-only/config.yaml
+++ /dev/null
@@ -1,41 +0,0 @@
-base_image:
-  image: docker.io/baseten/triton_trt_llm:4062d46_20240401
-  python_executable_path: /usr/bin/python3
-description: Mixtral 8x22B Instruct, with INT8 weights only quantization, optimized
-  with TRT-LLM!
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  engine_repository: baseten/mixtral-8x22B_i60000_o4000_bs2_tp4_int8_weights_only_A100-tllm_0.9.0.dev2024022000
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: What is your favourite condiment?
-      role: user
-    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
-        adds just the right amount of zesty flavour to whatever I'm cooking up in
-        the kitchen!
-      role: assistant
-    - content: Do you have mayonnaise recipes?
-      role: user
-  tags:
-  - text-generation
-  - openai-compatible
-  tensor_parallelism: 4
-  tokenizer_repository: mistralai/Mixtral-8x22B-Instruct-v0.1
-model_name: Mixtral 8x22B Instruct TRT-LLM Weights Only Quantized
-python_version: py311
-requirements:
-- tritonclient[all]
-- transformers==4.42.3
-resources:
-  accelerator: A100:4
-  use_gpu: true
-runtime:
-  num_workers: 1
-  predict_concurrency: 256
-secrets:
-  hf_access_token: "your-hf-access-token"
-system_packages: []
diff --git a/mistral/mixtral-8x22b/README.md b/mistral/mixtral-8x22b/README.md
deleted file mode 100644
index 026856769..000000000
--- a/mistral/mixtral-8x22b/README.md
+++ /dev/null
@@ -1,78 +0,0 @@
-# Mixtral 8x22B Truss
-
-This is a [Truss](https://truss.baseten.co/) for the community edition of [Mixtral 8x22B](https://huggingface.co/mistral-community/Mixtral-8x22B-v0.1).
-This is not an optimized model. If you would like to have a more optimized version that has lower latency + higher throughput, please contact our team.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mistral/mixtral-8x22b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x22b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-You need four A100s to run Mixtral at `fp16`. If you need access to A100s, please [contact us](mailto:support@baseten.co).
-
-## Mixtral 8x22B API documentation
-
-This section provides an overview of the Mixtral 8x22B API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The `predict` route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __stream__ (optional, default=True): A boolean determining whether the model should stream a response back. When `True`, the API returns generated text as it becomes available.
-- __max_tokens__ (optional, default=128): Determines the maximum number of tokens to generate
-- __temperature__ (optional, default=1.0): Controls the strength of the generation. The higher the temperature, the more diverse and creative the output would be.
-- __top_p__ (optional, default=0.95): Parameter used to control the randomness of the output.
-- __top_k__ (optional, default=50): Controls the vocab size considered during the generation.
-
-## Example usage
-
-```python
-import requests
-import os
-
-# Replace the empty string with your model id below
-model_id = ""
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-
-data = {
-    "prompt": "What is mistral wind?",
-    "stream": True,
-    "max_tokens": 256,
-    "temperature": 0.9
-}
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/production/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=data,
-    stream=True
-)
-
-# Print the generated tokens as they get streamed
-for content in res.iter_content():
-    print(content.decode("utf-8"), end="", flush=True)
-```
diff --git a/mistral/mixtral-8x22b/config.yaml b/mistral/mixtral-8x22b/config.yaml
deleted file mode 100644
index d469b4b35..000000000
--- a/mistral/mixtral-8x22b/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  repo_id: mistralai/Mixtral-8x22B-Instruct-v0.1
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  example_model_input:
-    prompt: What is the Mistral wind?
-  pretty_name: Mistral 8x22B
-  tags:
-  - text-generation
-model_name: Mixtral 8x22
-python_version: py310
-requirements:
-  - accelerate
-  - transformers==4.42.3
-  - torch==2.2.0
-resources:
-  accelerator: A100:4
-  use_gpu: true
-secrets:
-  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
-system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/README.md b/mistral/mixtral-8x7b-instruct-trt-llm-h100/README.md
deleted file mode 100644
index 23e3f9142..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm-h100/README.md
+++ /dev/null
@@ -1,54 +0,0 @@
-# Mixtral-8x7B-Instruct Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mixtral 8x7B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Mistral 7B Instruct.
-
-**Warning: This example is intended for usage on two H100s, changing your resource type for this deployment will result in unsupported behavior**
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten. Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-## Deploying Mixtral-8x7B-Instruct
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mixtral-8x7b-instruct-trt-llm-h100
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x7b-instruct-trt-llm-h100` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Mixtral 8x7B Instruct API documentation
-This section provides an overview of the Mistral 7B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided instruction.
-
-### API route: `predict`
-
-This model is designed for our ChatCompletions endpoint:
-
-- [ChatCompletions endpoint tutorial](https://www.baseten.co/blog/gpt-vs-mistral-migrate-to-open-source-llms-with-minor-code-changes/)
-- [ChatCompletions endpoint reference docs](https://docs.baseten.co/api-reference/openai)
-
-We expect requests will the following information:
-
-- ```messages``` (str): The prompt you'd like to complete
-- ```max_tokens``` (int, default: 50): The max token count. This includes the number of tokens in your prompt so if this value is less than your prompt, you'll just recieve a truncated version of the prompt.
-- ```beam_width``` (int, default:50): The number of beams to compute. This must be 1 for this version of TRT-LLM. Inflight-batching does not support beams > 1.
-- ```bad_words_list``` (list, default:[]): A list of words to not include in generated output.
-- ```stop_words_list``` (list, default:[]): A list of words to stop generation upon encountering.
-- ```repetition_penalty``` (float, defualt: 1.0): A repetition penalty to incentivize not repeating tokens.
-
-This Truss will stream responses back. Responses will be buffered chunks of text.
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-h100/config.yaml b/mistral/mixtral-8x7b-instruct-trt-llm-h100/config.yaml
deleted file mode 100644
index 8889d1a52..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm-h100/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-base_image:
-  image: docker.io/baseten/trtllm-server:r23.12_baseten_v0.7.1
-  python_executable_path: /usr/bin/python3
-description: Mixtral 8x7B Instruct optimized with TRT-LLM! Compatible with OpenAI
-  Client
-environment_variables:
-  HF_HUB_ENABLE_HF_TRANSFER: 1
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  engine_repository: baseten/mixtral-h100-0.7.1
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: What is your favourite condiment?
-      role: user
-    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
-        adds just the right amount of zesty flavour to whatever I'm cooking up in
-        the kitchen!
-      role: assistant
-    - content: Do you have mayonnaise recipes?
-      role: user
-  tags:
-  - text-generation
-  - openai-compatible
-  tensor_parallelism: 2
-  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
-model_name: Mixtral 8x7B Instruct TRT-LLM for H100
-python_version: py311
-requirements:
-- tritonclient[all]
-- transformers==4.42.3
-- jinja2==3.1.3
-- hf_transfer==0.1.5
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 256
-secrets: {}
-system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/README.md b/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/README.md
deleted file mode 100644
index 3e4984352..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/README.md
+++ /dev/null
@@ -1,54 +0,0 @@
-# Mixtral-8x7B-Instruct Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mixtral 8x7B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Mistral 7B Instruct.
-
-**Warning: This example is only intended for usage on a single H100, changing your resource type for this deployment will result in unsupported behavior**
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten. Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-## Deploying Mixtral-8x7B-Instruct-Chat
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x7b-instruct-trt-llm-weights-only-quant` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Mixtral 8x7B Instruct API documentation
-This section provides an overview of the Mistral 7B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided instruction.
-
-### API route: `predict`
-
-This model is designed for our ChatCompletions endpoint:
-
-- [ChatCompletions endpoint tutorial](https://www.baseten.co/blog/gpt-vs-mistral-migrate-to-open-source-llms-with-minor-code-changes/)
-- [ChatCompletions endpoint reference docs](https://docs.baseten.co/api-reference/openai)
-
-We expect requests will the following information:
-
-- ```messages``` (str): The prompt you'd like to complete
-- ```max_tokens``` (int, default: 50): The max token count. This includes the number of tokens in your prompt so if this value is less than your prompt, you'll just recieve a truncated version of the prompt.
-- ```beam_width``` (int, default:50): The number of beams to compute. This must be 1 for this version of TRT-LLM. Inflight-batching does not support beams > 1.
-- ```bad_words_list``` (list, default:[]): A list of words to not include in generated output.
-- ```stop_words_list``` (list, default:[]): A list of words to stop generation upon encountering.
-- ```repetition_penalty``` (float, defualt: 1.0): A repetition penalty to incentivize not repeating tokens.
-
-This Truss will stream responses back. Responses will be buffered chunks of text.
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/config.yaml b/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/config.yaml
deleted file mode 100644
index 68a244d8e..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant-h100/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-base_image:
-  image: docker.io/baseten/trtllm-server:r23.12_baseten_v0.7.1
-  python_executable_path: /usr/bin/python3
-description: Mixtral 8x7B Instruct, with INT8 weights only quantization, optimized
-  with TRT-LLM! Compatible with OpenAI Client
-environment_variables:
-  HF_HUB_ENABLE_HF_TRANSFER: 1
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  engine_repository: baseten/mixtral-weights-only-quantized-h100-0.7.1
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: What is your favourite condiment?
-      role: user
-    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
-        adds just the right amount of zesty flavour to whatever I'm cooking up in
-        the kitchen!
-      role: assistant
-    - content: Do you have mayonnaise recipes?
-      role: user
-  tags:
-  - text-generation
-  - openai-compatible
-  tensor_parallelism: 1
-  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
-model_name: Mixtral 8x7B Instruct TRT-LLM Weights Only Quantized for H100
-python_version: py311
-requirements:
-- tritonclient[all]
-- transformers==4.42.3
-- jinja2==3.1.3
-- hf_transfer==0.1.5
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency: 256
-secrets: {}
-system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/README.md b/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/README.md
deleted file mode 100644
index 51acaa368..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/README.md
+++ /dev/null
@@ -1,54 +0,0 @@
-# Mixtral-8x7B-Instruct Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mixtral 8x7B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Mistral 7B Instruct.
-
-**Warning: This example is only intended for usage on a single A100, changing your resource type for this deployment will result in unsupported behavior**
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten. Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-## Deploying Mixtral-8x7B-Instruct-Chat
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mixtral-8x7b-instruct-trt-llm-weights-only-quant
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x7b-instruct-trt-llm-weights-only-quant` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Mixtral 8x7B Instruct API documentation
-This section provides an overview of the Mistral 7B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided instruction.
-
-### API route: `predict`
-
-This model is designed for our ChatCompletions endpoint:
-
-- [ChatCompletions endpoint tutorial](https://www.baseten.co/blog/gpt-vs-mistral-migrate-to-open-source-llms-with-minor-code-changes/)
-- [ChatCompletions endpoint reference docs](https://docs.baseten.co/api-reference/openai)
-
-We expect requests will the following information:
-
-- ```messages``` (str): The prompt you'd like to complete
-- ```max_tokens``` (int, default: 50): The max token count. This includes the number of tokens in your prompt so if this value is less than your prompt, you'll just recieve a truncated version of the prompt.
-- ```beam_width``` (int, default:50): The number of beams to compute. This must be 1 for this version of TRT-LLM. Inflight-batching does not support beams > 1.
-- ```bad_words_list``` (list, default:[]): A list of words to not include in generated output.
-- ```stop_words_list``` (list, default:[]): A list of words to stop generation upon encountering.
-- ```repetition_penalty``` (float, defualt: 1.0): A repetition penalty to incentivize not repeating tokens.
-
-This Truss will stream responses back. Responses will be buffered chunks of text.
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/config.yaml b/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/config.yaml
deleted file mode 100644
index cfafdd705..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm-weights-only-quant/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-base_image:
-  image: docker.io/baseten/triton_trt_llm:main-20231215
-  python_executable_path: /usr/bin/python3
-description: Mixtral 8x7B Instruct, with INT8 weights only quantization, optimized
-  with TRT-LLM! Compatible with OpenAI Client
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  engine_repository: baseten/mixtral-weights-only-quantized
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: What is your favourite condiment?
-      role: user
-    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
-        adds just the right amount of zesty flavour to whatever I'm cooking up in
-        the kitchen!
-      role: assistant
-    - content: Do you have mayonnaise recipes?
-      role: user
-  tags:
-  - text-generation
-  - openai-compatible
-  tensor_parallelism: 1
-  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
-  repo_id: mistralai/Mixtral-8x7B-v0.1
-model_name: Mixtral 8x7B Instruct TRT-LLM Weights Only Quantized
-python_version: py311
-requirements:
-- tritonclient[all]
-- transformers==4.42.3
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  num_workers: 1
-  predict_concurrency: 256
-secrets:
-  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
-system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/README.md b/mistral/mixtral-8x7b-instruct-trt-llm/README.md
deleted file mode 100644
index e15f8d1ac..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm/README.md
+++ /dev/null
@@ -1,54 +0,0 @@
-# Mixtral-8x7B-Instruct Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mixtral 8x7B Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Mistral 7B Instruct.
-
-**Warning: This example is only intended for usage on a single A100, changing your resource type for this deployment will result in unsupported behavior**
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten. Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-## Deploying Mixtral-8x7B-Instruct
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mixtral-8x7b-instruct-trt-llm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x7b-instruct-trt-llm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Mixtral 8x7B Instruct API documentation
-This section provides an overview of the Mistral 7B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided instruction.
-
-### API route: `predict`
-
-This model is designed for our ChatCompletions endpoint:
-
-- [ChatCompletions endpoint tutorial](https://www.baseten.co/blog/gpt-vs-mistral-migrate-to-open-source-llms-with-minor-code-changes/)
-- [ChatCompletions endpoint reference docs](https://docs.baseten.co/api-reference/openai)
-
-We expect requests will the following information:
-
-- ```messages``` (str): The prompt you'd like to complete
-- ```max_tokens``` (int, default: 50): The max token count. This includes the number of tokens in your prompt so if this value is less than your prompt, you'll just recieve a truncated version of the prompt.
-- ```beam_width``` (int, default:50): The number of beams to compute. This must be 1 for this version of TRT-LLM. Inflight-batching does not support beams > 1.
-- ```bad_words_list``` (list, default:[]): A list of words to not include in generated output.
-- ```stop_words_list``` (list, default:[]): A list of words to stop generation upon encountering.
-- ```repetition_penalty``` (float, defualt: 1.0): A repetition penalty to incentivize not repeating tokens.
-
-This Truss will stream responses back. Responses will be buffered chunks of text.
diff --git a/mistral/mixtral-8x7b-instruct-trt-llm/config.yaml b/mistral/mixtral-8x7b-instruct-trt-llm/config.yaml
deleted file mode 100644
index 15bff9bab..000000000
--- a/mistral/mixtral-8x7b-instruct-trt-llm/config.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-base_image:
-  image: docker.io/baseten/triton_trt_llm:main-20231215
-  python_executable_path: /usr/bin/python3
-description: Mixtral 8x7B Instruct optimized with TRT-LLM! Compatible with OpenAI
-  Client
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  engine_repository: baseten/mixtral
-  example_model_input:
-    max_tokens: 512
-    messages:
-    - content: What is your favourite condiment?
-      role: user
-    - content: Well, I'm quite partial to a good squeeze of fresh lemon juice. It
-        adds just the right amount of zesty flavour to whatever I'm cooking up in
-        the kitchen!
-      role: assistant
-    - content: Do you have mayonnaise recipes?
-      role: user
-  tags:
-  - text-generation
-  - openai-compatible
-  tensor_parallelism: 2
-  tokenizer_repository: mistralai/Mixtral-8x7B-v0.1
-model_name: Mixtral 8x7B Instruct TRT-LLM
-python_version: py311
-requirements:
-- tritonclient[all]
-- transformers==4.42.3
-resources:
-  accelerator: A100:2
-  use_gpu: true
-runtime:
-  num_workers: 1
-  predict_concurrency: 256
-secrets: {}
-system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/README.md b/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/README.md
deleted file mode 100644
index d21ff0eda..000000000
--- a/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/README.md
+++ /dev/null
@@ -1,62 +0,0 @@
-# Mixtral 8x7B Instruct Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mixtral 8x7B Instruct. Mixtral 8x7B Instruct parameter language model released by [Mistral AI](https://mistral.ai/). It is a mixture-of-experts (MoE) model. This README will walk you through how to deploy this Truss on Baseten to get your own instance of it.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mixtral-8x7b-instruct-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x7b-instruct-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-You need two A100s to run Mixtral at `fp16`. If you need access to A100s, please [contact us](mailto:support@baseten.co).
-
-## Mixtral 8x7B Instruct API documentation
-
-This section provides an overview of the Mixtral 8x7B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The `predict` route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __stream__ (optional, default=False): A boolean determining whether the model should stream a response back. When `True`, the API returns generated text as it becomes available.
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "What is the Mistral wind?"}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "What is the meaning of life? Answer in substantial detail with multiple examples from famous philosophies, religions, and schools of thought.",
-           "stream": true,
-           "max_tokens": 4096
-         }' --no-buffer
-```
diff --git a/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/config.yaml b/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/config.yaml
deleted file mode 100644
index ccc476bea..000000000
--- a/mistral/mixtral-8x7b-instruct-vllm-a100-t-tp2/config.yaml
+++ /dev/null
@@ -1,13 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: Mixtral 8x7B — VLLM TP2 — A100:2
-python_version: py310
-requirements:
-- vllm
-resources:
-  accelerator: A100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets: {}
-system_packages: []
diff --git a/mistral/mixtral-8x7b-instruct-vllm/README.md b/mistral/mixtral-8x7b-instruct-vllm/README.md
deleted file mode 100644
index d21ff0eda..000000000
--- a/mistral/mixtral-8x7b-instruct-vllm/README.md
+++ /dev/null
@@ -1,62 +0,0 @@
-# Mixtral 8x7B Instruct Truss
-
-This is a [Truss](https://truss.baseten.co/) for Mixtral 8x7B Instruct. Mixtral 8x7B Instruct parameter language model released by [Mistral AI](https://mistral.ai/). It is a mixture-of-experts (MoE) model. This README will walk you through how to deploy this Truss on Baseten to get your own instance of it.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mixtral-8x7b-instruct-vllm
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `mixtral-8x7b-instruct-vllm` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-You need two A100s to run Mixtral at `fp16`. If you need access to A100s, please [contact us](mailto:support@baseten.co).
-
-## Mixtral 8x7B Instruct API documentation
-
-This section provides an overview of the Mixtral 8x7B Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The `predict` route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __stream__ (optional, default=False): A boolean determining whether the model should stream a response back. When `True`, the API returns generated text as it becomes available.
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "What is the Mistral wind?"}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "What is the meaning of life? Answer in substantial detail with multiple examples from famous philosophies, religions, and schools of thought.",
-           "stream": true,
-           "max_tokens": 4096
-         }' --no-buffer
-```
diff --git a/mistral/mixtral-8x7b-instruct-vllm/config.yaml b/mistral/mixtral-8x7b-instruct-vllm/config.yaml
deleted file mode 100644
index f219fc452..000000000
--- a/mistral/mixtral-8x7b-instruct-vllm/config.yaml
+++ /dev/null
@@ -1,13 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: Mixtral 8x7B
-python_version: py310
-requirements:
-- vllm==0.2.5
-resources:
-  accelerator: A100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets: {}
-system_packages: []
diff --git a/mistral/pixtral-12b/README.md b/mistral/pixtral-12b/README.md
deleted file mode 100644
index 5f6420ff6..000000000
--- a/mistral/pixtral-12b/README.md
+++ /dev/null
@@ -1,84 +0,0 @@
-# Pixtral 12B Truss
-
-This is a [Truss](https://truss.baseten.co/) for Pixtral 12B. Pixtral 12B parameter language model released by [Mistral AI](https://mistral.ai/) and is a multimodal (text + vision) LLM. This README will walk you through how to deploy this Truss on Baseten to get your own instance of it.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd mistral/pixtral-12b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. Accept the terms of service of the Pixtral model [here](https://huggingface.co/mistralai/Pixtral-12B-2409).
-4. Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens).
-5. Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`. Note that you will *not* be able to successfully deploy Pixtral without doing this.
-
-With `pixtral-12b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --publish --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-You need one A100 to run Pixtral at `bf16`.
-
-## Pixtral 12B API documentation
-
-This section provides an overview of the Pixtral 12B API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The `predict` route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __messages__: The input in OpenAI API format (see below for examples)
-- __stream__ (optional, default=False): A boolean determining whether the model should stream a response back. When `True`, the API returns generated text as it becomes available.
-- __max_tokens__ (optional, default=512): Maximum number of tokens to generate
-- __temperature__ (optional, default=0.7): A float between 0 and 1. Higher values means the generated output is more random while lower values means the generated output is more determenistic
-
-## Example usage
-
-```sh
-truss predict -d '{"messages": [{"role": "user", "content": "Tell me about yourself"}]}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "messages": [
-            {
-                "role": "user",
-                "content": [
-                {
-                    "type": "text",
-                    "text": "What type of animal is this? Answer in French only"
-                },
-                {
-                    "type": "image_url",
-                    "image_url": {
-                    "url": "https://vetmed.illinois.edu/wp-content/uploads/2021/04/pc-keller-hedgehog.jpg"
-                    }
-                }
-                ]
-            }
-            ],
-           "stream": true,
-           "max_tokens": 64,
-           "temperature": 0.2
-         }' --no-buffer
-```
diff --git a/mistral/pixtral-12b/config.yaml b/mistral/pixtral-12b/config.yaml
deleted file mode 100644
index 5e1d0dba8..000000000
--- a/mistral/pixtral-12b/config.yaml
+++ /dev/null
@@ -1,43 +0,0 @@
-model_metadata:
-  repo_id: mistral-community/pixtral-12b-240910
-  avatar_url: https://cdn.baseten.co/production/static/explore/mistral_logo.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/mistral.png
-  example_model_input: {
-    messages: [
-      {
-        role: user,
-        content: [
-          {
-            type: text,
-            text: "Describe this image in one sentence."
-          },
-          {
-            type: image_url,
-            image_url: {
-              url: "https://picsum.photos/id/237/200/300"
-            }
-          }
-        ]
-      }
-    ],
-    stream: false,
-    max_tokens: 512,
-    temperature: 0.5
-  }
-  vllm_config:
-    tensor_parallel_size: 1
-    max_model_len: 16384
-    max_num_batched_tokens: 16384
-    limit_mm_per_prompt: {image: 5}
-  tags:
-    - text-generation
-    - multimodal
-model_name: Pixtral 12B
-python_version: py311
-secrets:
-  hf_access_token: null
-requirements:
-  - vllm==0.6.1
-resources:
-  accelerator: A100
-  use_gpu: true
diff --git a/mistral/voxtral-streaming-4b/config.yaml b/mistral/voxtral-streaming-4b/config.yaml
deleted file mode 100644
index f7afea296..000000000
--- a/mistral/voxtral-streaming-4b/config.yaml
+++ /dev/null
@@ -1,38 +0,0 @@
-model_name: Voxtral-Mini-4B-Realtime-2602
-secrets:
-  hf_access_token: null
-environment_variables:
-  VLLM_DISABLE_COMPILE_CACHE: "1"
-base_image:
-  image: vllm/vllm-openai:nightly-d88a1df699f68e5284fe3a3170f8ae292a3e9c3f
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) VLLM_DISABLE_COMPILE_CACHE=1 vllm serve mistralai/Voxtral-Mini-4B-Realtime-2602 --compilation-config '{\"cudagraph_mode\":\"PIECEWISE\"}' --host 0.0.0.0 --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/realtime
-  server_port: 8000
-resources:
-  accelerator: H100_40GB:1
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-requirements:
-  - --pre --extra-index-url https://wheels.vllm.ai/nightly
-  - vllm[audio]
-  - librosa
-  - torch
-  - torchaudio
-  - pynvml
-  - ffmpeg-python
-  - websockets
-system_packages:
-  - python3.10-venv
-  - ffmpeg
-  - openmpi-bin
-  - libopenmpi-dev
-runtime:
-  is_websocket_endpoint: true
-  transport:
-    kind: websocket
-    ping_interval_seconds: null
-    ping_timeout_seconds: null
diff --git a/model_cach_gcs/config.yaml b/model_cach_gcs/config.yaml
deleted file mode 100644
index 953ca8e2c..000000000
--- a/model_cach_gcs/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-model_name: Hello Model Cache GCS
-python_version: py311
-requirements: [""]
-resources:
-  accelerator: null
-  cpu: "1"
-  memory: 2Gi
-  use_gpu: false
-build:
-  secret_to_path_mapping:
-    gcs-service-account-jsn: /secrets/gcs-service-account-jsn
-secrets: { gcs-service-account-jsn: null } # null is encouraged, as this will automatically use the one provided by baseten.co
-model_cache:
-  # this is a simple private bucket, uploaded llama-3-2, as is from huggingface.
-  # The repo contains e.g. gs://llama-3-2-1b-instruct/config.json / gs://llama-3-2-1b-instruct/model.safetensors
-  - repo_id: gs://llama-3-2-1b-instruct/
-    revision: main
-    use_volume: true
-    volume_folder: llama
-    runtime_secret_name: "gcs-service-account-jsn"
-    kind: "gcs"
diff --git a/model_cach_gcs/model/model.py b/model_cach_gcs/model/model.py
deleted file mode 100644
index e949d452b..000000000
--- a/model_cach_gcs/model/model.py
+++ /dev/null
@@ -1,32 +0,0 @@
-# <- download is invoked before here.
-import pathlib
-
-
-class Model:
-    """example usage of `model_cache` in truss"""
-
-    def __init__(self, *args, **kwargs):
-        # `lazy_data_resolver` is passed as keyword-argument in init
-        self._lazy_data_resolver = kwargs["lazy_data_resolver"]
-        self.tensor_size = None
-
-    def load(self):
-        # work that does not require the download may be done beforehand
-        # important to collect the download before using any incomplete data
-        self._lazy_data_resolver.block_until_download_complete()
-        # after the call, you may use the /app/model_cache directory and the contents
-        # torch.load(
-        #     "/app/model_cache/stable-diffusion-xl-base/vae_1_0/diffusion_pytorch_model.fp16.safetensors",
-        #     weights_only=True
-        # )
-        self.tensor_size = (
-            pathlib.Path("/app/model_cache/llama/model.safetensors").stat().st_size
-        )
-        print(
-            "Model loaded successfully with size of {} bytes".format(self.tensor_size)
-        )
-
-    def predict(self, input_data):
-        # this method will be called by the serving container
-        # you may use the model here, after the download is complete
-        return {"input": input_data, "tensor_size": self.tensor_size}
diff --git a/model_cache/config.yaml b/model_cache/config.yaml
deleted file mode 100644
index 5129c2096..000000000
--- a/model_cache/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_name: Hello Model Cache Qwen
-python_version: py311
-requirements: ["torch"]
-resources:
-  accelerator: null
-  cpu: "1"
-  memory: 8Gi
-  use_gpu: false
-secrets: { hf_access_token: null } # null is encouraged, as this will automatically use the one provided by baseten.co
-model_cache:
-  - repo_id: madebyollin/sdxl-vae-fp16-fix
-    revision: 207b116dae70ace3637169f1ddd2434b91b3a8cd
-    use_volume: true
-    volume_folder: sdxl-vae-fp16
-    allow_patterns:
-      - config.json
-      - diffusion_pytorch_model.safetensors
-    runtime_secret_name: hf_access_token
-    kind: "hf"
-  - repo_id: stabilityai/stable-diffusion-xl-base-1.0
-    revision: 462165984030d82259a11f4367a4eed129e94a7b
-    use_volume: true
-    volume_folder: stable-diffusion-xl-base
-    allow_patterns:
-      - "*.json"
-      - "*.fp16.safetensors"
-      - sd_xl_base_1.0.safetensors
-    runtime_secret_name: hf_access_token
-    kind: "hf"
diff --git a/multiprocessing/README.md b/multiprocessing/README.md
deleted file mode 100644
index 359dd10dc..000000000
--- a/multiprocessing/README.md
+++ /dev/null
@@ -1,17 +0,0 @@
-# Using Multiprocessing for pre- or post-processing steps
-
-Truss server executes the steps `preprocssing`, `predict` and `postprocess` (or a sub set) using python asyncio. This allows concurrency - and therefore higher throughput - while waiting for I/O-bound tasks (if asyncio-supporting libraries are used e.g. fetching data with async HTTP requrests).
-However, it does not increase throughput, if the work in any of those steps is CPU-bound (either by blocking I/O functions or compute intense pre/post-processing operations).
-
-Specifically, if pre/post-processing is CPU-bound, it can lead to underutilization of the server's GPU resources and suboptimal throughput.
-In this case, simply increasing the number server replicas is not efficient, because it would likewise increase the GPU footprint.
-
-A possible solution is to move the CPU work into separate processes that do not block the main server from using the GPU for *different* requests in the meantime.
-
-This dummy model illustrates such a setup. To adopt it, the overall inference pipeline must  be structured such that most of the CPU work goes into the pre- or post-processing methods and the `predict` method contains mostly GPU work.
-
-
-## Comments on the example code:
-* Multiprocessing only demonstrated for `preprocessing`, for `postprocessing` the same pattern (with same process pool) can be used.
-* Multiprocessing depends on pickling the function to run (and all its context) to the process pool. If the function depends on unpickleable objects or "large" objects this can be problematic and a different design is needed (see comments in code).
-* The worker pool size is determined by the number of CPUs visible to the server.
diff --git a/multiprocessing/config.yaml b/multiprocessing/config.yaml
deleted file mode 100644
index d8652b4e3..000000000
--- a/multiprocessing/config.yaml
+++ /dev/null
@@ -1,13 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: Model with multiprocessing pre/post-process
-python_version: py310
-requirements:
-- torch
-resources:
-  accelerator: A10G
-  cpu: '8'
-  memory: 8Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/musicgen-large/README.md b/musicgen-large/README.md
deleted file mode 100644
index 236e18c70..000000000
--- a/musicgen-large/README.md
+++ /dev/null
@@ -1,87 +0,0 @@
-# MusicGen Truss
-
-This repository packages [MusicGen](https://github.com/facebookresearch/audiocraft/) as a [Truss](https://truss.baseten.co).
-
-MusicGen is a simple and controllable model for music generation developed by Facebook AI Research.
-
-Utilizing this model for inference can be challenging given the hardware requirements. With Baseten and Truss, inference is dead simple.
-
-## Deploying MusicGen
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd musicgen-large-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `musicgen-large-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-We found this model runs reasonably fast on A10Gs; you can configure the hardware you'd like in the config.yaml.
-
-```yaml
-resources:
-  cpu: "3"
-  memory: 14Gi
-  use_gpu: true
-  accelerator: A10G
-```
-
-## Invoking MusicGen
-
-MusicGen takes a list of prompts and a duration in seconds. It will generate one clip per prompt and return each clip as a base64 encoded WAV file.
-
-```sh
-truss predict -d '{"prompts": ["happy rock" "energetic EDM", "sad jazz"], "duration": 8}'
-```
-
-You'll want to pipe your results into a script such as:
-
-```python
-import json
-import base64
-import os, sys
-
-model_output = json.loads(sys.stdin.read())
-
-for idx, clip in enumerate(model_output["data"]):
-  with open(f"clip_{idx}.wav", "wb") as f:
-    f.write(base64.b64decode(clip))
-```
-
-You can also invoke your model via a REST API
-
-```
-curl -X POST " https://app.baseten.co/models/YOUR_MODEL_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompts": ["happy rock" "energetic EDM", "sad jazz"], "duration": 8
-         }'
-```
-
-## Model sizes
-
-MusicGen supports four model sizes:
-
-- `small`: 300M model, text to music only
-- `medium`: 1.5B model, text to music only
-- `melody`: 1.5B model, text to music and text+melody to music
-- `large`: 3.3B model, text to music only
-
-This truss can been configured to run the large size but you can easily select other versions by changing the `MODEL_SIZE` constant in `model/model.py`.
diff --git a/musicgen-large/config.yaml b/musicgen-large/config.yaml
deleted file mode 100644
index 3f6a7e9f7..000000000
--- a/musicgen-large/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-description: MusicGen is a simple and controllable model for music generation developed
-  by Facebook AI Research.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/musicgen-cover.png
-  example_model_input:
-    duration: 8
-    prompts:
-    - happy rock
-    - energetic EDM
-    - sad jazz
-  tags:
-  - text-to-music
-model_name: MusicGen large
-python_version: py39
-requirements:
-- torch>=2
-- audiocraft
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/musicgen-melody/README.md b/musicgen-melody/README.md
deleted file mode 100644
index 54c132af6..000000000
--- a/musicgen-melody/README.md
+++ /dev/null
@@ -1,78 +0,0 @@
-# MusicGen Truss
-
-This repository packages the `melody` model from [MusicGen](https://github.com/facebookresearch/audiocraft/) as a [Truss](https://truss.baseten.co).
-
-MusicGen is a simple and controllable suite of models for music generation developed by Facebook AI Research. The `melody` model accepts both text and audio to condition it's outputs.
-
-Utilizing this model for inference can be challenging given the hardware requirements. With Baseten and Truss, inference is dead simple.
-
-## Deploying MusicGen
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd musicgen-melody-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `musicgen-melody-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-We found this model runs reasonably fast on A10Gs; you can configure the hardware you'd like in the config.yaml.
-
-```yaml
-resources:
-  cpu: "3"
-  memory: 14Gi
-  use_gpu: true
-  accelerator: A10G
-```
-
-## Invoking MusicGen
-
-MusicGen takes a list of prompts and a duration in seconds. You may also, optionally, provide a base64 encoded WAV file as the melody to condition on. It will generate one clip per prompt and return each clip as a base64 encoded WAV file.
-
-```sh
-truss predict -d '{"prompts": ["happy rock" "energetic EDM", "sad jazz"], "melody" : "b64_encoded_melody", "duration": 8}'
-```
-
-You'll want to pipe your results into a script such as:
-
-```python
-import json
-import base64
-import os, sys
-
-model_output = json.loads(sys.stdin.read())
-
-for idx, clip in enumerate(model_output["data"]):
-  with open(f"clip_{idx}.wav", "wb") as f:
-    f.write(base64.b64decode(clip))
-```
-
-You can also invoke your model via a REST API
-
-```
-curl -X POST " https://app.baseten.co/models/YOUR_MODEL_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompts": ["happy rock" "energetic EDM", "sad jazz"],
-           "melody" : "b64_encoded_melody",
-           "duration": 8
-         }'
-```
diff --git a/musicgen-melody/config.yaml b/musicgen-melody/config.yaml
deleted file mode 100644
index ad71431a4..000000000
--- a/musicgen-melody/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-description: MusicGen Melody is a simple and controllable model for music generation
-  conditioned on text and audio. It is developed by Facebook AI Research.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/meta.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/musicgen-cover.png
-  example_model_input:
-    duration: 8
-    prompts:
-    - happy rock
-    - energetic EDM
-    - sad jazz
-  tags:
-  - text-to-music
-model_name: MusicGen Melody
-python_version: py39
-requirements:
-- torch>=2
-- audiocraft
-- protobuf
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/nemotron/Llama-3-1-Nemotron-Nano-VL-8B-V1/config.yaml b/nemotron/Llama-3-1-Nemotron-Nano-VL-8B-V1/config.yaml
deleted file mode 100644
index f4d752aa4..000000000
--- a/nemotron/Llama-3-1-Nemotron-Nano-VL-8B-V1/config.yaml
+++ /dev/null
@@ -1,47 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.11.0
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful vision-language assistant."
-      - role: user
-        content:
-          - type: image
-            url: "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
-          - type: text
-            text: "Describe this image in detail."
-    stream: true
-    model: "nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1"
-    max_tokens: 1024
-    temperature: 0.7
-  tags:
-    - openai-compatible
-model_name: Llama 3.1 Nemotron Nano VL 8B V1
-requirements:
-  - transformers>=4.55.0
-  - accelerate
-  - timm
-  - einops
-  - open-clip-torch
-  - pillow
-python_version: py312
-model_cache:
-  - repo_id: nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1
-    revision: main
-    use_volume: true
-    volume_folder: "llama-3-1-nemotron-nano-vl-8b-v1"
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-docker_server:
-  start_command: vllm serve nvidia/Llama-3.1-Nemotron-Nano-VL-8B-V1 --tensor-parallel-size 1 --served-model-name llama-3-1-nemotron-nano-vl-8b-v1 --trust-remote-code --max-model-len 16384 --gpu-memory-utilization 0.9
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 16
diff --git a/nemotron/Nemotron-3-Nano-NVFP4/config.yaml b/nemotron/Nemotron-3-Nano-NVFP4/config.yaml
deleted file mode 100644
index 87ffeb8bb..000000000
--- a/nemotron/Nemotron-3-Nano-NVFP4/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.12.0
-model_metadata:
-  example_model_input: {
-    messages: [
-      {
-        role: "user",
-        content: "Write me a short story about a cat."
-      }
-    ],
-    stream: true,
-    max_tokens: 512,
-    temperature: 0.6
-  }
-  repo_id: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: sh -c "wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py -O /app/nano_v3_reasoning_parser.py && HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-NVFP4 --max-num-seqs 8 --tensor-parallel-size 1 --max-model-len 262144 --port 8000 --trust-remote-code --tool-call-parser qwen3_coder --reasoning-parser-plugin /app/nano_v3_reasoning_parser.py --reasoning-parser nano_v3"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: B200
-  use_gpu: true
-runtime:
-  predict_concurrency : 8
-model_name: Nemotron 3 Nano
-environment_variables:
-  hf_access_token: null
-system_packages:
-  - wget
diff --git a/nemotron/Nemotron-3-Nano/config.yaml b/nemotron/Nemotron-3-Nano/config.yaml
deleted file mode 100644
index cb506ed2d..000000000
--- a/nemotron/Nemotron-3-Nano/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.12.0
-model_metadata:
-  example_model_input: {
-    messages: [
-      {
-        role: "user",
-        content: "Write me a short story about a cat."
-      }
-    ],
-    stream: true,
-    max_tokens: 512,
-    temperature: 0.6
-  }
-  repo_id: nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: sh -c "wget https://huggingface.co/nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16/resolve/main/nano_v3_reasoning_parser.py -O /app/nano_v3_reasoning_parser.py && HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve nvidia/NVIDIA-Nemotron-3-Nano-30B-A3B-BF16 --max-num-seqs 8 --tensor-parallel-size 1 --max-model-len 262144 --port 8000 --trust-remote-code --tool-call-parser qwen3_coder --reasoning-parser-plugin /app/nano_v3_reasoning_parser.py --reasoning-parser nano_v3"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency : 8
-model_name: Nemotron 3 Nano
-environment_variables:
-  hf_access_token: null
-system_packages:
-  - wget
diff --git a/nemotron/Nemotron-Nano-12B-v2-VL-BF16/config.yaml b/nemotron/Nemotron-Nano-12B-v2-VL-BF16/config.yaml
deleted file mode 100644
index 77dafd856..000000000
--- a/nemotron/Nemotron-Nano-12B-v2-VL-BF16/config.yaml
+++ /dev/null
@@ -1,44 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.11.0
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    model: ""
-    messages:
-      - role: user
-        content:
-          - type: image_url
-            image_url:
-              url: "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
-          - type: text
-            text: "Describe this image in detail."
-    stream: true
-  tags:
-    - openai-compatible
-model_name: NVIDIA Nemotron Nano 12B v2 VL BF16
-requirements:
-  - transformers>=4.55.0
-  - accelerate
-  - timm
-  - einops
-  - open-clip-torch
-  - pillow
-python_version: py312
-model_cache:
-  - repo_id: nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16
-    revision: main
-    use_volume: true
-    volume_folder: "nvidia-nemotron-nano-12b-v2-vl-bf16"
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-docker_server:
-  start_command: vllm serve nvidia/NVIDIA-Nemotron-Nano-12B-v2-VL-BF16 --tensor-parallel-size 1 --served-model-name nvidia-nemotron-nano-12b-v2-vl-bf16 --trust-remote-code --max-model-len 16384 --gpu-memory-utilization 0.9
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 16
diff --git a/nemotron/llama-3-1-nemotron-70b-instruct/README.md b/nemotron/llama-3-1-nemotron-70b-instruct/README.md
deleted file mode 100644
index 33bd03802..000000000
--- a/nemotron/llama-3-1-nemotron-70b-instruct/README.md
+++ /dev/null
@@ -1,6 +0,0 @@
-
-# Llama-3.1-Nemotron-70B-Instruct
-
-This deployment of Llama-3.1-Nemotron-70B-Instruct uses the TensorRT-LLM Engine Builder.
-
-For details, see: https://docs.baseten.co/performance/examples/llama-trt
diff --git a/nemotron/llama-3-1-nemotron-70b-instruct/config.yaml b/nemotron/llama-3-1-nemotron-70b-instruct/config.yaml
deleted file mode 100644
index a56d044fa..000000000
--- a/nemotron/llama-3-1-nemotron-70b-instruct/config.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    {
-      messages: [{ role: "user", content: "How many r in strawberry?" }],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.6,
-    }
-  repo_id: nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
-model_name: Llama-3.1-Nemotron-70B-Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:2
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: nvidia/Llama-3.1-Nemotron-70B-Instruct-HF
-      source: HF
-    num_builder_gpus: 4
-    quantization_type: fp8_kv
-    max_seq_len: 131072
-    tensor_parallel_count: 2
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: true
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 131072
diff --git a/nemotron/llama-nemoretriever-colembed-3b-v1/README.md b/nemotron/llama-nemoretriever-colembed-3b-v1/README.md
deleted file mode 100644
index 9e8c14b48..000000000
--- a/nemotron/llama-nemoretriever-colembed-3b-v1/README.md
+++ /dev/null
@@ -1,9 +0,0 @@
-# NVIDIA Llama NemoRetriever ColEmbed 3B V1
-
-This is a traditional [Truss](https://truss.baseten.co/) implementation for [NVIDIA's Llama NemoRetriever ColEmbed 3B V1](https://huggingface.co/nvidia/llama-nemoretriever-colembed-3b-v1) cross-modal embedding model. This implementation uses the standard `load()` and `predict()` pattern with a `model.py` file.
-
-## Support
-
-For questions or issues, please refer to:
-- [Truss Documentation](https://docs.baseten.co)
-- [Model Card](https://huggingface.co/nvidia/llama-nemoretriever-colembed-3b-v1)
diff --git a/nemotron/nemotron-ultra-253b/config.yaml b/nemotron/nemotron-ultra-253b/config.yaml
deleted file mode 100644
index c4a35d1a0..000000000
--- a/nemotron/nemotron-ultra-253b/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-      - content: detailed thinking on
-        role: system
-      - content: Tell me everything you know about optimized inference.
-        role: user
-    stream: true
-    temperature: 0.5
-  tags:
-    - openai-compatible
-model_name: Briton-nemotron-253b-tp8-fp8
-resources:
-  accelerator: H100:8
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      # pre-quanitzed checkpoint in plain FP8
-      repo: michaelfeil/nemotron-251b-ultra-v2-tp8-fp8-tllm
-      source: HF
-    max_batch_size: 64
-    max_seq_len: 65536
-    quantization_type: fp8
-    tensor_parallel_count: 8
-    speculator:
-      lookahead_ngram_size: 5
-      lookahead_verification_set_size: 5
-      lookahead_windows_size: 7
-      num_draft_tokens: 47
-      speculative_decoding_mode: LOOKAHEAD_DECODING
diff --git a/ngram-speculator/truss/config.yaml b/ngram-speculator/truss/config.yaml
deleted file mode 100644
index ff9ea5ff7..000000000
--- a/ngram-speculator/truss/config.yaml
+++ /dev/null
@@ -1,14 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata: {}
-model_name: ngram-speculator
-python_version: py310
-requirements:
-- vllm==0.6.5
-- transformers==4.47.1
-resources:
-  accelerator: H100
-  use_gpu: True
-secrets: {}
-system_packages: []
diff --git a/ngram-speculator/trussless/config.yaml b/ngram-speculator/trussless/config.yaml
deleted file mode 100644
index c755bfa58..000000000
--- a/ngram-speculator/trussless/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:nightly
-model_metadata:
-  repo_id: Qwen/Qwen3-0.6B
-  example_model_input: {
-    "model": "llama",
-    "messages": [
-      {
-      "role": "user",
-      "content": [
-        {
-        "type": "text",
-        "text": "What do llamas dream of?"
-        }
-      ]
-      }
-    ],
-    "stream": false,
-    "max_tokens": 512,
-  }
-docker_server:
-  start_command: sh -c "vllm serve distilbert/distilgpt2 --served-model-name llama --tensor-parallel-size 1 --max-model-len 64 --max-num-seqs 1 --max-num-batched-tokens 64 --gpu-memory-utilization 0.7"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-runtime:
-  predict_concurrency : 16
-resources:
-  accelerator: H100
-  use_gpu: true
-model_name: ngram-speculator
-environment_variables:
-  hf_access_token: null
diff --git a/nous-capybara/nous-capybara-34b-openai/README.md b/nous-capybara/nous-capybara-34b-openai/README.md
deleted file mode 100644
index a37905174..000000000
--- a/nous-capybara/nous-capybara-34b-openai/README.md
+++ /dev/null
@@ -1,75 +0,0 @@
-# Nous Capybara 34B Truss (OpenAI Client Compatible)
-
-This is a [Truss](https://truss.baseten.co/) for [Nous Capybara 34B](https://huggingface.co/NousResearch/Nous-Capybara-34B), compatible with our [bridge endpoint for OpenAI ChatCompletion users](https://docs.baseten.co/api-reference/openai).
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd nous-capybara/nous-capybara-34b-openai
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `nous-capybara/nous-capybara-34b-openai` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Model API reference
-
-This model is designed for our ChatCompletions endpoint:
-
-- [ChatCompletions endpoint tutorial](https://www.baseten.co/blog/gpt-vs-mistral-migrate-to-open-source-llms-with-minor-code-changes/)
-- [ChatCompletions endpoint reference docs](https://docs.baseten.co/api-reference/openai)
-
-Note that Nous Capybara currently does not support system messages (see [here](https://huggingface.co/NousResearch/Nous-Capybara-34B/discussions/5)).
-
-See the script below for an example of calling the model, with and without streaming:
-```python
-from openai import OpenAI
-import os
-
-model_id = "YOUR_MODEL_ID"
-client = OpenAI(
-    api_key="YOUR_BASETEN_API_KEY",
-    base_url=f"https://bridge.baseten.co/{model_id}/v1"
-)
-
-# Non-streaming example
-response = client.chat.completions.create(
-    model="nous-capybara-34b",
-    messages=[
-            {"role": "user", "content": "What happens if I go to the top of the tallest mountain in California with a bucket of water and tip it over the highest cliff?"},
-    ],
-    stream=False,
-)
-
-print(response.choices[0].message.content)
-
-# Streaming example
-response = client.chat.completions.create(
-    model="nous-capybara-34b",
-    messages=[
-            {"role": "user", "content": "Who won the world series in 2020?"},
-            {"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
-            {"role": "user", "content": "Where was it played?"}
-    ],
-    stream=True,
-)
-
-for chunk in response:
-    content = chunk.choices[0].delta.content
-    if content:
-        print(content, end="")
-```
diff --git a/nous-capybara/nous-capybara-34b-openai/config.yaml b/nous-capybara/nous-capybara-34b-openai/config.yaml
deleted file mode 100644
index f071d9d60..000000000
--- a/nous-capybara/nous-capybara-34b-openai/config.yaml
+++ /dev/null
@@ -1,24 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_cache:
-- allow_patterns:
-  - '*.json'
-  - '*.bin'
-  repo_id: NousResearch/Nous-Capybara-34B
-  use_volume: false
-model_name: Nous Capybara 34B OpenAI
-python_version: py310
-requirements:
-- accelerate==0.25.0
-- transformers==4.35.2
-- torch==2.1.0
-- bitsandbytes==0.41.3
-- scipy==1.11.4
-- sentencepiece==0.1.99
-resources:
-  accelerator: A100
-  cpu: '3'
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/nous-capybara/nous-capybara-34b-openai/model/model.py b/nous-capybara/nous-capybara-34b-openai/model/model.py
deleted file mode 100644
index d0d124d00..000000000
--- a/nous-capybara/nous-capybara-34b-openai/model/model.py
+++ /dev/null
@@ -1,115 +0,0 @@
-from threading import Thread
-
-import torch
-from transformers import (
-    AutoModelForCausalLM,
-    AutoTokenizer,
-    GenerationConfig,
-    TextIteratorStreamer,
-)
-
-MODEL_NAME = "NousResearch/Nous-Capybara-34B"
-
-DEFAULT_STREAM = False
-
-
-def _format_prompt(messages: list[dict], add_generation_prompt: bool = True) -> str:
-    """Given a list of messages in the form: [{'role': 'user', 'content': 'hello world'}], returns
-    the messages as a string in the form that Nous-Capybara expects:
-    'USER: hello world
-    ASSISTANT:'
-
-    This is a workaround for Nous-Capybara not being configured with a chat template (see https://huggingface.co/NousResearch/Nous-Capybara-34B/discussions/5)
-    """
-    formatted_prompts = []
-    for message in messages:
-        if message["role"] == "user":
-            formatted_prompts.append(f"USER: {message['content']}")
-        elif message["role"] == "assistant":
-            formatted_prompts.append(f"ASSISTANT: {message['content']}")
-        # Note: Capybara doesn't support system messages. See https://huggingface.co/NousResearch/Nous-Capybara-34B/discussions/8
-    if add_generation_prompt:
-        formatted_prompts.append("ASSISTANT:")
-
-    return "\n".join(formatted_prompts)
-
-
-class Model:
-    def __init__(self, **kwargs):
-        self.model = None
-        self.tokenizer = None
-
-    def load(self):
-        self.model = AutoModelForCausalLM.from_pretrained(
-            MODEL_NAME,
-            device_map="auto",
-            torch_dtype=torch.float16,
-            trust_remote_code=True,
-        )
-        self.tokenizer = AutoTokenizer.from_pretrained(
-            MODEL_NAME, trust_remote_code=True
-        )
-
-    def preprocess(self, request: dict):
-        # Set generate_args to default values.
-        default_generate_args = {
-            "max_new_tokens": 256,
-            "temperature": 0.7,
-            "top_p": 0.8,
-            "top_k": 40,
-            "repetition_penalty": 1.3,
-            "no_repeat_ngram_size": 5,
-            "use_cache": True,
-            "do_sample": True,
-            "eos_token_id": self.tokenizer.eos_token_id,
-            "pad_token_id": self.tokenizer.pad_token_id,
-        }
-        request["generate_args"] = default_generate_args
-
-        # Override generate_args with values provided in the user's request.
-        for k in default_generate_args:
-            if k in request and request[k] is not None:
-                request["generate_args"][k] = request[k]
-
-        return request
-
-    def stream(self, input_ids: list, generation_args: dict):
-        streamer = TextIteratorStreamer(self.tokenizer)
-        generation_config = GenerationConfig(**generation_args)
-        generation_kwargs = {
-            "input_ids": input_ids,
-            "generation_config": generation_config,
-            "return_dict_in_generate": True,
-            "output_scores": True,
-            "max_new_tokens": generation_args["max_new_tokens"],
-            "streamer": streamer,
-        }
-
-        with torch.no_grad():
-            # Begin generation in a separate thread
-            thread = Thread(target=self.model.generate, kwargs=generation_kwargs)
-            thread.start()
-
-            # Yield generated text as it becomes available
-            def inner():
-                for text in streamer:
-                    yield text
-                thread.join()
-
-        return inner()
-
-    def predict(self, request: dict):
-        stream = request.pop("stream", DEFAULT_STREAM)
-        messages = request.pop("messages")
-        input_ids = self.tokenizer(
-            _format_prompt(messages), return_tensors="pt"
-        ).input_ids.cuda()
-
-        generation_args = request.pop("generate_args")
-
-        if stream:
-            return self.stream(input_ids, generation_args)
-
-        with torch.no_grad():
-            outputs = self.model.generate(inputs=input_ids, **generation_args)
-            return self.tokenizer.decode(outputs[0], skip_special_tokens=True)
diff --git a/nous-capybara/nous-capybara-34b/README.md b/nous-capybara/nous-capybara-34b/README.md
deleted file mode 100644
index cf72d77f6..000000000
--- a/nous-capybara/nous-capybara-34b/README.md
+++ /dev/null
@@ -1,97 +0,0 @@
-# Nous Capybara 34B Truss
-
-This is a [Truss](https://truss.baseten.co/) for [Nous Capybara 34B](https://huggingface.co/NousResearch/Nous-Capybara-34B). This model is a fine-tuned version of Yi-34B with a 200K context length.
-
-## Truss
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd nous-capybara/nous-capybara-34b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `nous-capybara/nous-capybara-34b` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-
-### Hardware notes
-
-This 34 billion parameter model requires an A100 GPU.
-
-## Nous Capybara 34B Chat API documentation
-
-This section provides an overview of the Nous-Capybara-34B model, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The input text that you want the model to generate a response for.
-- __max_tokens__ (optional, default=256): The maximum number of tokens to return, counting input tokens.
-- __temperature__ (optional, default=0.7): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results.
-- __top_p__ (optional, default=0.8): The cumulative probability threshold for token sampling. The model will only consider tokens whose cumulative probability is below this threshold.
-- __top_k__ (optional, default=40): The number of top tokens to consider when sampling. The model will only consider the top_k highest-probability tokens.
-- __repetition_penalty__ (optional, default=1.3): Helps the model generate more diverse content instead of repeating previous phrases.
-- __no_repeat_ngram_size__ (optional, default=5): Specifies the length of token sets that are completely blocked from repeating at all.
-- __stream__ (optional, default=True): Allows you to receive the tokens are they are generated in a stream like fashion.
-
-The API also supports passing any parameter supported by HuggingFace's `Transformers.generate`.
-
-The output of the model is a JSON object which only have one key called `output`. Here is an example of what that JSON object looks like:
-```json
-{"output": "There's a place where time stands still. A place of breath taking wonder, but also deep mystery and danger; the ocean floor.."}
-```
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "There is a place where time stands still. A place of breath taking wonder, but also", "max_tokens": 512}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "What happens if I go to the top of the tallest mountian in california with a bucket of water and tip it over the highest cliff?",
-           "max_tokens": 512,
-           "stream": False
-         }'
-```
-
-By default streaming the tokens is enabled. Here is an example of how to invoke the model with streaming in Python:
-```python
-import requests
-headers = {"Authorization": f"Api-Key BASETEN-API-KEY"}
-
-res = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers=headers,
-    json={"prompt": "What happens if I go to the top of the tallest mountian in california with a bucket of water and tip it over the highest cliff?",
-          "max_tokens": 512, "temperature": 0.9, "stream": True},
-    stream=True
-)
-res.raise_for_status()
-
-for word in res:
-    print(word.decode("utf-8"))
-```
diff --git a/nous-capybara/nous-capybara-34b/config.yaml b/nous-capybara/nous-capybara-34b/config.yaml
deleted file mode 100644
index fc567d33e..000000000
--- a/nous-capybara/nous-capybara-34b/config.yaml
+++ /dev/null
@@ -1,22 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    prompt: What happens if I go to the top of the tallest mountian in california
-      with a bucket of water and tip it over the highest cliff?
-model_name: Nous Capybara 34B
-python_version: py310
-requirements:
-- accelerate==0.25.0
-- transformers==4.35.2
-- torch==2.1.0
-- bitsandbytes==0.41.3
-- scipy==1.11.4
-- sentencepiece==0.1.99
-resources:
-  accelerator: A100
-  cpu: '3'
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/nous-capybara/nous-capybara-34b/model/model.py b/nous-capybara/nous-capybara-34b/model/model.py
deleted file mode 100644
index 0b8f91bff..000000000
--- a/nous-capybara/nous-capybara-34b/model/model.py
+++ /dev/null
@@ -1,98 +0,0 @@
-from threading import Thread
-from typing import Dict
-
-import torch
-from transformers import (
-    AutoModelForCausalLM,
-    AutoTokenizer,
-    GenerationConfig,
-    TextIteratorStreamer,
-)
-
-MODEL_NAME = "NousResearch/Nous-Capybara-34B"
-MAX_LENGTH = 256
-DO_SAMPLE = True
-REPETITION_PENALTY = 1.3
-NO_REPEAT_NGRAM_SIZE = 5
-TEMPERATURE = 0.7
-TOP_K = 40
-TOP_P = 0.8
-DEFAULT_STREAM = True
-
-
-class Model:
-    def __init__(self, **kwargs):
-        self.model = None
-        self.tokenizer = None
-
-    def load(self):
-        self.model = AutoModelForCausalLM.from_pretrained(
-            MODEL_NAME,
-            device_map="auto",
-            torch_dtype=torch.float16,
-            trust_remote_code=True,
-        )
-        self.tokenizer = AutoTokenizer.from_pretrained(
-            MODEL_NAME, trust_remote_code=True
-        )
-
-    def preprocess(self, request: dict):
-        generate_args = {
-            "max_length": request.get("max_tokens", MAX_LENGTH),
-            "temperature": request.get("temperature", TEMPERATURE),
-            "top_p": request.get("top_p", TOP_P),
-            "top_k": request.get("top_k", TOP_K),
-            "repetition_penalty": request.get("repetition_penalty", REPETITION_PENALTY),
-            "no_repeat_ngram_size": request.get(
-                "no_repeat_ngram_size", NO_REPEAT_NGRAM_SIZE
-            ),
-            "do_sample": request.get("do_sample", DO_SAMPLE),
-            "use_cache": True,
-            "eos_token_id": self.tokenizer.eos_token_id,
-            "pad_token_id": self.tokenizer.pad_token_id,
-        }
-        request["generate_args"] = generate_args
-        return request
-
-    def stream(self, input_ids: list, generation_args: dict):
-        streamer = TextIteratorStreamer(self.tokenizer)
-        generation_config = GenerationConfig(**generation_args)
-        generation_kwargs = {
-            "input_ids": input_ids,
-            "generation_config": generation_config,
-            "return_dict_in_generate": True,
-            "output_scores": True,
-            "max_new_tokens": generation_args["max_length"],
-            "streamer": streamer,
-        }
-
-        with torch.no_grad():
-            # Begin generation in a separate thread
-            thread = Thread(target=self.model.generate, kwargs=generation_kwargs)
-            thread.start()
-
-            # Yield generated text as it becomes available
-            def inner():
-                for text in streamer:
-                    yield text
-                thread.join()
-
-        return inner()
-
-    def predict(self, model_input: Dict):
-        prompt = model_input.get("prompt")
-        stream = model_input.get("stream", DEFAULT_STREAM)
-        generation_args = model_input.pop("generate_args")
-
-        formatted_prompt = f"USER: {prompt}\n ASSISTANT:"
-        input_ids = self.tokenizer(
-            formatted_prompt, return_tensors="pt"
-        ).input_ids.cuda()
-
-        if stream:
-            return self.stream(input_ids, generation_args)
-
-        with torch.no_grad():
-            outputs = self.model.generate(inputs=input_ids, **generation_args)
-            model_output = self.tokenizer.decode(outputs[0], skip_special_tokens=True)
-            return {"output": model_output}
diff --git a/nsql/README.md b/nsql/README.md
deleted file mode 100644
index 74d2b5f82..000000000
--- a/nsql/README.md
+++ /dev/null
@@ -1,94 +0,0 @@
-# NSQL Truss
-
-This is a [Truss](https://truss.baseten.co/) for [Number Station](https://www.numbersstation.ai/)'s 350M parameter NSQL model. NSQL is a text-to-SQL foundation model, enabling users to query their databases using natual language. There are also 2B and 6B NSQL variants available, which you can alternatively deploy by editing the HuggingFace paths in `model/model.py`.
-
-This README will walk you through how to deploy this Truss on Baseten to get your own instance of NSQL 350M.
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-## Deploying NSQL
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd nsql-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `nsql-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## NSQL API documentation
-
-This section provides an overview of the NSQL API, its parameters, and how to use it. The API consists of a single route named `predict`, which you can invoke to generate text based on the provided instruction.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given instruction. It takes several parameters:
-
-- **schema**: An SQL schema for the table you want to query. You can provide multiple schemas as a single string.
-- **query**: A natural language query over the provided database schemas.
-
-## Example usage
-
-You can use the `baseten` model package to invoke your model from Python
-
-```
-import baseten
-# You can retrieve your deployed model ID from the UI
-model = baseten.deployed_model_version_id('YOUR_MODEL_ID')
-
-schema = """CREATE TABLE stadium (
-    stadium_id number,
-    location text,
-    name text,
-    capacity number,
-    highest number,
-    lowest number,
-    average number
-)
-
-CREATE TABLE singer (
-    singer_id number,
-    name text,
-    country text,
-    song_name text,
-    song_release_year text,
-    age number,
-    is_male others
-)
-
-CREATE TABLE concert (
-    concert_id number,
-    concert_name text,
-    theme text,
-    stadium_id text,
-    year text
-)
-
-CREATE TABLE singer_in_concert (
-    concert_id number,
-    singer_id text
-)"""
-
-request = {
-    "schema": schema,
-    "query": "What is the maximum, the average, and the minimum capacity of stadiums?"
-}
-
-response = model.predict(request)
-```
diff --git a/nsql/config.yaml b/nsql/config.yaml
deleted file mode 100644
index e3863ecfe..000000000
--- a/nsql/config.yaml
+++ /dev/null
@@ -1,20 +0,0 @@
-description: NSQL is an open-source text-to-SQL AI model developed by Numbers Station.
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://aeiljuispo.cloudimg.io/v7/https://cdn-uploads.huggingface.co/production/uploads/649c7ee8f97bd6fd710a9eb5/nBg1Fyo22RrqRJrkz9IYB.png
-  cover_image_url: https://global-uploads.webflow.com/6348b2d49808811e3f7a0fff/640690727b722a05771960ec_graphic-data-p-800.png
-  tags:
-  - code-generation
-model_name: NSQL 350M
-python_version: py39
-requirements:
-- torch
-- transformers>=4.29.0
-resources:
-  accelerator: A10G
-  cpu: '8'
-  memory: 30Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/nvidia/parakeet-tdt-0_6b-v2/config.yaml b/nvidia/parakeet-tdt-0_6b-v2/config.yaml
deleted file mode 100644
index a8ffc2aa8..000000000
--- a/nvidia/parakeet-tdt-0_6b-v2/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-description: Parakeet TDT 0.6B V2 is a 600-million-parameter automatic speech recognition (ASR) model designed for high-quality English transcription.
-python_version: py312
-model_metadata:
-  repo_id: nvidia/parakeet-tdt-0.6b-v2
-  avatar_url: https://cdn-avatars.huggingface.co/v1/production/uploads/1613114437487-60262a8e0703121c822a80b6.png
-  example_model_input:
-    {
-      "audio_url": "https://dldata-public.s3.us-east-2.amazonaws.com/2086-149220-0033.wav",
-      "timestamps": false
-    }
-system_packages:
-  - ffmpeg
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-runtime:
-  predict_concurrency: 8
-model_name: Parakeet TDT 0.6B V2
-secrets:
-  hf_access_token: null
-requirements:
-  - nemo_toolkit[asr]
-  - requests
-  - pyarrow==20.0.0
-  - cuda-python>=12.3
diff --git a/openai/gpt-oss-120b/README.md b/openai/gpt-oss-120b/README.md
deleted file mode 100644
index b960a93d6..000000000
--- a/openai/gpt-oss-120b/README.md
+++ /dev/null
@@ -1,72 +0,0 @@
-# GPT OSS 120B with BISv2 — High-Throughput Template
-
-GPT OSS 120B is OpenAI's open source model designed for powerful reasoning, agentic tasks and other developer use cases. It uses their open source response format, Harmony.
-
-This directory contains a **[Truss](https://truss.baseten.co/)** template for deploying **GPT OSS 120B** with **Baseten Inference Stack v2 (TensorRT-LLM + PyTorch backend)** on 4 H100 GPUs. This truss fully abstracts OpenAI's harmony response format, so everything works outside of the box. You can simply use it like a regular OpenAI compatible server. This stack maximizes both inference and throughput.
-
----
-
-# Requirements
-
-`truss==0.10.5`
-
-You also need the file in data, which downloads GPT's harmony encoding ahead of time, because once deployed, the deployment will be unable to download from internet.
-
-The environment variable `TIKTOKEN_RS_CACHE_DIR: /app/data` in `config.yaml` points `openai_harmony` to the local encoding file. See this(discussion)[https://huggingface.co/openai/gpt-oss-120b/discussions/39] for details.
-
----
-
-
-## Core TRT-LLM `runtime` parameters
-
-| Property (YAML path)  | Value                | Why it matters |
-| --------------------- | -------------------- | -------------- |
-| `tensor_parallel_size`| **4** | Shards every weight matrix across the 2 H100s |
-| `moe_expert_parallel_size` | **4** | Shards each expert across 2 H100s |
-| `max_batch_size`      | **64** | Up to 64 concurrent requests per forward pass |
-| `max_seq_len`         | **98304** | 98304 token context length |
-| `enable_chunked_prefill` | `true` | Chunks long prompts to reduce memory usage |
-| `max_num_tokens`      | **8192** | Upper limit on total tokens per chunk |
-| `served_model_name`   | `openai/gpt-oss-120b` | `model: openai/gpt-oss-120b` to call this model in OpenAI Compatible server |
-
----
-
-## Important Advanced **`runtime.patch_kwargs`** parameters
-
-These map 1-to-1 to TensorRT-LLM flags for extra performance tuning.
-
-| Property (YAML path)                    | Value / Setting | Effect |
-| --------------------------------------- | --------------- | ------ |
-| `cuda_graph_config.enable_padding`      | `true`          | Pad to fixed shape so one CUDA Graph is reused every step |
-| `kv_cache_config.free_gpu_memory_fraction` | **0.8** | 80 % of post-load VRAM reserved for paged KV-cache |
-| `kv_cache_config.enable_block_reuse`    | `true`          | Identical prefixes share cache blocks → faster TTFT |
-| `kv_cache_config.enable_block_reuse`    | `true`          | Identical prefixes share cache blocks → faster TTFT |
-| `chat_processor`                        | `harmony`       | GPT OSS uses Harmony response format |
-
----
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd openai/gpt-oss-120b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `openai/gpt-oss-120b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted --publish
-```
-
-Paste your Baseten API key if prompted. Also ensure the `hf_access_token` secret is properly setup in your Baseten Account to access this model.
-
-**Note**: TensorRT-LLM with PyTorch Backend will only work under a Baseten production deployment
-
-For more information, refer to the [Truss documentation](https://docs.baseten.co/performance/engine-builder-overview).
diff --git a/openai/gpt-oss-120b/config.yaml b/openai/gpt-oss-120b/config.yaml
deleted file mode 100644
index c76fdaf12..000000000
--- a/openai/gpt-oss-120b/config.yaml
+++ /dev/null
@@ -1,69 +0,0 @@
-model_name: GPT OSS 120B
-build_commands:
-  - python -c 'from openai_harmony import load_harmony_encoding; load_harmony_encoding("HarmonyGptOss")'
-model_metadata:
-  repo_id: openai/gpt-oss-120b
-  example_model_input: {
-    "model": "openai/gpt-oss-120b",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    "stream": true,
-    "max_tokens": 2048,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-resources:
-  accelerator: B200
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-model_cache:
-  - repo_id: openai/gpt-oss-120b
-    revision: refs/pr/35
-    use_volume: true
-    volume_folder: trt_model
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: michaelfeil/empty-model
-      revision: main
-      source: HF
-  inference_stack: v2
-  runtime:
-    enable_chunked_prefill: true
-    max_batch_size: 64
-    max_num_tokens: 8192
-    max_seq_len: 131072
-    patch_kwargs:
-      model_path: /app/model_cache/trt_model
-      chat_processor: harmony
-      moe_expert_parallel_size: 1
-      backend: pytorch
-      cuda_graph_config:
-        enable_padding: true
-      disable_overlap_scheduler: 1
-      enable_autotuner: 0
-      enable_iter_perf_stats: 0
-      enable_trtllm_sampler: 1
-      guided_decoding_backend: xgrammar
-      kv_cache_config:
-        enable_block_reuse: true
-        free_gpu_memory_fraction: 0.8
-        event_buffer_max_size: 1024
-      max_beam_width: 1
-      max_input_len: 131072
-      model_level_stop_words:
-      - "<|call|>"
-      tokenizer_limit_length: 131072
-      trust_remote_code: 1
-      moe_config:
-        backend: TRTLLM
-    served_model_name: openai/gpt-oss-120b
-    tensor_parallel_size: 1
-  version_overrides:
-    v2_llm_version: null
diff --git a/openai/gpt-oss-20b/README.md b/openai/gpt-oss-20b/README.md
deleted file mode 100644
index 4ef0a85c2..000000000
--- a/openai/gpt-oss-20b/README.md
+++ /dev/null
@@ -1,72 +0,0 @@
-# GPT OSS 20B with BISv2 — High-Throughput Template
-
-GPT OSS 20B is OpenAI's open source model designed for powerful reasoning, agentic tasks and other developer use cases. It uses their open source response format, Harmony.
-
-This directory contains a **[Truss](https://truss.baseten.co/)** template for deploying **GPT OSS 20B** with **Baseten Inference Stack v2 (TensorRT-LLM + PyTorch backend)** on 4 H100 GPUs. This truss fully abstracts OpenAI's harmony response format, so everything works outside of the box. You can simply use it like a regular OpenAI compatible server. This stack maximizes both inference and throughput.
-
----
-
-# Requirements
-
-`truss==0.10.5`
-
-You also need the file in data, which downloads GPT's harmony encoding ahead of time, because once deployed, the deployment will be unable to download from internet.
-
-The environment variable `TIKTOKEN_RS_CACHE_DIR: /app/data` in `config.yaml` points `openai_harmony` to the local encoding file. See this(discussion)[https://huggingface.co/openai/gpt-oss-20b/discussions/39] for details.
-
----
-
-
-## Core TRT-LLM `runtime` parameters
-
-| Property (YAML path)  | Value                | Why it matters |
-| --------------------- | -------------------- | -------------- |
-| `tensor_parallel_size`| **4** | Shards every weight matrix across the 2 H100s |
-| `moe_expert_parallel_size` | **4** | Shards each expert across 2 H100s |
-| `max_batch_size`      | **64** | Up to 64 concurrent requests per forward pass |
-| `max_seq_len`         | **98304** | 98304 token context length |
-| `enable_chunked_prefill` | `true` | Chunks long prompts to reduce memory usage |
-| `max_num_tokens`      | **8192** | Upper limit on total tokens per chunk |
-| `served_model_name`   | `openai/gpt-oss-20b` | `model: openai/gpt-oss-20b` to call this model in OpenAI Compatible server |
-
----
-
-## Important Advanced **`runtime.patch_kwargs`** parameters
-
-These map 1-to-1 to TensorRT-LLM flags for extra performance tuning.
-
-| Property (YAML path)                    | Value / Setting | Effect |
-| --------------------------------------- | --------------- | ------ |
-| `cuda_graph_config.enable_padding`      | `true`          | Pad to fixed shape so one CUDA Graph is reused every step |
-| `kv_cache_config.free_gpu_memory_fraction` | **0.8** | 80 % of post-load VRAM reserved for paged KV-cache |
-| `kv_cache_config.enable_block_reuse`    | `true`          | Identical prefixes share cache blocks → faster TTFT |
-| `kv_cache_config.enable_block_reuse`    | `true`          | Identical prefixes share cache blocks → faster TTFT |
-| `chat_processor`                        | `harmony`       | GPT OSS uses Harmony response format |
-
----
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd openai/gpt-oss-20b
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `openai/gpt-oss-20b` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted --publish
-```
-
-Paste your Baseten API key if prompted. Also ensure the `hf_access_token` secret is properly setup in your Baseten Account to access this model.
-
-**Note**: TensorRT-LLM with PyTorch Backend will only work under a Baseten production deployment
-
-For more information, refer to the [Truss documentation](https://docs.baseten.co/performance/engine-builder-overview).
diff --git a/openai/gpt-oss-20b/config.yaml b/openai/gpt-oss-20b/config.yaml
deleted file mode 100644
index a6d4edbc0..000000000
--- a/openai/gpt-oss-20b/config.yaml
+++ /dev/null
@@ -1,69 +0,0 @@
-model_name: GPT OSS 20B
-build_commands:
-  - python -c 'from openai_harmony import load_harmony_encoding; load_harmony_encoding("HarmonyGptOss")'
-model_metadata:
-  repo_id: openai/gpt-oss-20b
-  example_model_input: {
-    "model": "openai/gpt-oss-20b",
-    "messages": [
-      {
-      "role": "user",
-      "content": "Given an array of integers nums and an integer target, return indices of the two numbers such that they add up to target. You may assume that each input would have exactly one solution, and you may not use the same element twice. You can return the answer in any order. class Solution: def twoSum(self, nums: List[int], target: int) -> List[int]:"
-      }
-    ],
-    "stream": true,
-    "max_tokens": 2048,
-    "temperature": 0.5
-  }
-  tags:
-  - openai-compatible
-resources:
-  accelerator: H100
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-model_cache:
-  - repo_id: openai/gpt-oss-20b
-    revision: refs/pr/36
-    use_volume: true
-    volume_folder: trt_model
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: michaelfeil/empty-model
-      revision: main
-      source: HF
-  inference_stack: v2
-  runtime:
-    enable_chunked_prefill: true
-    max_batch_size: 64
-    max_num_tokens: 8192
-    max_seq_len: 131072
-    patch_kwargs:
-      model_path: /app/model_cache/trt_model
-      chat_processor: harmony
-      moe_expert_parallel_size: 1
-      backend: pytorch
-      cuda_graph_config:
-        enable_padding: true
-      disable_overlap_scheduler: 1
-      enable_autotuner: 0
-      enable_iter_perf_stats: 0
-      enable_trtllm_sampler: 1
-      guided_decoding_backend: xgrammar
-      kv_cache_config:
-        enable_block_reuse: true
-        free_gpu_memory_fraction: 0.8
-        event_buffer_max_size: 1024
-      max_beam_width: 1
-      max_input_len: 131072
-      model_level_stop_words:
-      - "<|call|>"
-      tokenizer_limit_length: 131072
-      trust_remote_code: 1
-      moe_config:
-        backend: CUTLASS
-    served_model_name: openai/gpt-oss-20b
-    tensor_parallel_size: 1
-  version_overrides:
-    v2_llm_version: null
diff --git a/optimized/README.md b/optimized/README.md
new file mode 100644
index 000000000..3a20de593
--- /dev/null
+++ b/optimized/README.md
@@ -0,0 +1,22 @@
+# Optimized Models
+
+Production-grade, TensorRT-LLM optimized model configurations autogenerated by `_internal/templating/generate_templates.py`. These templates are not intended to be edited by hand.
+
+| Directory | Models | Description |
+|-----------|--------|-------------|
+| [briton](briton/) | 33 | Baseten Runtime Optimized Networks -- TRT-LLM configs for Llama, Qwen, Gemma, DeepSeek, Mistral, Phi, and Falcon models with FP4/FP8 quantization and speculative decoding variants |
+| [bisv2](bisv2/) | 11 | Baseten Inference Server v2 -- optimized serving configs for Llama, Qwen, DeepSeek, and NVIDIA models with FP4/FP8 quantization |
+
+These configurations are generated from templates. To regenerate them, run:
+
+```bash
+python _internal/templating/generate_templates.py
+```
+
+## Deploying
+
+Each optimized model can be deployed to Baseten with:
+
+```bash
+truss push
+```
diff --git a/optimized/bisv2/deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md b/optimized/bisv2/deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md
new file mode 100644
index 000000000..1dbd604af
--- /dev/null
+++ b/optimized/bisv2/deepseek-ai-deepseek-r1-distill-llama-70b-fp4/README.md
@@ -0,0 +1,56 @@
+# BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4-truss-example
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4_kv**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml b/optimized/bisv2/deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml
new file mode 100644
index 000000000..e973e54cb
--- /dev/null
+++ b/optimized/bisv2/deepseek-ai-deepseek-r1-distill-llama-70b-fp4/config.yaml
@@ -0,0 +1,30 @@
+description: "DeepSeek-R1-Distill-Llama-70B optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-deepseek-ai-deepseek-r1-distill-llama-70b-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
+      revision: main
+      source: HF
+    quantization_type: fp4_kv
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md
new file mode 100644
index 000000000..949dec9a0
--- /dev/null
+++ b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/README.md
@@ -0,0 +1,56 @@
+# BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only-truss-example
+
+Deploy [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 MLP ONLY |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.2-3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.2-3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4_mlp_only**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml
new file mode 100644
index 000000000..0ec59fab9
--- /dev/null
+++ b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp4-mlp-only/config.yaml
@@ -0,0 +1,30 @@
+description: "Llama-3.2-3B-Instruct optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-meta-llama-llama-3.2-3b-instruct-fp4-mlp-only-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.2-3B-Instruct
+      revision: main
+      source: HF
+    quantization_type: fp4_mlp_only
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp8/README.md b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp8/README.md
new file mode 100644
index 000000000..275a5d4a0
--- /dev/null
+++ b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp8/README.md
@@ -0,0 +1,56 @@
+# BISV2-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.2-3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.2-3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp8/config.yaml b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
new file mode 100644
index 000000000..fdfff00a8
--- /dev/null
+++ b/optimized/bisv2/meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
@@ -0,0 +1,30 @@
+description: "Llama-3.2-3B-Instruct optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.2-3B-Instruct
+      revision: main
+      source: HF
+    quantization_type: fp8_kv
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/meta-llama-llama-3.3-70b-instruct-fp4/README.md b/optimized/bisv2/meta-llama-llama-3.3-70b-instruct-fp4/README.md
new file mode 100644
index 000000000..8c071a6d4
--- /dev/null
+++ b/optimized/bisv2/meta-llama-llama-3.3-70b-instruct-fp4/README.md
@@ -0,0 +1,56 @@
+# BISV2-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
+
+Deploy [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.3-70B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.3-70B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/meta-llama-llama-3.3-70b-instruct-fp4/config.yaml b/optimized/bisv2/meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
new file mode 100644
index 000000000..7310a7367
--- /dev/null
+++ b/optimized/bisv2/meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
@@ -0,0 +1,30 @@
+description: "Llama-3.3-70B-Instruct optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.3-70B-Instruct
+      revision: main
+      source: HF
+    quantization_type: fp4
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/nvidia-llama-3.1-8b-instruct-fp4/README.md b/optimized/bisv2/nvidia-llama-3.1-8b-instruct-fp4/README.md
new file mode 100644
index 000000000..f0d209789
--- /dev/null
+++ b/optimized/bisv2/nvidia-llama-3.1-8b-instruct-fp4/README.md
@@ -0,0 +1,56 @@
+# BISV2-nvidia-llama-3.1-8b-instruct-fp4-truss-example
+
+Deploy [nvidia/Llama-3.1-8B-Instruct-FP4](https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP4) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/Llama-3.1-8B-Instruct-FP4](https://huggingface.co/nvidia/Llama-3.1-8B-Instruct-FP4) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="nvidia/Llama-3.1-8B-Instruct-FP4",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "nvidia/Llama-3.1-8B-Instruct-FP4", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/nvidia-llama-3.1-8b-instruct-fp4/config.yaml b/optimized/bisv2/nvidia-llama-3.1-8b-instruct-fp4/config.yaml
new file mode 100644
index 000000000..47a311760
--- /dev/null
+++ b/optimized/bisv2/nvidia-llama-3.1-8b-instruct-fp4/config.yaml
@@ -0,0 +1,30 @@
+description: "Llama-3.1-8B-Instruct-FP4 optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-nvidia-llama-3.1-8b-instruct-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: nvidia/Llama-3.1-8B-Instruct-FP4
+      revision: main
+      source: HF
+    quantization_type: no_quant
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/nvidia-qwen3-30b-a3b-fp4/README.md b/optimized/bisv2/nvidia-qwen3-30b-a3b-fp4/README.md
new file mode 100644
index 000000000..b6507308c
--- /dev/null
+++ b/optimized/bisv2/nvidia-qwen3-30b-a3b-fp4/README.md
@@ -0,0 +1,56 @@
+# BISV2-nvidia-qwen3-30b-a3b-fp4-truss-example
+
+Deploy [nvidia/Qwen3-30B-A3B-FP4](https://huggingface.co/nvidia/Qwen3-30B-A3B-FP4) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/Qwen3-30B-A3B-FP4](https://huggingface.co/nvidia/Qwen3-30B-A3B-FP4) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="nvidia/Qwen3-30B-A3B-FP4",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "nvidia/Qwen3-30B-A3B-FP4", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/nvidia-qwen3-30b-a3b-fp4/config.yaml b/optimized/bisv2/nvidia-qwen3-30b-a3b-fp4/config.yaml
new file mode 100644
index 000000000..5e335790f
--- /dev/null
+++ b/optimized/bisv2/nvidia-qwen3-30b-a3b-fp4/config.yaml
@@ -0,0 +1,30 @@
+description: "Qwen3-30B-A3B-FP4 optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-nvidia-qwen3-30b-a3b-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: nvidia/Qwen3-30B-A3B-FP4
+      revision: main
+      source: HF
+    quantization_type: no_quant
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/nvidia-qwen3-8b-fp4/README.md b/optimized/bisv2/nvidia-qwen3-8b-fp4/README.md
new file mode 100644
index 000000000..7ffc7a616
--- /dev/null
+++ b/optimized/bisv2/nvidia-qwen3-8b-fp4/README.md
@@ -0,0 +1,56 @@
+# BISV2-nvidia-qwen3-8b-fp4-truss-example
+
+Deploy [nvidia/Qwen3-8B-FP4](https://huggingface.co/nvidia/Qwen3-8B-FP4) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [nvidia/Qwen3-8B-FP4](https://huggingface.co/nvidia/Qwen3-8B-FP4) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="nvidia/Qwen3-8B-FP4",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "nvidia/Qwen3-8B-FP4", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/nvidia-qwen3-8b-fp4/config.yaml b/optimized/bisv2/nvidia-qwen3-8b-fp4/config.yaml
new file mode 100644
index 000000000..0ec18889b
--- /dev/null
+++ b/optimized/bisv2/nvidia-qwen3-8b-fp4/config.yaml
@@ -0,0 +1,30 @@
+description: "Qwen3-8B-FP4 optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-nvidia-qwen3-8b-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: nvidia/Qwen3-8B-FP4
+      revision: main
+      source: HF
+    quantization_type: no_quant
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct-fp4/README.md b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct-fp4/README.md
new file mode 100644
index 000000000..89769337c
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct-fp4/README.md
@@ -0,0 +1,56 @@
+# BISV2-qwen-qwen2.5-coder-7b-instruct-fp4-truss-example
+
+Deploy [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Coder-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Coder-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml
new file mode 100644
index 000000000..179430659
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct-fp4/config.yaml
@@ -0,0 +1,33 @@
+description: "Qwen2.5-Coder-7B-Instruct optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-qwen-qwen2.5-coder-7b-instruct-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Coder-7B-Instruct
+      revision: main
+      source: HF
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp4
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct/README.md b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct/README.md
new file mode 100644
index 000000000..36404cc09
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct/README.md
@@ -0,0 +1,56 @@
+# BISV2-qwen-qwen2.5-coder-7b-instruct-truss-example
+
+Deploy [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Coder-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Coder-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct/config.yaml b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct/config.yaml
new file mode 100644
index 000000000..d987b764b
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen2.5-coder-7b-instruct/config.yaml
@@ -0,0 +1,33 @@
+description: "Qwen2.5-Coder-7B-Instruct optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-qwen-qwen2.5-coder-7b-instruct-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Coder-7B-Instruct
+      revision: main
+      source: HF
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: no_quant
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/qwen-qwen3-32b-fp4/README.md b/optimized/bisv2/qwen-qwen3-32b-fp4/README.md
new file mode 100644
index 000000000..a164b7149
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen3-32b-fp4/README.md
@@ -0,0 +1,56 @@
+# BISV2-qwen-qwen3-32b-fp4-truss-example
+
+Deploy [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4_kv**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/qwen-qwen3-32b-fp4/config.yaml b/optimized/bisv2/qwen-qwen3-32b-fp4/config.yaml
new file mode 100644
index 000000000..486d1b63d
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen3-32b-fp4/config.yaml
@@ -0,0 +1,30 @@
+description: "Qwen3-32B optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-qwen-qwen3-32b-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: Qwen/Qwen3-32B
+      revision: main
+      source: HF
+    quantization_type: fp4_kv
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/bisv2/qwen-qwen3-4b-fp8/README.md b/optimized/bisv2/qwen-qwen3-4b-fp8/README.md
new file mode 100644
index 000000000..3ed75764c
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen3-4b-fp8/README.md
@@ -0,0 +1,56 @@
+# BISV2-qwen-qwen3-4b-fp8-truss-example
+
+Deploy [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-4B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-4B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Streaming: **enabled**
diff --git a/optimized/bisv2/qwen-qwen3-4b-fp8/config.yaml b/optimized/bisv2/qwen-qwen3-4b-fp8/config.yaml
new file mode 100644
index 000000000..6cf0cf1b6
--- /dev/null
+++ b/optimized/bisv2/qwen-qwen3-4b-fp8/config.yaml
@@ -0,0 +1,30 @@
+description: "Qwen3-4B optimized with BISv2 (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: BISV2-qwen-qwen3-4b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  inference_stack: v2
+  build:
+    checkpoint_repository:
+      repo: Qwen/Qwen3-4B
+      revision: main
+      source: HF
+    quantization_type: fp8_kv
+  runtime:
+    max_batch_size: 32
+    max_num_tokens: 32768
+    max_seq_len: 32768
diff --git a/optimized/briton/deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md b/optimized/briton/deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md
new file mode 100644
index 000000000..b913cb6e8
--- /dev/null
+++ b/optimized/briton/deepseek-ai-deepseek-r1-distill-llama-70b-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8-truss-example
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Llama-70B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Llama-70B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Llama-70B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Llama-70B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml b/optimized/briton/deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml
new file mode 100644
index 000000000..1d4a2d4a9
--- /dev/null
+++ b/optimized/briton/deepseek-ai-deepseek-r1-distill-llama-70b-fp8/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "DeepSeek-R1-Distill-Llama-70B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-deepseek-ai-deepseek-r1-distill-llama-70b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100:2
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Llama-70B
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 2
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md b/optimized/briton/deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md
new file mode 100644
index 000000000..ba0bcc28c
--- /dev/null
+++ b/optimized/briton/deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/README.md
@@ -0,0 +1,58 @@
+# Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8-truss-example
+
+Deploy [deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [deepseek-ai/DeepSeek-R1-Distill-Qwen-32B](https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="deepseek-ai/DeepSeek-R1-Distill-Qwen-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "deepseek-ai/DeepSeek-R1-Distill-Qwen-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml b/optimized/briton/deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml
new file mode 100644
index 000000000..97627c271
--- /dev/null
+++ b/optimized/briton/deepseek-ai-deepseek-r1-distill-qwen-32b-fp8/config.yaml
@@ -0,0 +1,35 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "DeepSeek-R1-Distill-Qwen-32B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-deepseek-ai-deepseek-r1-distill-qwen-32b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: deepseek-ai/DeepSeek-R1-Distill-Qwen-32B
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    num_builder_gpus: 4
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp8
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/google-gemma-3-1b-it/README.md b/optimized/briton/google-gemma-3-1b-it/README.md
new file mode 100644
index 000000000..4882f6ef0
--- /dev/null
+++ b/optimized/briton/google-gemma-3-1b-it/README.md
@@ -0,0 +1,59 @@
+# Briton-google-gemma-3-1b-it-truss-example
+
+Deploy [unsloth/gemma-3-1b-it](https://huggingface.co/unsloth/gemma-3-1b-it) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [unsloth/gemma-3-1b-it](https://huggingface.co/unsloth/gemma-3-1b-it) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="unsloth/gemma-3-1b-it",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "unsloth/gemma-3-1b-it", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Streaming: **enabled**
diff --git a/optimized/briton/google-gemma-3-1b-it/config.yaml b/optimized/briton/google-gemma-3-1b-it/config.yaml
new file mode 100644
index 000000000..2c35bd3f2
--- /dev/null
+++ b/optimized/briton/google-gemma-3-1b-it/config.yaml
@@ -0,0 +1,32 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "gemma-3-1b-it optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-google-gemma-3-1b-it-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: unsloth/gemma-3-1b-it
+      revision: main
+      source: HF
+    max_seq_len: 32768
+    quantization_type: no_quant
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/google-gemma-3-270m-it/README.md b/optimized/briton/google-gemma-3-270m-it/README.md
new file mode 100644
index 000000000..3d0f1c0ef
--- /dev/null
+++ b/optimized/briton/google-gemma-3-270m-it/README.md
@@ -0,0 +1,59 @@
+# Briton-google-gemma-3-270m-it-truss-example
+
+Deploy [google/gemma-3-270m-it](https://huggingface.co/google/gemma-3-270m-it) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [google/gemma-3-270m-it](https://huggingface.co/google/gemma-3-270m-it) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="google/gemma-3-270m-it",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "google/gemma-3-270m-it", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Streaming: **enabled**
diff --git a/optimized/briton/google-gemma-3-270m-it/config.yaml b/optimized/briton/google-gemma-3-270m-it/config.yaml
new file mode 100644
index 000000000..3f674b9b7
--- /dev/null
+++ b/optimized/briton/google-gemma-3-270m-it/config.yaml
@@ -0,0 +1,32 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "gemma-3-270m-it optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-google-gemma-3-270m-it-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: google/gemma-3-270m-it
+      revision: main
+      source: HF
+    max_seq_len: 32768
+    quantization_type: no_quant
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/google-gemma-3-27b-it-speculative-lookahead/README.md b/optimized/briton/google-gemma-3-27b-it-speculative-lookahead/README.md
new file mode 100644
index 000000000..667a4a56f
--- /dev/null
+++ b/optimized/briton/google-gemma-3-27b-it-speculative-lookahead/README.md
@@ -0,0 +1,60 @@
+# Briton-google-gemma-3-27b-it-speculative-lookahead-truss-example
+
+Deploy [baseten/gemma-3-27b-causallm-it](https://huggingface.co/baseten/gemma-3-27b-causallm-it) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/gemma-3-27b-causallm-it](https://huggingface.co/baseten/gemma-3-27b-causallm-it) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="baseten/gemma-3-27b-causallm-it",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "baseten/gemma-3-27b-causallm-it", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Streaming: **enabled**
diff --git a/optimized/briton/google-gemma-3-27b-it-speculative-lookahead/config.yaml b/optimized/briton/google-gemma-3-27b-it-speculative-lookahead/config.yaml
new file mode 100644
index 000000000..e33398c36
--- /dev/null
+++ b/optimized/briton/google-gemma-3-27b-it-speculative-lookahead/config.yaml
@@ -0,0 +1,41 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "gemma-3-27b-causallm-it optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-google-gemma-3-27b-it-speculative-lookahead-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: baseten/gemma-3-27b-causallm-it
+      revision: main
+      source: HF
+    max_batch_size: 64
+    max_num_tokens: 131072
+    max_seq_len: 131072
+    quantization_type: no_quant
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 8
+      lookahead_verification_set_size: 3
+      lookahead_windows_size: 3
+      num_draft_tokens: 41
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/google-gemma-3-27b-it/README.md b/optimized/briton/google-gemma-3-27b-it/README.md
new file mode 100644
index 000000000..7d92cb183
--- /dev/null
+++ b/optimized/briton/google-gemma-3-27b-it/README.md
@@ -0,0 +1,59 @@
+# Briton-google-gemma-3-27b-it-truss-example
+
+Deploy [baseten/gemma-3-27b-causallm-it](https://huggingface.co/baseten/gemma-3-27b-causallm-it) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/gemma-3-27b-causallm-it](https://huggingface.co/baseten/gemma-3-27b-causallm-it) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="baseten/gemma-3-27b-causallm-it",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "baseten/gemma-3-27b-causallm-it", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Streaming: **enabled**
diff --git a/optimized/briton/google-gemma-3-27b-it/config.yaml b/optimized/briton/google-gemma-3-27b-it/config.yaml
new file mode 100644
index 000000000..a315df80d
--- /dev/null
+++ b/optimized/briton/google-gemma-3-27b-it/config.yaml
@@ -0,0 +1,32 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "gemma-3-27b-causallm-it optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-google-gemma-3-27b-it-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: baseten/gemma-3-27b-causallm-it
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    quantization_type: no_quant
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.1-405b-fp8/README.md b/optimized/briton/meta-llama-llama-3.1-405b-fp8/README.md
new file mode 100644
index 000000000..05a97c571
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.1-405b-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-meta-llama-llama-3.1-405b-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.1-405B](https://huggingface.co/meta-llama/Llama-3.1-405B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-405B](https://huggingface.co/meta-llama/Llama-3.1-405B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:8 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.1-405B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-405B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **8** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.1-405b-fp8/config.yaml b/optimized/briton/meta-llama-llama-3.1-405b-fp8/config.yaml
new file mode 100644
index 000000000..62f208929
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.1-405b-fp8/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.1-405B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.1-405b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100:8
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.1-405B
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 8
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md b/optimized/briton/meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md
new file mode 100644
index 000000000..67b77b5c4
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.1-8B-Instruct](https://huggingface.co/meta-llama/Llama-3.1-8B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.1-8B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.1-8B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml b/optimized/briton/meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
new file mode 100644
index 000000000..574d56993
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
@@ -0,0 +1,43 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.1-8B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.1-8b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.1-8B-Instruct
+      revision: main
+      source: HF
+    max_batch_size: 64
+    max_num_tokens: 131072
+    max_seq_len: 131072
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 8
+      lookahead_verification_set_size: 3
+      lookahead_windows_size: 3
+      num_draft_tokens: 41
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.2-1b-instruct-fp8/README.md b/optimized/briton/meta-llama-llama-3.2-1b-instruct-fp8/README.md
new file mode 100644
index 000000000..ddcc09767
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-1b-instruct-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-meta-llama-llama-3.2-1b-instruct-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.2-1B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.2-1B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.2-1B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.2-1b-instruct-fp8/config.yaml b/optimized/briton/meta-llama-llama-3.2-1b-instruct-fp8/config.yaml
new file mode 100644
index 000000000..d6ae88c29
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-1b-instruct-fp8/config.yaml
@@ -0,0 +1,35 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.2-1B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.2-1b-instruct-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.2-1B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md b/optimized/briton/meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md
new file mode 100644
index 000000000..0bf9f6118
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.2-3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.2-3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml b/optimized/briton/meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml
new file mode 100644
index 000000000..e5c656fce
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8/config.yaml
@@ -0,0 +1,36 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.2-3B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.2-3b-instruct-calib-dataset-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.2-3B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_config:
+      calib_dataset: baseten/quant_calibration_dataset_v1
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.2-3b-instruct-fp8/README.md b/optimized/briton/meta-llama-llama-3.2-3b-instruct-fp8/README.md
new file mode 100644
index 000000000..88ebaa661
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-3b-instruct-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.2-3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.2-3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.2-3b-instruct-fp8/config.yaml b/optimized/briton/meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
new file mode 100644
index 000000000..8f2ae1078
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-3b-instruct-fp8/config.yaml
@@ -0,0 +1,34 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.2-3B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.2-3b-instruct-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.2-3B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.2-3b-instruct/README.md b/optimized/briton/meta-llama-llama-3.2-3b-instruct/README.md
new file mode 100644
index 000000000..c27af5c18
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-3b-instruct/README.md
@@ -0,0 +1,58 @@
+# Briton-meta-llama-llama-3.2-3b-instruct-truss-example
+
+Deploy [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.2-3B-Instruct](https://huggingface.co/meta-llama/Llama-3.2-3B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100_40GB |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.2-3B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.2-3B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.2-3b-instruct/config.yaml b/optimized/briton/meta-llama-llama-3.2-3b-instruct/config.yaml
new file mode 100644
index 000000000..da4b58628
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.2-3b-instruct/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.2-3B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.2-3b-instruct-truss-example
+python_version: py39
+resources:
+  accelerator: H100_40GB
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.2-3B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    quantization_type: no_quant
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp4/README.md b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp4/README.md
new file mode 100644
index 000000000..18eba6144
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp4/README.md
@@ -0,0 +1,58 @@
+# Briton-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
+
+Deploy [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.3-70B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.3-70B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4**
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp4/config.yaml b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
new file mode 100644
index 000000000..ca02362a8
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp4/config.yaml
@@ -0,0 +1,32 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.3-70B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.3-70b-instruct-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.3-70B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    num_builder_gpus: 4
+    quantization_type: fp4
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp8/README.md b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp8/README.md
new file mode 100644
index 000000000..4669284ab
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-meta-llama-llama-3.3-70b-instruct-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.3-70B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.3-70B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp8/config.yaml b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp8/config.yaml
new file mode 100644
index 000000000..108a5a224
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.3-70b-instruct-fp8/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.3-70B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.3-70b-instruct-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100:2
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.3-70B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 2
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md b/optimized/briton/meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md
new file mode 100644
index 000000000..cadd22884
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.3-70b-instruct-tp4-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8-truss-example
+
+Deploy [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [meta-llama/Llama-3.3-70B-Instruct](https://huggingface.co/meta-llama/Llama-3.3-70B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:4 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="meta-llama/Llama-3.3-70B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "meta-llama/Llama-3.3-70B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **4** GPUs
+- Max sequence length: **131,072**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml b/optimized/briton/meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml
new file mode 100644
index 000000000..6abfddd53
--- /dev/null
+++ b/optimized/briton/meta-llama-llama-3.3-70b-instruct-tp4-fp8/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Llama-3.3-70B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-meta-llama-llama-3.3-70b-instruct-tp4-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100:4
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: meta-llama/Llama-3.3-70B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 131072
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 4
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/microsoft-phi-4-fp8/README.md b/optimized/briton/microsoft-phi-4-fp8/README.md
new file mode 100644
index 000000000..c68d9a9f9
--- /dev/null
+++ b/optimized/briton/microsoft-phi-4-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-microsoft-phi-4-fp8-truss-example
+
+Deploy [unsloth/phi-4](https://huggingface.co/unsloth/phi-4) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [unsloth/phi-4](https://huggingface.co/unsloth/phi-4) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | L4:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="unsloth/phi-4",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "unsloth/phi-4", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **16,384**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/microsoft-phi-4-fp8/config.yaml b/optimized/briton/microsoft-phi-4-fp8/config.yaml
new file mode 100644
index 000000000..51f8bc64d
--- /dev/null
+++ b/optimized/briton/microsoft-phi-4-fp8/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "phi-4 optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-microsoft-phi-4-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: L4:2
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: unsloth/phi-4
+      revision: main
+      source: HF
+    max_seq_len: 16384
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 2
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/mistralai-mistral-7b-instruct-v0.3/README.md b/optimized/briton/mistralai-mistral-7b-instruct-v0.3/README.md
new file mode 100644
index 000000000..580d57d4f
--- /dev/null
+++ b/optimized/briton/mistralai-mistral-7b-instruct-v0.3/README.md
@@ -0,0 +1,59 @@
+# Briton-mistralai-mistral-7b-instruct-v0.3-truss-example
+
+Deploy [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mistral-7B-Instruct-v0.3](https://huggingface.co/mistralai/Mistral-7B-Instruct-v0.3) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | A10G:2 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mistral-7B-Instruct-v0.3",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mistral-7B-Instruct-v0.3", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/mistralai-mistral-7b-instruct-v0.3/config.yaml b/optimized/briton/mistralai-mistral-7b-instruct-v0.3/config.yaml
new file mode 100644
index 000000000..9827dd88e
--- /dev/null
+++ b/optimized/briton/mistralai-mistral-7b-instruct-v0.3/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Mistral-7B-Instruct-v0.3 optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-mistralai-mistral-7b-instruct-v0.3-truss-example
+python_version: py39
+resources:
+  accelerator: A10G:2
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: mistralai/Mistral-7B-Instruct-v0.3
+      revision: main
+      source: HF
+    max_seq_len: 32768
+    quantization_type: no_quant
+    tensor_parallel_count: 2
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/mistralai-mistral-small-24b-instruct-2501-fp8/README.md b/optimized/briton/mistralai-mistral-small-24b-instruct-2501-fp8/README.md
new file mode 100644
index 000000000..93a8cd2c6
--- /dev/null
+++ b/optimized/briton/mistralai-mistral-small-24b-instruct-2501-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-mistralai-mistral-small-24b-instruct-2501-fp8-truss-example
+
+Deploy [mistralai/Mistral-Small-24B-Instruct-2501](https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [mistralai/Mistral-Small-24B-Instruct-2501](https://huggingface.co/mistralai/Mistral-Small-24B-Instruct-2501) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="mistralai/Mistral-Small-24B-Instruct-2501",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "mistralai/Mistral-Small-24B-Instruct-2501", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml b/optimized/briton/mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml
new file mode 100644
index 000000000..3df60f735
--- /dev/null
+++ b/optimized/briton/mistralai-mistral-small-24b-instruct-2501-fp8/config.yaml
@@ -0,0 +1,34 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Mistral-Small-24B-Instruct-2501 optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-mistralai-mistral-small-24b-instruct-2501-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: mistralai/Mistral-Small-24B-Instruct-2501
+      revision: main
+      source: HF
+    max_seq_len: 32768
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen2.5-72b-instruct-tp2-fp8/README.md b/optimized/briton/qwen-qwen2.5-72b-instruct-tp2-fp8/README.md
new file mode 100644
index 000000000..48922e097
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-72b-instruct-tp2-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen2.5-72b-instruct-tp2-fp8-truss-example
+
+Deploy [Qwen/Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-72B-Instruct](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-72B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-72B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml b/optimized/briton/qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml
new file mode 100644
index 000000000..09c8b007d
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-72b-instruct-tp2-fp8/config.yaml
@@ -0,0 +1,34 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen2.5-72B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen2.5-72b-instruct-tp2-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100:2
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-72B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 32768
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp8
+    tensor_parallel_count: 2
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md b/optimized/briton/qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md
new file mode 100644
index 000000000..6d35b7298
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
+
+Deploy [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml b/optimized/briton/qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
new file mode 100644
index 000000000..5bee03956
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8/config.yaml
@@ -0,0 +1,44 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen2.5-7B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen2.5-7b-instruct-with-speculative-lookahead-decoding-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-7B-Instruct
+      revision: main
+      source: HF
+    max_batch_size: 64
+    max_num_tokens: 32768
+    max_seq_len: 32768
+    num_builder_gpus: 4
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp8
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 8
+      lookahead_verification_set_size: 3
+      lookahead_windows_size: 3
+      num_draft_tokens: 41
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md
new file mode 100644
index 000000000..22b2291a6
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/README.md
@@ -0,0 +1,58 @@
+# Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only-truss-example
+
+Deploy [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 MLP ONLY |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Coder-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Coder-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4_mlp_only**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml
new file mode 100644
index 000000000..961e00329
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only/config.yaml
@@ -0,0 +1,35 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen2.5-Coder-7B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen2.5-coder-7b-instruct-calib-dataset-fp4-mlp-only-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Coder-7B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 32768
+    num_builder_gpus: 4
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp4_mlp_only
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md
new file mode 100644
index 000000000..17b7b953b
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8-truss-example
+
+Deploy [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen2.5-Coder-7B-Instruct](https://huggingface.co/Qwen/Qwen2.5-Coder-7B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen2.5-Coder-7B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen2.5-Coder-7B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml
new file mode 100644
index 000000000..e3e3956d0
--- /dev/null
+++ b/optimized/briton/qwen-qwen2.5-coder-7b-instruct-min-latency-fp8/config.yaml
@@ -0,0 +1,44 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen2.5-Coder-7B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen2.5-coder-7b-instruct-min-latency-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen2.5-Coder-7B-Instruct
+      revision: main
+      source: HF
+    max_batch_size: 64
+    max_num_tokens: 32768
+    max_seq_len: 32768
+    num_builder_gpus: 4
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp8
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 32
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      num_draft_tokens: 61
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md b/optimized/briton/qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md
new file mode 100644
index 000000000..a391a012d
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-235b-a22b-instruct-2507-fp8/README.md
@@ -0,0 +1,62 @@
+# Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8-truss-example
+
+Deploy [Qwen/Qwen3-235B-A22B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-235B-A22B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-235B-A22B-Instruct-2507) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:8 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-235B-A22B-Instruct-2507",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-235B-A22B-Instruct-2507", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **8** GPUs
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **262,144**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml b/optimized/briton/qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml
new file mode 100644
index 000000000..314fd33ee
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-235b-a22b-instruct-2507-fp8/config.yaml
@@ -0,0 +1,43 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-235B-A22B-Instruct-2507 optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-235b-a22b-instruct-2507-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100:8
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-235B-A22B-Instruct-2507
+      revision: main
+      source: HF
+    max_batch_size: 64
+    max_num_tokens: 262144
+    max_seq_len: 262144
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 32
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      num_draft_tokens: 61
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 8
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-30b-a3b-fp8/README.md b/optimized/briton/qwen-qwen3-30b-a3b-fp8/README.md
new file mode 100644
index 000000000..161735e8a
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-30b-a3b-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen3-30b-a3b-fp8-truss-example
+
+Deploy [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-30B-A3B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-30B-A3B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-30b-a3b-fp8/config.yaml b/optimized/briton/qwen-qwen3-30b-a3b-fp8/config.yaml
new file mode 100644
index 000000000..c26c05360
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-30b-a3b-fp8/config.yaml
@@ -0,0 +1,34 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-30B-A3B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-30b-a3b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-30B-A3B
+      revision: main
+      source: HF
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md b/optimized/briton/qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md
new file mode 100644
index 000000000..acb19d843
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-30b-a3b-instruct-2507-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8-truss-example
+
+Deploy [Qwen/Qwen3-30B-A3B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-30B-A3B-Instruct-2507](https://huggingface.co/Qwen/Qwen3-30B-A3B-Instruct-2507) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-30B-A3B-Instruct-2507",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-30B-A3B-Instruct-2507", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **262,144**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml b/optimized/briton/qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml
new file mode 100644
index 000000000..6f94aa9da
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-30b-a3b-instruct-2507-fp8/config.yaml
@@ -0,0 +1,34 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-30B-A3B-Instruct-2507 optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-30b-a3b-instruct-2507-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-30B-A3B-Instruct-2507
+      revision: main
+      source: HF
+    max_seq_len: 262144
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-30b-a3b/README.md b/optimized/briton/qwen-qwen3-30b-a3b/README.md
new file mode 100644
index 000000000..3932c6a5d
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-30b-a3b/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen3-30b-a3b-truss-example
+
+Deploy [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-30B-A3B](https://huggingface.co/Qwen/Qwen3-30B-A3B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100:2 |
+| Quantization | NO QUANT |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-30B-A3B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-30B-A3B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **no_quant**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-30b-a3b/config.yaml b/optimized/briton/qwen-qwen3-30b-a3b/config.yaml
new file mode 100644
index 000000000..386807b37
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-30b-a3b/config.yaml
@@ -0,0 +1,31 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-30B-A3B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-30b-a3b-truss-example
+python_version: py39
+resources:
+  accelerator: H100:2
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-30B-A3B
+      revision: main
+      source: HF
+    max_seq_len: 40960
+    quantization_type: no_quant
+    tensor_parallel_count: 2
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-32b-fp4-mlp-only/README.md b/optimized/briton/qwen-qwen3-32b-fp4-mlp-only/README.md
new file mode 100644
index 000000000..9686ddbcd
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-32b-fp4-mlp-only/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen3-32b-fp4-mlp-only-truss-example
+
+Deploy [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 MLP ONLY |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4_mlp_only**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-32b-fp4-mlp-only/config.yaml b/optimized/briton/qwen-qwen3-32b-fp4-mlp-only/config.yaml
new file mode 100644
index 000000000..5161c5b02
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-32b-fp4-mlp-only/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-32B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-32b-fp4-mlp-only-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-32B
+      revision: main
+      source: HF
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    quantization_type: fp4_mlp_only
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-32b-fp4/README.md b/optimized/briton/qwen-qwen3-32b-fp4/README.md
new file mode 100644
index 000000000..55ab80362
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-32b-fp4/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwen3-32b-fp4-truss-example
+
+Deploy [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | B200 |
+| Quantization | FP4 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp4_kv**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-32b-fp4/config.yaml b/optimized/briton/qwen-qwen3-32b-fp4/config.yaml
new file mode 100644
index 000000000..f91afaaf4
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-32b-fp4/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-32B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-32b-fp4-truss-example
+python_version: py39
+resources:
+  accelerator: B200
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-32B
+      revision: main
+      source: HF
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    quantization_type: fp4_kv
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-32b-fp8/README.md b/optimized/briton/qwen-qwen3-32b-fp8/README.md
new file mode 100644
index 000000000..241c36276
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-32b-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-qwen-qwen3-32b-fp8-truss-example
+
+Deploy [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-32B](https://huggingface.co/Qwen/Qwen3-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Batch scheduler policy: **max_utilization**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-32b-fp8/config.yaml b/optimized/briton/qwen-qwen3-32b-fp8/config.yaml
new file mode 100644
index 000000000..9065d08b8
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-32b-fp8/config.yaml
@@ -0,0 +1,35 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-32B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-32b-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-32B
+      revision: main
+      source: HF
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 1
+  runtime:
+    batch_scheduler_policy: max_utilization
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwen3-8b-min-latency-fp8/README.md b/optimized/briton/qwen-qwen3-8b-min-latency-fp8/README.md
new file mode 100644
index 000000000..4a2f18da3
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-8b-min-latency-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-qwen-qwen3-8b-min-latency-fp8-truss-example
+
+Deploy [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/Qwen3-8B](https://huggingface.co/Qwen/Qwen3-8B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/Qwen3-8B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/Qwen3-8B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwen3-8b-min-latency-fp8/config.yaml b/optimized/briton/qwen-qwen3-8b-min-latency-fp8/config.yaml
new file mode 100644
index 000000000..2d33f24e9
--- /dev/null
+++ b/optimized/briton/qwen-qwen3-8b-min-latency-fp8/config.yaml
@@ -0,0 +1,43 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Qwen3-8B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwen3-8b-min-latency-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/Qwen3-8B
+      revision: main
+      source: HF
+    max_batch_size: 64
+    max_num_tokens: 40960
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 32
+      lookahead_verification_set_size: 1
+      lookahead_windows_size: 1
+      num_draft_tokens: 61
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwq-32b-reasoning-fp8/README.md b/optimized/briton/qwen-qwq-32b-reasoning-fp8/README.md
new file mode 100644
index 000000000..7c22b16ef
--- /dev/null
+++ b/optimized/briton/qwen-qwq-32b-reasoning-fp8/README.md
@@ -0,0 +1,58 @@
+# Briton-qwen-qwq-32b-reasoning-fp8-truss-example
+
+Deploy [Qwen/QwQ-32B](https://huggingface.co/Qwen/QwQ-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/QwQ-32B](https://huggingface.co/Qwen/QwQ-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/QwQ-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/QwQ-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwq-32b-reasoning-fp8/config.yaml b/optimized/briton/qwen-qwq-32b-reasoning-fp8/config.yaml
new file mode 100644
index 000000000..6c4eeac74
--- /dev/null
+++ b/optimized/briton/qwen-qwq-32b-reasoning-fp8/config.yaml
@@ -0,0 +1,35 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "QwQ-32B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwq-32b-reasoning-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/QwQ-32B
+      revision: main
+      source: HF
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp8
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/qwen-qwq-32b-reasoning-with-speculative-fp8/README.md b/optimized/briton/qwen-qwq-32b-reasoning-with-speculative-fp8/README.md
new file mode 100644
index 000000000..779fb7e24
--- /dev/null
+++ b/optimized/briton/qwen-qwq-32b-reasoning-with-speculative-fp8/README.md
@@ -0,0 +1,59 @@
+# Briton-qwen-qwq-32b-reasoning-with-speculative-fp8-truss-example
+
+Deploy [Qwen/QwQ-32B](https://huggingface.co/Qwen/QwQ-32B) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [Qwen/QwQ-32B](https://huggingface.co/Qwen/QwQ-32B) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | H100 |
+| Quantization | FP8 |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="Qwen/QwQ-32B",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "Qwen/QwQ-32B", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8**
+- Speculative decoding: **LOOKAHEAD_DECODING**
+- Max sequence length: **40,960**
+- Chunked context: **enabled**
+- Streaming: **enabled**
diff --git a/optimized/briton/qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml b/optimized/briton/qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml
new file mode 100644
index 000000000..f7fa10268
--- /dev/null
+++ b/optimized/briton/qwen-qwq-32b-reasoning-with-speculative-fp8/config.yaml
@@ -0,0 +1,44 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "QwQ-32B optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-qwen-qwq-32b-reasoning-with-speculative-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: Qwen/QwQ-32B
+      revision: main
+      source: HF
+    max_batch_size: 64
+    max_num_tokens: 40960
+    max_seq_len: 40960
+    num_builder_gpus: 4
+    quantization_config:
+      calib_max_seq_length: 2048
+      calib_size: 2048
+    quantization_type: fp8
+    speculator:
+      enable_b10_lookahead: true
+      lookahead_ngram_size: 8
+      lookahead_verification_set_size: 3
+      lookahead_windows_size: 3
+      num_draft_tokens: 41
+      speculative_decoding_mode: LOOKAHEAD_DECODING
+    tensor_parallel_count: 1
+  runtime:
+    enable_chunked_context: true
diff --git a/optimized/briton/tiiuae-falcon3-10b-instruct-fp8/README.md b/optimized/briton/tiiuae-falcon3-10b-instruct-fp8/README.md
new file mode 100644
index 000000000..6441a13d4
--- /dev/null
+++ b/optimized/briton/tiiuae-falcon3-10b-instruct-fp8/README.md
@@ -0,0 +1,60 @@
+# Briton-tiiuae-falcon3-10b-instruct-fp8-truss-example
+
+Deploy [tiiuae/Falcon3-10B-Instruct](https://huggingface.co/tiiuae/Falcon3-10B-Instruct) for text generation using a TRT-LLM engine on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [tiiuae/Falcon3-10B-Instruct](https://huggingface.co/tiiuae/Falcon3-10B-Instruct) |
+| Task | Text generation |
+| Engine | TRT-LLM |
+| GPU | L4:2 |
+| Quantization | FP8 KV |
+| OpenAI compatible | Yes |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+This model is OpenAI-compatible. You can use the OpenAI Python client or curl.
+
+**Python (OpenAI SDK):**
+
+```python
+from openai import OpenAI
+
+client = OpenAI(
+    api_key="YOUR_BASETEN_API_KEY",
+    base_url="https://model-<model_id>.api.baseten.co/v1",
+)
+
+response = client.chat.completions.create(
+    model="tiiuae/Falcon3-10B-Instruct",
+    messages=[{"role": "user", "content": "What is machine learning?"}],
+    max_tokens=512,
+)
+
+print(response.choices[0].message.content)
+```
+
+**curl:**
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/v1/chat/completions \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"model": "tiiuae/Falcon3-10B-Instruct", "messages": [{"role": "user", "content": "What is machine learning?"}], "max_tokens": 512}'
+```
+
+## Configuration highlights
+
+- Quantization: **fp8_kv**
+- Tensor parallelism: **2** GPUs
+- Max sequence length: **32,768**
+- Chunked context: **enabled**
+- Plugin: **use_fp8_context_fmha**
+- Streaming: **enabled**
diff --git a/optimized/briton/tiiuae-falcon3-10b-instruct-fp8/config.yaml b/optimized/briton/tiiuae-falcon3-10b-instruct-fp8/config.yaml
new file mode 100644
index 000000000..23a260c14
--- /dev/null
+++ b/optimized/briton/tiiuae-falcon3-10b-instruct-fp8/config.yaml
@@ -0,0 +1,33 @@
+# this file was autogenerated by `generate_templates.py` - please do change via template only
+description: "Falcon3-10B-Instruct optimized with Briton (TRT-LLM)"
+model_metadata:
+  example_model_input:
+    max_tokens: 512
+    messages:
+    - content: Tell me everything you know about optimized inference.
+      role: user
+    stream: true
+    temperature: 0.5
+  tags:
+  - openai-compatible
+model_name: Briton-tiiuae-falcon3-10b-instruct-fp8-truss-example
+python_version: py39
+resources:
+  accelerator: L4:2
+  cpu: '1'
+  memory: 10Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: decoder
+    checkpoint_repository:
+      repo: tiiuae/Falcon3-10B-Instruct
+      revision: main
+      source: HF
+    max_seq_len: 32768
+    plugin_configuration:
+      use_fp8_context_fmha: true
+    quantization_type: fp8_kv
+    tensor_parallel_count: 2
+  runtime:
+    enable_chunked_context: true
diff --git a/orpheus-3b-websockets/config.yaml b/orpheus-3b-websockets/config.yaml
deleted file mode 100644
index eae3cae0b..000000000
--- a/orpheus-3b-websockets/config.yaml
+++ /dev/null
@@ -1,59 +0,0 @@
-build_commands:
-  - apt-get update && apt-get install git git-lfs -y
-  - git lfs install
-  - git clone https://huggingface.co/hubertsiuzdak/snac_24khz /app/snac_24khz
-environment_variables:
-  ENABLE_EXECUTOR_API: "1"
-model_metadata:
-  example_model_input:
-    max_tokens: 10000
-    prompt:
-      "In todays fast-paced world, finding balance between work and personal
-      life is more important than ever. With the constant demands of technology, remote
-      communication, "
-    voice: tara
-  tags:
-    - force-legacy-api-non-openai-compatible
-model_name: Orpheus-3b Websockets
-python_version: py39
-requirements:
-  - snac==1.2.1
-  - torch==2.7.0
-  - batched==0.1.4
-  - httpx
-  - websockets
-  - pysbd
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
-runtime:
-  is_websocket_endpoint: true
-  transport:
-    kind: websocket
-    ping_interval_seconds: null
-    ping_timeout_seconds: null
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: baseten/orpheus-3b-0.1-ft
-      revision: b9eb57a06083cb9e5a083885fad991aa79c0bd24
-      source: HF
-    max_batch_size: 256
-    # set higher, so we can always use the max batch size in a single iter.
-    max_num_tokens: 16384
-    # 65536 would be around 600s of audio, typically model produces max 120s.
-    max_seq_len: 65536
-    num_builder_gpus: 1
-    quantization_config:
-      # TODO: Generate a typical dataset (input + output tokens) in target language
-      # or disable quantization for other languages
-      calib_dataset: "cnn_dailymail"
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
diff --git a/orpheus-best-performance/README.md b/orpheus-best-performance/README.md
deleted file mode 100644
index 3feae8dff..000000000
--- a/orpheus-best-performance/README.md
+++ /dev/null
@@ -1,8 +0,0 @@
-# Orpheus TTS
-
-To deploy this performance-optimized implementation of Orpheus TTS, first ensure everything is correct in config.yaml then run:
-
-```
-pip install --upgrade truss
-truss push --publish orpheus-best-performance
-```
diff --git a/orpheus-best-performance/config.yaml b/orpheus-best-performance/config.yaml
deleted file mode 100644
index 322d77e10..000000000
--- a/orpheus-best-performance/config.yaml
+++ /dev/null
@@ -1,58 +0,0 @@
-build_commands:
-  - apt-get update && apt-get install git git-lfs -y
-  - git lfs install
-  - git clone https://huggingface.co/hubertsiuzdak/snac_24khz /app/snac_24khz
-environment_variables:
-  ENABLE_EXECUTOR_API: "1"
-model_metadata:
-  repo_id: canopylabs/orpheus-3b-0.1-ft
-  example_model_input:
-    max_tokens: 10000
-    prompt:
-      "In todays fast-paced world, finding balance between work and personal
-      life is more important than ever. With the constant demands of technology, remote
-      communication, "
-    voice: tara
-  tags:
-    - force-legacy-api-non-openai-compatible
-model_name: Orpheus-3b Best Performance
-python_version: py39
-requirements:
-  - --extra-index-url https://download.pytorch.org/whl/cu128
-  - torch==2.7.1
-  - snac==1.2.1
-  - batched==0.1.4
-resources:
-  # NOTE: Model is bottlenecked by CPU clock speed
-  # H100 upgrade is not really effective
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: baseten/orpheus-3b-0.1-ft
-      revision: b9eb57a06083cb9e5a083885fad991aa79c0bd24
-      source: HF
-    max_batch_size: 256
-    # set higher, so we can always use the max batch size in a single iter.
-    max_num_tokens: 16384
-    # 32768 would be around 300s of audio, typically model produces max 120s.
-    max_seq_len: 32768
-    num_builder_gpus: 1
-    quantization_config:
-      # TODO: Generate a typical dataset (input + output tokens) in target language
-      # or disable quantization for other languages
-      calib_dataset: "cnn_dailymail"
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
-    kv_cache_free_gpu_mem_fraction: 0.90
-    batch_scheduler_policy: max_utilization
diff --git a/paddlepaddle/PaddleOCR_VL/config.yaml b/paddlepaddle/PaddleOCR_VL/config.yaml
deleted file mode 100644
index 37a105a2c..000000000
--- a/paddlepaddle/PaddleOCR_VL/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-base_image:
-  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:0bf29fadf5f8b28817fbccb037fb70adaef3f7f1
-# build_commands:
-#   - pip uninstall -y vllm
-#   - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git
-model_metadata:
-  repo_id: PaddlePaddle/PaddleOCR-VL
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: user
-        content:
-          - type: image_url
-            image_url:
-              url: "https://ofasys-multimodal-wlcb-3-toshanghai.oss-accelerate.aliyuncs.com/wpf272043/keepme/image/receipt.png"
-          - type: text
-            text: "OCR:"
-    model: "PaddlePaddle/PaddleOCR-VL"
-    max_tokens: 4096
-    temperature: 0.0
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: vllm serve PaddlePaddle/PaddleOCR-VL --trust-remote-code --max-num-batched-tokens 16384 --no-enable-prefix-caching --mm-processor-cache-gb 0 --tensor-parallel-size 1 --served-model-name PaddlePaddle/PaddleOCR-VL --host 0.0.0.0 --port 8000
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-model_name: PaddleOCR-VL
diff --git a/personaplex-7b-v1/README.md b/personaplex-7b-v1/README.md
deleted file mode 100644
index dafe3b0b2..000000000
--- a/personaplex-7b-v1/README.md
+++ /dev/null
@@ -1,117 +0,0 @@
-# PersonaPlex 7V - Real-time Speech-to-Speech
-
-This example demonstrates deploying NVIDIA's PersonaPlex 7B v1 model using Baseten's Bring Your Own Image (BYOI) feature with a custom Docker image containing a forked version of PersonaPlex.
-
-## Overview
-
-PersonaPlex is a speech-to-speech AI model that enables real-time voice conversations. This deployment:
-
-1. **Custom Docker Image**: We built a Docker image (`basetenservice/personaplex-7v:fork`) that builds a fork of PersonaPlex from [basetenlabs/personaplex-baseten](https://github.com/basetenlabs/personaplex-baseten)
-2. **Baseten BYOI**: The Docker image is deployed using Baseten's bring your own image capability, which allows us to use custom base images with the Baseten platform
-3. **WebSocket Protocol**: The model communicates via WebSocket for low-latency bidirectional audio streaming
-
-The dockerfile is included in this folder. The `config.yaml` brings this image and runs the Moshi server on port 8998.
-
-## Server Setup
-
-To deploy PersonaPlex 7V on Baseten, you'll use the Truss CLI to push this folder
-
-### Requirements
-
-- Truss CLI installed (`pip install truss`), cd into this directory, and `truss push --publish`
-- Baseten account and API key
-- Hugging Face account and Read Access Token
-```
-
-### Deploy the Model
-
-From the `personaplex-7b-v1` directory, push your Truss:
-
-```bash
-truss push . --byoi --env HUGGINGFACE_TOKEN=$HUGGINGFACE_TOKEN
-```
-
-This command builds your custom Docker image, passes the Hugging Face token as a runtime environment variable, and deploys your model to Baseten.
-
-After deployment, your WebSocket server will be available and ready to accept client connections.
-
-## Client Setup
-
-### Prerequisites
-
-The client to connect to the server requires the Opus audio codec and Python dependencies.
-
-#### macOS
-
-```bash
-brew install opus
-```
-
-#### Linux (Ubuntu/Debian)
-
-```bash
-sudo apt-get install libopus-dev
-```
-
-### Install Python Dependencies
-
-```bash
-pip install -r requirements-client.txt
-```
-
-This installs:
-- `websockets` - WebSocket client library
-- `numpy` - Array operations
-- `sphn` - Opus audio codec Python bindings
-- `sounddevice` - Audio I/O
-
-### Set API Key
-
-```bash
-export BASETEN_API_KEY="your_api_key_here"
-```
-
-## Running the Client
-
-```bash
-python client.py
-```
-
-The client will:
-1. Connect to the deployed model via WebSocket
-2. Start your microphone and speakers
-3. Enable real-time voice conversation with the AI
-
-## WebSocket Protocol
-
-The communication protocol uses a simple binary format:
-
-### Initial Handshake
-
-1. **Client sends config** (first message): JSON configuration
-   ```json
-   {
-     "voice_prompt": "NATF0.pt",
-     "text_prompt": "You are a helpful assistant.",
-     "seed": -1
-   }
-   ```
-
-2. **Server responds**: Single byte `0x00` to confirm readiness
-
-### Streaming Messages
-
-After the handshake, all messages are binary with a single-byte header:
-
-- **`0x01` + Opus audio data**: Audio frames (sent by both client and server)
-- **`0x02` + UTF-8 text**: Text transcriptions/responses (sent by server)
-
-The first byte of each message indicates the payload type, and the remaining bytes contain the Opus-encoded audio or UTF-8 text.
-
-## Configuration
-
-The model is configured in `config.yaml`:
-- **Image**: Contains the built model at nvidia/personaplex-7b-v1
-- **Accelerator**: H100 GPU
-- **Transport**: WebSocket
-- **Secrets**: Requires HuggingFace access token for model download
diff --git a/personaplex-7b-v1/config.yaml b/personaplex-7b-v1/config.yaml
deleted file mode 100644
index b4b778edf..000000000
--- a/personaplex-7b-v1/config.yaml
+++ /dev/null
@@ -1,22 +0,0 @@
-description: Speech-to-speech model powered by NVIDIA Personaplex
-base_image:
-  image: basetenservice/personaplex-7v:fork
-model_metadata:
-  repo_id: nvidia/personaplex-7b-v1
-  tags:
-    - speech-to-speech
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) python3 -m moshi.server --host 0.0.0.0 --port 8998"
-  readiness_endpoint: /
-  liveness_endpoint: /
-  predict_endpoint: /api/chat
-  server_port: 8998
-resources:
-  accelerator: H100
-  use_gpu: true
-model_name: Personaplex 7V H100
-secrets:
-  hf_access_token: null
-runtime:
-  transport:
-    kind: websocket
diff --git a/phi/phi-3-mini-128k-instruct/README.md b/phi/phi-3-mini-128k-instruct/README.md
deleted file mode 100644
index a8cf1d6b8..000000000
--- a/phi/phi-3-mini-128k-instruct/README.md
+++ /dev/null
@@ -1,57 +0,0 @@
-# Phi-3-Mini-128K-Instruct
-
-This is a [Truss](https://truss.baseten.co/) for Phi-3-Mini-128K-Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Llama 3 8B.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd phi/phi-3-mini-128k-instruct
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `phi/phi-3-mini-128k-instruct` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Phi-3-Mini-128K-Instruct API documentation
-
-This section provides an overview of the Phi-3-Mini-128K-Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __messages__: The input text that you want the model to generate a response for.
-- __max_tokens__ (optional, default=512): The maximum number of tokens to return, counting input tokens. Maximum of 4096.
-- __temperature__ (optional, default=1.0): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results.
-- __top_p__ (optional, default=0.75): The cumulative probability threshold for token sampling. The model will only consider tokens whose cumulative probability is below this threshold.
-- __top_k__ (optional, default=40): The number of top tokens to consider when sampling. The model will only consider the top_k highest-probability tokens.
-
-The API also supports passing any parameter supported by HuggingFace's `Transformers.generate`.
-
-## Example usage
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "messages": [{"role": "user", "content": "What even is AGI?"}],
-           "max_tokens": 256
-         }'
-```
diff --git a/phi/phi-3-mini-128k-instruct/config.yaml b/phi/phi-3-mini-128k-instruct/config.yaml
deleted file mode 100644
index db4324c6b..000000000
--- a/phi/phi-3-mini-128k-instruct/config.yaml
+++ /dev/null
@@ -1,15 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata: {}
-model_name: Phi-3-Mini-128K-Instruct
-python_version: py39
-requirements:
-  - accelerate
-  - einops
-  - transformers==4.40.1
-  - torch==2.3.0
-resources:
-  accelerator: T4
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/phi/phi-3-mini-4k-instruct/README.md b/phi/phi-3-mini-4k-instruct/README.md
deleted file mode 100644
index 19d4ddda8..000000000
--- a/phi/phi-3-mini-4k-instruct/README.md
+++ /dev/null
@@ -1,57 +0,0 @@
-# Phi-3-Mini-4K-Instruct
-
-This is a [Truss](https://truss.baseten.co/) for Phi-3-Mini-4K-Instruct. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Llama 3 8B.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd phi/phi-3-mini-4k-instruct
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `phi/phi-3-mini-4k-instruct` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Phi-3-Mini-4K-Instruct API documentation
-
-This section provides an overview of the Phi-3-Mini-4K-Instruct API, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __messages__: The input text that you want the model to generate a response for.
-- __max_tokens__ (optional, default=512): The maximum number of tokens to return, counting input tokens. Maximum of 4096.
-- __temperature__ (optional, default=1.0): Controls the randomness of the generated text. Higher values produce more diverse results, while lower values produce more deterministic results.
-- __top_p__ (optional, default=0.75): The cumulative probability threshold for token sampling. The model will only consider tokens whose cumulative probability is below this threshold.
-- __top_k__ (optional, default=40): The number of top tokens to consider when sampling. The model will only consider the top_k highest-probability tokens.
-
-The API also supports passing any parameter supported by HuggingFace's `Transformers.generate`.
-
-## Example usage
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://app.baseten.co/model_versions/YOUR_MODEL_VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "messages": [{"role": "user", "content": "What even is AGI?"}],
-           "max_tokens": 256
-         }'
-```
diff --git a/phi/phi-3-mini-4k-instruct/config.yaml b/phi/phi-3-mini-4k-instruct/config.yaml
deleted file mode 100644
index 20f2c30e8..000000000
--- a/phi/phi-3-mini-4k-instruct/config.yaml
+++ /dev/null
@@ -1,15 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata: {}
-model_name: Phi-3-Mini-4K-Instruct
-python_version: py39
-requirements:
-  - accelerate
-  - einops
-  - transformers==4.40.1
-  - torch==2.3.0
-resources:
-  accelerator: T4
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/phi/phi-3.5-mini/README.md b/phi/phi-3.5-mini/README.md
deleted file mode 100644
index 87be85f77..000000000
--- a/phi/phi-3.5-mini/README.md
+++ /dev/null
@@ -1,98 +0,0 @@
-# Phi 3.5 Mini Instruct
-
-This is a [Truss](https://truss.baseten.co/) example using our general purpose [vLLM Template](https://github.com/basetenlabs/truss-examples/tree/main/vllm) but for Phi-3.5-Mini-instruct, one of the [compatible chat completion models](https://docs.vllm.ai/en/latest/models/supported_models.html).
-
-> Note: `prefix_caching` is not supported by vLLM for this model, please do not include `enable_prefix_caching` as part of the vllm config in `config.yaml`
-
-
-
-## Deploy your Truss
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. With `vllm` as your working directory, you can deploy the model with:
-
-    ```sh
-    truss push --trusted
-    ```
-
-    Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Call your model
-
-Once your deployment is up, there are [many ways](https://docs.baseten.co/invoke/quickstart) to call your model.
-
-### curl command
-
-#### If you are NOT using OpenAI compatible server
-
-```
-curl -X POST https://model-<YOUR_MODEL_ID>.api.baseten.co/development/predict \
-     -H "Authorization: Api-Key $BASETEN_API_KEY" \
-     -d '{"prompt": "what is the meaning of life"}'
-```
-
-
-#### If you are using OpenAI compatible server
-
-```
-curl -X POST "https://model-<YOUR_MODEL_ID>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {BASETEN_API_KEY}' \
-     -d '{
-           "messages": [{"role": "user", "content": "What even is AGI?"}],
-           "max_tokens": 256
-         }'
-```
-
-To access [production metrics](https://docs.vllm.ai/en/latest/serving/metrics.html) reported by OpenAI compatible server, simply add `metrics: true` to the request.
-
-```
-curl -X POST "https://model-<YOUR_MODEL_ID>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {BASETEN_API_KEY}' \
-     -d '{
-           "metrics": true
-         }'
-```
-
-### OpenAI SDK (if you are using OpenAI compatible server)
-
-```
-from openai import OpenAI
-import os
-
-model_id = "abcd1234" # Replace with your model ID
-deployment_id = "4321cbda" # [Optional] Replace with your deployment ID
-
-client = OpenAI(
-    api_key=os.environ["BASETEN_API_KEY"],
-    base_url=f"https://bridge.baseten.co/v1/direct"
-)
-
-response = client.chat.completions.create(
-  model="microsoft/Phi-3.5-mini-instruct",
-  messages=[
-    {"role": "user", "content": "Who won the world series in 2020?"},
-    {"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
-    {"role": "user", "content": "Where was it played?"}
-  ],
-  extra_body={
-    "baseten": {
-      "model_id": model_id,
-      "deployment_id": deployment_id
-    }
-  }
-)
-
-print(response.choices[0].message.content)
-
-```
-
-For more information, see [API reference](https://docs.baseten.co/api-reference/openai).
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/phi/phi-3.5-mini/config.yaml b/phi/phi-3.5-mini/config.yaml
deleted file mode 100644
index 9beacb1f1..000000000
--- a/phi/phi-3.5-mini/config.yaml
+++ /dev/null
@@ -1,18 +0,0 @@
-model_name: "Phi 3.5 Mini Instruct VLLM openai compatible"
-python_version: py311
-model_metadata:
-  example_model_input: {"messages": [{"role": "user", "content": "what is the meaning of life"}]}
-  repo_id: microsoft/Phi-3.5-mini-instruct
-  openai_compatible: true
-  vllm_config:
-    tensor_parallel_size: 1
-    max_model_len: 10000
-requirements:
-  - vllm==0.5.4
-resources:
-  accelerator: A10G
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
diff --git a/piper-tts/README.md b/piper-tts/README.md
deleted file mode 100644
index 7bb375fce..000000000
--- a/piper-tts/README.md
+++ /dev/null
@@ -1,65 +0,0 @@
-# Piper TTS Truss
-
-This repository packages [Piper-TTS](https://github.com/rhasspy/piper) as a [Truss](https://truss.baseten.co/).
-
-Piper TTS is a generative audio model for text-to-speech generation. This model can be trained on various voices to create realistic output audio. This model has very low latency and is optimized to run on a Raspberry Pi!
-
-## Deploying Piper TTS
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd piper-tts
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `piper-tts` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-
-## Hardware Requirements
-This model does not need a GPU to run. However, the configuration provided with this Truss does have a GPU enabled for extra performance.
-
-## Invoking the model
-
-Here are the following inputs for the model:
-1. `text`: The text that needs to be converted into speech
-
-Here is an example of how to invoke this model:
-
-```python
-import base64
-import requests
-import os
-
-def base64_to_wav(base64_string, output_file_path):
-  binary_data = base64.b64decode(base64_string)
-  with open(output_file_path, "wb") as wav_file:
-    wav_file.write(binary_data)
-
-text = "Listen up, people. Life's a wild ride, and sometimes you gotta grab it by the horns and steer it where you want to go. You can't just sit around waiting for things to happen – you gotta make 'em happen. Yeah, it's gonna get tough, but that's when you dig deep, find that inner badass, and come out swinging. Remember, success ain't handed to you on a silver platter; you gotta snatch it like it owes you money. So, lace up your boots, square those shoulders, and let the world know that you're here to play, and you're playing for keeps"
-data = {"text": text}
-headers = {"Authorization": f"Api-Key <BASETEN-API-KEY>"}
-res = requests.post("https://model-<model-id>.api.baseten.co/development/predict", headers=headers, json=data)
-res = res.json()
-output = base64_to_wav(res.get('output'), "piper-tts-output.wav")
-os.system("open piper-tts-output.wav")
-```
-
-The output of the model is a base64 string, so you can convert it to a wav file using the `base64_to_wav` function.
-
-Here is the output from the model using the input above:
-
-https://github.com/htrivedi99/truss-examples/assets/15642666/d35a9da3-f49c-4820-ab1e-08982d893598
diff --git a/piper-tts/config.yaml b/piper-tts/config.yaml
deleted file mode 100644
index 63b890b52..000000000
--- a/piper-tts/config.yaml
+++ /dev/null
@@ -1,23 +0,0 @@
-environment_variables: {}
-external_data:
-- local_data_path: models/model.onnx
-  url: https://huggingface.co/rhasspy/piper-voices/resolve/v1.0.0/en/en_US/lessac/medium/en_US-lessac-medium.onnx
-- local_data_path: models/model.onnx.json
-  url: https://huggingface.co/rhasspy/piper-voices/raw/v1.0.0/en/en_US/lessac/medium/en_US-lessac-medium.onnx.json
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    text: I love robots. Robots are cool!
-  tags:
-  - text-to-speech
-model_name: Piper TTS
-python_version: py310
-requirements:
-- piper-tts==1.2.0
-resources:
-  accelerator: T4
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/playground-v2-aesthetic/README.md b/playground-v2-aesthetic/README.md
deleted file mode 100644
index 15327a4e8..000000000
--- a/playground-v2-aesthetic/README.md
+++ /dev/null
@@ -1,97 +0,0 @@
-# Playground V2 Aesthetic Truss
-
-This is a [Truss](https://truss.baseten.co/welcome) for [Playground V2 Aesthetic](https://huggingface.co/playgroundai/playground-v2-1024px-aesthetic)
-
-Playground v2 is a diffusion-based text-to-image generative model. The model was trained from scratch by the research team at [Playground](https://playground.com). The research team curated 3,000 samples from Midjourney per category to train their model. According to their human preference study, people prefers the output of this model 2.5X more than Stable Diffusion XL.
-
-Here are some examples of the outputs from this model:
-
-![aesthetic_images](https://github.com/htrivedi99/truss-examples/assets/15642666/6c3fc815-73ec-4ab1-abab-7723884791f1)
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd playground-v2-aesthetic
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `playground-v2-aesthetic` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### API route: `predict`
-
-The predict route is the primary method for generating images based on a given prompt. It takes several parameters:
-
-- `prompt` (required): The input text required for image generation.
-- `negative_prompt` (optional): Use this to refine the image generation by discarding unwanted items.
-- `scheduler` (optional): This controls which scheduler to use resulting in more image variations.
-- `steps` (optional): The number of iterations the model runs through.
-- `guidance_scale` (optional): Used to control how closely the image generation follows the prompt.
-- `seed` (optional): Random number used to control image variations.
-
-The output of the model is an image in the form of a base64 string.
-Example model output: `{"output": "BASE64-STRING"}`
-
-## Example usage
-
-```sh
-truss predict -d '{"prompt": "An astronaut snowboarding on an alient planet, highly detailed, 8K"}'
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST " https://model-<model-id>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "An astronaut snowboarding on an alient planet, highly detailed, 8K"
-         }'
-```
-
-You can also use Python to invoke the model:
-
-``` python
-import os
-import json
-import base64
-import requests
-
-# Set essential values
-model_id = ""
-baseten_api_key = ""
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/development/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json={
-        "prompt": "a futuristic motorcycle, neon colors, cyberpunk city, detailed, 8K",
-        "steps": 50
-    }
-)
-# Get output image
-res = res.json()
-output = json.loads(res)["output"]
-image = base64.b64decode(output)
-# Save image to file
-img_file = open("playground.png", "wb")
-img_file.write(image)
-img_file.close()
-os.system("open playground.png")
-```
diff --git a/poetry.lock b/poetry.lock
deleted file mode 100644
index 4d3793a89..000000000
--- a/poetry.lock
+++ /dev/null
@@ -1,609 +0,0 @@
-# This file is automatically @generated by Poetry 1.8.4 and should not be changed by hand.
-
-[[package]]
-name = "asttokens"
-version = "3.0.0"
-description = "Annotate AST trees with source code positions"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "asttokens-3.0.0-py3-none-any.whl", hash = "sha256:e3078351a059199dd5138cb1c706e6430c05eff2ff136af5eb4790f9d28932e2"},
-    {file = "asttokens-3.0.0.tar.gz", hash = "sha256:0dcd8baa8d62b0c1d118b399b2ddba3c4aff271d0d7a9e0d4c1681c79035bbc7"},
-]
-
-[package.extras]
-astroid = ["astroid (>=2,<4)"]
-test = ["astroid (>=2,<4)", "pytest", "pytest-cov", "pytest-xdist"]
-
-[[package]]
-name = "black"
-version = "23.12.1"
-description = "The uncompromising code formatter."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "black-23.12.1-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:e0aaf6041986767a5e0ce663c7a2f0e9eaf21e6ff87a5f95cbf3675bfd4c41d2"},
-    {file = "black-23.12.1-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:c88b3711d12905b74206227109272673edce0cb29f27e1385f33b0163c414bba"},
-    {file = "black-23.12.1-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:a920b569dc6b3472513ba6ddea21f440d4b4c699494d2e972a1753cdc25df7b0"},
-    {file = "black-23.12.1-cp310-cp310-win_amd64.whl", hash = "sha256:3fa4be75ef2a6b96ea8d92b1587dd8cb3a35c7e3d51f0738ced0781c3aa3a5a3"},
-    {file = "black-23.12.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:8d4df77958a622f9b5a4c96edb4b8c0034f8434032ab11077ec6c56ae9f384ba"},
-    {file = "black-23.12.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:602cfb1196dc692424c70b6507593a2b29aac0547c1be9a1d1365f0d964c353b"},
-    {file = "black-23.12.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9c4352800f14be5b4864016882cdba10755bd50805c95f728011bcb47a4afd59"},
-    {file = "black-23.12.1-cp311-cp311-win_amd64.whl", hash = "sha256:0808494f2b2df923ffc5723ed3c7b096bd76341f6213989759287611e9837d50"},
-    {file = "black-23.12.1-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:25e57fd232a6d6ff3f4478a6fd0580838e47c93c83eaf1ccc92d4faf27112c4e"},
-    {file = "black-23.12.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:2d9e13db441c509a3763a7a3d9a49ccc1b4e974a47be4e08ade2a228876500ec"},
-    {file = "black-23.12.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6d1bd9c210f8b109b1762ec9fd36592fdd528485aadb3f5849b2740ef17e674e"},
-    {file = "black-23.12.1-cp312-cp312-win_amd64.whl", hash = "sha256:ae76c22bde5cbb6bfd211ec343ded2163bba7883c7bc77f6b756a1049436fbb9"},
-    {file = "black-23.12.1-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:1fa88a0f74e50e4487477bc0bb900c6781dbddfdfa32691e780bf854c3b4a47f"},
-    {file = "black-23.12.1-cp38-cp38-macosx_11_0_arm64.whl", hash = "sha256:a4d6a9668e45ad99d2f8ec70d5c8c04ef4f32f648ef39048d010b0689832ec6d"},
-    {file = "black-23.12.1-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:b18fb2ae6c4bb63eebe5be6bd869ba2f14fd0259bda7d18a46b764d8fb86298a"},
-    {file = "black-23.12.1-cp38-cp38-win_amd64.whl", hash = "sha256:c04b6d9d20e9c13f43eee8ea87d44156b8505ca8a3c878773f68b4e4812a421e"},
-    {file = "black-23.12.1-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:3e1b38b3135fd4c025c28c55ddfc236b05af657828a8a6abe5deec419a0b7055"},
-    {file = "black-23.12.1-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:4f0031eaa7b921db76decd73636ef3a12c942ed367d8c3841a0739412b260a54"},
-    {file = "black-23.12.1-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:97e56155c6b737854e60a9ab1c598ff2533d57e7506d97af5481141671abf3ea"},
-    {file = "black-23.12.1-cp39-cp39-win_amd64.whl", hash = "sha256:dd15245c8b68fe2b6bd0f32c1556509d11bb33aec9b5d0866dd8e2ed3dba09c2"},
-    {file = "black-23.12.1-py3-none-any.whl", hash = "sha256:78baad24af0f033958cad29731e27363183e140962595def56423e626f4bee3e"},
-    {file = "black-23.12.1.tar.gz", hash = "sha256:4ce3ef14ebe8d9509188014d96af1c456a910d5b5cbf434a09fef7e024b3d0d5"},
-]
-
-[package.dependencies]
-click = ">=8.0.0"
-mypy-extensions = ">=0.4.3"
-packaging = ">=22.0"
-pathspec = ">=0.9.0"
-platformdirs = ">=2"
-tomli = {version = ">=1.1.0", markers = "python_version < \"3.11\""}
-typing-extensions = {version = ">=4.0.1", markers = "python_version < \"3.11\""}
-
-[package.extras]
-colorama = ["colorama (>=0.4.3)"]
-d = ["aiohttp (>=3.7.4)", "aiohttp (>=3.7.4,!=3.9.0)"]
-jupyter = ["ipython (>=7.8.0)", "tokenize-rt (>=3.2.0)"]
-uvloop = ["uvloop (>=0.15.2)"]
-
-[[package]]
-name = "cfgv"
-version = "3.4.0"
-description = "Validate configuration and produce human readable error messages."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "cfgv-3.4.0-py2.py3-none-any.whl", hash = "sha256:b7265b1f29fd3316bfcd2b330d63d024f2bfd8bcb8b0272f8e19a504856c48f9"},
-    {file = "cfgv-3.4.0.tar.gz", hash = "sha256:e52591d4c5f5dead8e0f673fb16db7949d2cfb3f7da4582893288f0ded8fe560"},
-]
-
-[[package]]
-name = "click"
-version = "8.1.8"
-description = "Composable command line interface toolkit"
-optional = false
-python-versions = ">=3.7"
-files = [
-    {file = "click-8.1.8-py3-none-any.whl", hash = "sha256:63c132bbbed01578a06712a2d1f497bb62d9c1c0d329b7903a866228027263b2"},
-    {file = "click-8.1.8.tar.gz", hash = "sha256:ed53c9d8990d83c2a27deae68e4ee337473f6330c040a31d4225c9574d16096a"},
-]
-
-[package.dependencies]
-colorama = {version = "*", markers = "platform_system == \"Windows\""}
-
-[[package]]
-name = "colorama"
-version = "0.4.6"
-description = "Cross-platform colored terminal text."
-optional = false
-python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
-files = [
-    {file = "colorama-0.4.6-py2.py3-none-any.whl", hash = "sha256:4f1d9991f5acc0ca119f9d443620b77f9d6b33703e51011c16baf57afb285fc6"},
-    {file = "colorama-0.4.6.tar.gz", hash = "sha256:08695f5cb7ed6e0531a20572697297273c47b8cae5a63ffc6d6ed5c201be6e44"},
-]
-
-[[package]]
-name = "decorator"
-version = "5.2.1"
-description = "Decorators for Humans"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "decorator-5.2.1-py3-none-any.whl", hash = "sha256:d316bb415a2d9e2d2b3abcc4084c6502fc09240e292cd76a76afc106a1c8e04a"},
-    {file = "decorator-5.2.1.tar.gz", hash = "sha256:65f266143752f734b0a7cc83c46f4618af75b8c5911b00ccb61d0ac9b6da0360"},
-]
-
-[[package]]
-name = "distlib"
-version = "0.3.9"
-description = "Distribution utilities"
-optional = false
-python-versions = "*"
-files = [
-    {file = "distlib-0.3.9-py2.py3-none-any.whl", hash = "sha256:47f8c22fd27c27e25a65601af709b38e4f0a45ea4fc2e710f65755fa8caaaf87"},
-    {file = "distlib-0.3.9.tar.gz", hash = "sha256:a60f20dea646b8a33f3e7772f74dc0b2d0772d2837ee1342a00645c81edf9403"},
-]
-
-[[package]]
-name = "exceptiongroup"
-version = "1.3.0"
-description = "Backport of PEP 654 (exception groups)"
-optional = false
-python-versions = ">=3.7"
-files = [
-    {file = "exceptiongroup-1.3.0-py3-none-any.whl", hash = "sha256:4d111e6e0c13d0644cad6ddaa7ed0261a0b36971f6d23e7ec9b4b9097da78a10"},
-    {file = "exceptiongroup-1.3.0.tar.gz", hash = "sha256:b241f5885f560bc56a59ee63ca4c6a8bfa46ae4ad651af316d4e81817bb9fd88"},
-]
-
-[package.dependencies]
-typing-extensions = {version = ">=4.6.0", markers = "python_version < \"3.13\""}
-
-[package.extras]
-test = ["pytest (>=6)"]
-
-[[package]]
-name = "executing"
-version = "2.2.0"
-description = "Get the currently executing AST node of a frame, and other information"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "executing-2.2.0-py2.py3-none-any.whl", hash = "sha256:11387150cad388d62750327a53d3339fad4888b39a6fe233c3afbb54ecffd3aa"},
-    {file = "executing-2.2.0.tar.gz", hash = "sha256:5d108c028108fe2551d1a7b2e8b713341e2cb4fc0aa7dcf966fa4327a5226755"},
-]
-
-[package.extras]
-tests = ["asttokens (>=2.1.0)", "coverage", "coverage-enable-subprocess", "ipython", "littleutils", "pytest", "rich"]
-
-[[package]]
-name = "filelock"
-version = "3.18.0"
-description = "A platform independent file lock."
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "filelock-3.18.0-py3-none-any.whl", hash = "sha256:c401f4f8377c4464e6db25fff06205fd89bdd83b65eb0488ed1b160f780e21de"},
-    {file = "filelock-3.18.0.tar.gz", hash = "sha256:adbc88eabb99d2fec8c9c1b229b171f18afa655400173ddc653d5d01501fb9f2"},
-]
-
-[package.extras]
-docs = ["furo (>=2024.8.6)", "sphinx (>=8.1.3)", "sphinx-autodoc-typehints (>=3)"]
-testing = ["covdefaults (>=2.3)", "coverage (>=7.6.10)", "diff-cover (>=9.2.1)", "pytest (>=8.3.4)", "pytest-asyncio (>=0.25.2)", "pytest-cov (>=6)", "pytest-mock (>=3.14)", "pytest-timeout (>=2.3.1)", "virtualenv (>=20.28.1)"]
-typing = ["typing-extensions (>=4.12.2)"]
-
-[[package]]
-name = "identify"
-version = "2.6.12"
-description = "File identification library for Python"
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "identify-2.6.12-py2.py3-none-any.whl", hash = "sha256:ad9672d5a72e0d2ff7c5c8809b62dfa60458626352fb0eb7b55e69bdc45334a2"},
-    {file = "identify-2.6.12.tar.gz", hash = "sha256:d8de45749f1efb108badef65ee8386f0f7bb19a7f26185f74de6367bffbaf0e6"},
-]
-
-[package.extras]
-license = ["ukkonen"]
-
-[[package]]
-name = "ipython"
-version = "8.18.1"
-description = "IPython: Productive Interactive Computing"
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "ipython-8.18.1-py3-none-any.whl", hash = "sha256:e8267419d72d81955ec1177f8a29aaa90ac80ad647499201119e2f05e99aa397"},
-    {file = "ipython-8.18.1.tar.gz", hash = "sha256:ca6f079bb33457c66e233e4580ebfc4128855b4cf6370dddd73842a9563e8a27"},
-]
-
-[package.dependencies]
-colorama = {version = "*", markers = "sys_platform == \"win32\""}
-decorator = "*"
-exceptiongroup = {version = "*", markers = "python_version < \"3.11\""}
-jedi = ">=0.16"
-matplotlib-inline = "*"
-pexpect = {version = ">4.3", markers = "sys_platform != \"win32\""}
-prompt-toolkit = ">=3.0.41,<3.1.0"
-pygments = ">=2.4.0"
-stack-data = "*"
-traitlets = ">=5"
-typing-extensions = {version = "*", markers = "python_version < \"3.10\""}
-
-[package.extras]
-all = ["black", "curio", "docrepr", "exceptiongroup", "ipykernel", "ipyparallel", "ipywidgets", "matplotlib", "matplotlib (!=3.2.0)", "nbconvert", "nbformat", "notebook", "numpy (>=1.22)", "pandas", "pickleshare", "pytest (<7)", "pytest (<7.1)", "pytest-asyncio (<0.22)", "qtconsole", "setuptools (>=18.5)", "sphinx (>=1.3)", "sphinx-rtd-theme", "stack-data", "testpath", "trio", "typing-extensions"]
-black = ["black"]
-doc = ["docrepr", "exceptiongroup", "ipykernel", "matplotlib", "pickleshare", "pytest (<7)", "pytest (<7.1)", "pytest-asyncio (<0.22)", "setuptools (>=18.5)", "sphinx (>=1.3)", "sphinx-rtd-theme", "stack-data", "testpath", "typing-extensions"]
-kernel = ["ipykernel"]
-nbconvert = ["nbconvert"]
-nbformat = ["nbformat"]
-notebook = ["ipywidgets", "notebook"]
-parallel = ["ipyparallel"]
-qtconsole = ["qtconsole"]
-test = ["pickleshare", "pytest (<7.1)", "pytest-asyncio (<0.22)", "testpath"]
-test-extra = ["curio", "matplotlib (!=3.2.0)", "nbformat", "numpy (>=1.22)", "pandas", "pickleshare", "pytest (<7.1)", "pytest-asyncio (<0.22)", "testpath", "trio"]
-
-[[package]]
-name = "isort"
-version = "5.13.2"
-description = "A Python utility / library to sort Python imports."
-optional = false
-python-versions = ">=3.8.0"
-files = [
-    {file = "isort-5.13.2-py3-none-any.whl", hash = "sha256:8ca5e72a8d85860d5a3fa69b8745237f2939afe12dbf656afbcb47fe72d947a6"},
-    {file = "isort-5.13.2.tar.gz", hash = "sha256:48fdfcb9face5d58a4f6dde2e72a1fb8dcaf8ab26f95ab49fab84c2ddefb0109"},
-]
-
-[package.extras]
-colors = ["colorama (>=0.4.6)"]
-
-[[package]]
-name = "jedi"
-version = "0.19.2"
-description = "An autocompletion tool for Python that can be used for text editors."
-optional = false
-python-versions = ">=3.6"
-files = [
-    {file = "jedi-0.19.2-py2.py3-none-any.whl", hash = "sha256:a8ef22bde8490f57fe5c7681a3c83cb58874daf72b4784de3cce5b6ef6edb5b9"},
-    {file = "jedi-0.19.2.tar.gz", hash = "sha256:4770dc3de41bde3966b02eb84fbcf557fb33cce26ad23da12c742fb50ecb11f0"},
-]
-
-[package.dependencies]
-parso = ">=0.8.4,<0.9.0"
-
-[package.extras]
-docs = ["Jinja2 (==2.11.3)", "MarkupSafe (==1.1.1)", "Pygments (==2.8.1)", "alabaster (==0.7.12)", "babel (==2.9.1)", "chardet (==4.0.0)", "commonmark (==0.8.1)", "docutils (==0.17.1)", "future (==0.18.2)", "idna (==2.10)", "imagesize (==1.2.0)", "mock (==1.0.1)", "packaging (==20.9)", "pyparsing (==2.4.7)", "pytz (==2021.1)", "readthedocs-sphinx-ext (==2.1.4)", "recommonmark (==0.5.0)", "requests (==2.25.1)", "six (==1.15.0)", "snowballstemmer (==2.1.0)", "sphinx (==1.8.5)", "sphinx-rtd-theme (==0.4.3)", "sphinxcontrib-serializinghtml (==1.1.4)", "sphinxcontrib-websupport (==1.2.4)", "urllib3 (==1.26.4)"]
-qa = ["flake8 (==5.0.4)", "mypy (==0.971)", "types-setuptools (==67.2.0.1)"]
-testing = ["Django", "attrs", "colorama", "docopt", "pytest (<9.0.0)"]
-
-[[package]]
-name = "matplotlib-inline"
-version = "0.1.7"
-description = "Inline Matplotlib backend for Jupyter"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "matplotlib_inline-0.1.7-py3-none-any.whl", hash = "sha256:df192d39a4ff8f21b1895d72e6a13f5fcc5099f00fa84384e0ea28c2cc0653ca"},
-    {file = "matplotlib_inline-0.1.7.tar.gz", hash = "sha256:8423b23ec666be3d16e16b60bdd8ac4e86e840ebd1dd11a30b9f117f2fa0ab90"},
-]
-
-[package.dependencies]
-traitlets = "*"
-
-[[package]]
-name = "mypy-extensions"
-version = "1.1.0"
-description = "Type system extensions for programs checked with the mypy type checker."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "mypy_extensions-1.1.0-py3-none-any.whl", hash = "sha256:1be4cccdb0f2482337c4743e60421de3a356cd97508abadd57d47403e94f5505"},
-    {file = "mypy_extensions-1.1.0.tar.gz", hash = "sha256:52e68efc3284861e772bbcd66823fde5ae21fd2fdb51c62a211403730b916558"},
-]
-
-[[package]]
-name = "nodeenv"
-version = "1.9.1"
-description = "Node.js virtual environment builder"
-optional = false
-python-versions = "!=3.0.*,!=3.1.*,!=3.2.*,!=3.3.*,!=3.4.*,!=3.5.*,!=3.6.*,>=2.7"
-files = [
-    {file = "nodeenv-1.9.1-py2.py3-none-any.whl", hash = "sha256:ba11c9782d29c27c70ffbdda2d7415098754709be8a7056d79a737cd901155c9"},
-    {file = "nodeenv-1.9.1.tar.gz", hash = "sha256:6ec12890a2dab7946721edbfbcd91f3319c6ccc9aec47be7c7e6b7011ee6645f"},
-]
-
-[[package]]
-name = "packaging"
-version = "25.0"
-description = "Core utilities for Python packages"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "packaging-25.0-py3-none-any.whl", hash = "sha256:29572ef2b1f17581046b3a2227d5c611fb25ec70ca1ba8554b24b0e69331a484"},
-    {file = "packaging-25.0.tar.gz", hash = "sha256:d443872c98d677bf60f6a1f2f8c1cb748e8fe762d2bf9d3148b5599295b0fc4f"},
-]
-
-[[package]]
-name = "parso"
-version = "0.8.4"
-description = "A Python Parser"
-optional = false
-python-versions = ">=3.6"
-files = [
-    {file = "parso-0.8.4-py2.py3-none-any.whl", hash = "sha256:a418670a20291dacd2dddc80c377c5c3791378ee1e8d12bffc35420643d43f18"},
-    {file = "parso-0.8.4.tar.gz", hash = "sha256:eb3a7b58240fb99099a345571deecc0f9540ea5f4dd2fe14c2a99d6b281ab92d"},
-]
-
-[package.extras]
-qa = ["flake8 (==5.0.4)", "mypy (==0.971)", "types-setuptools (==67.2.0.1)"]
-testing = ["docopt", "pytest"]
-
-[[package]]
-name = "pathspec"
-version = "0.12.1"
-description = "Utility library for gitignore style pattern matching of file paths."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "pathspec-0.12.1-py3-none-any.whl", hash = "sha256:a0d503e138a4c123b27490a4f7beda6a01c6f288df0e4a8b79c7eb0dc7b4cc08"},
-    {file = "pathspec-0.12.1.tar.gz", hash = "sha256:a482d51503a1ab33b1c67a6c3813a26953dbdc71c31dacaef9a838c4e29f5712"},
-]
-
-[[package]]
-name = "pexpect"
-version = "4.9.0"
-description = "Pexpect allows easy control of interactive console applications."
-optional = false
-python-versions = "*"
-files = [
-    {file = "pexpect-4.9.0-py2.py3-none-any.whl", hash = "sha256:7236d1e080e4936be2dc3e326cec0af72acf9212a7e1d060210e70a47e253523"},
-    {file = "pexpect-4.9.0.tar.gz", hash = "sha256:ee7d41123f3c9911050ea2c2dac107568dc43b2d3b0c7557a33212c398ead30f"},
-]
-
-[package.dependencies]
-ptyprocess = ">=0.5"
-
-[[package]]
-name = "platformdirs"
-version = "4.3.8"
-description = "A small Python package for determining appropriate platform-specific dirs, e.g. a `user data dir`."
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "platformdirs-4.3.8-py3-none-any.whl", hash = "sha256:ff7059bb7eb1179e2685604f4aaf157cfd9535242bd23742eadc3c13542139b4"},
-    {file = "platformdirs-4.3.8.tar.gz", hash = "sha256:3d512d96e16bcb959a814c9f348431070822a6496326a4be0911c40b5a74c2bc"},
-]
-
-[package.extras]
-docs = ["furo (>=2024.8.6)", "proselint (>=0.14)", "sphinx (>=8.1.3)", "sphinx-autodoc-typehints (>=3)"]
-test = ["appdirs (==1.4.4)", "covdefaults (>=2.3)", "pytest (>=8.3.4)", "pytest-cov (>=6)", "pytest-mock (>=3.14)"]
-type = ["mypy (>=1.14.1)"]
-
-[[package]]
-name = "pre-commit"
-version = "3.8.0"
-description = "A framework for managing and maintaining multi-language pre-commit hooks."
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "pre_commit-3.8.0-py2.py3-none-any.whl", hash = "sha256:9a90a53bf82fdd8778d58085faf8d83df56e40dfe18f45b19446e26bf1b3a63f"},
-    {file = "pre_commit-3.8.0.tar.gz", hash = "sha256:8bb6494d4a20423842e198980c9ecf9f96607a07ea29549e180eef9ae80fe7af"},
-]
-
-[package.dependencies]
-cfgv = ">=2.0.0"
-identify = ">=1.0.0"
-nodeenv = ">=0.11.1"
-pyyaml = ">=5.1"
-virtualenv = ">=20.10.0"
-
-[[package]]
-name = "prompt-toolkit"
-version = "3.0.51"
-description = "Library for building powerful interactive command lines in Python"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "prompt_toolkit-3.0.51-py3-none-any.whl", hash = "sha256:52742911fde84e2d423e2f9a4cf1de7d7ac4e51958f648d9540e0fb8db077b07"},
-    {file = "prompt_toolkit-3.0.51.tar.gz", hash = "sha256:931a162e3b27fc90c86f1b48bb1fb2c528c2761475e57c9c06de13311c7b54ed"},
-]
-
-[package.dependencies]
-wcwidth = "*"
-
-[[package]]
-name = "ptyprocess"
-version = "0.7.0"
-description = "Run a subprocess in a pseudo terminal"
-optional = false
-python-versions = "*"
-files = [
-    {file = "ptyprocess-0.7.0-py2.py3-none-any.whl", hash = "sha256:4b41f3967fce3af57cc7e94b888626c18bf37a083e3651ca8feeb66d492fef35"},
-    {file = "ptyprocess-0.7.0.tar.gz", hash = "sha256:5c5d0a3b48ceee0b48485e0c26037c0acd7d29765ca3fbb5cb3831d347423220"},
-]
-
-[[package]]
-name = "pure-eval"
-version = "0.2.3"
-description = "Safely evaluate AST nodes without side effects"
-optional = false
-python-versions = "*"
-files = [
-    {file = "pure_eval-0.2.3-py3-none-any.whl", hash = "sha256:1db8e35b67b3d218d818ae653e27f06c3aa420901fa7b081ca98cbedc874e0d0"},
-    {file = "pure_eval-0.2.3.tar.gz", hash = "sha256:5f4e983f40564c576c7c8635ae88db5956bb2229d7e9237d03b3c0b0190eaf42"},
-]
-
-[package.extras]
-tests = ["pytest"]
-
-[[package]]
-name = "pygments"
-version = "2.19.2"
-description = "Pygments is a syntax highlighting package written in Python."
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "pygments-2.19.2-py3-none-any.whl", hash = "sha256:86540386c03d588bb81d44bc3928634ff26449851e99741617ecb9037ee5ec0b"},
-    {file = "pygments-2.19.2.tar.gz", hash = "sha256:636cb2477cec7f8952536970bc533bc43743542f70392ae026374600add5b887"},
-]
-
-[package.extras]
-windows-terminal = ["colorama (>=0.4.6)"]
-
-[[package]]
-name = "pyyaml"
-version = "6.0.2"
-description = "YAML parser and emitter for Python"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "PyYAML-6.0.2-cp310-cp310-macosx_10_9_x86_64.whl", hash = "sha256:0a9a2848a5b7feac301353437eb7d5957887edbf81d56e903999a75a3d743086"},
-    {file = "PyYAML-6.0.2-cp310-cp310-macosx_11_0_arm64.whl", hash = "sha256:29717114e51c84ddfba879543fb232a6ed60086602313ca38cce623c1d62cfbf"},
-    {file = "PyYAML-6.0.2-cp310-cp310-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:8824b5a04a04a047e72eea5cec3bc266db09e35de6bdfe34c9436ac5ee27d237"},
-    {file = "PyYAML-6.0.2-cp310-cp310-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:7c36280e6fb8385e520936c3cb3b8042851904eba0e58d277dca80a5cfed590b"},
-    {file = "PyYAML-6.0.2-cp310-cp310-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:ec031d5d2feb36d1d1a24380e4db6d43695f3748343d99434e6f5f9156aaa2ed"},
-    {file = "PyYAML-6.0.2-cp310-cp310-musllinux_1_1_aarch64.whl", hash = "sha256:936d68689298c36b53b29f23c6dbb74de12b4ac12ca6cfe0e047bedceea56180"},
-    {file = "PyYAML-6.0.2-cp310-cp310-musllinux_1_1_x86_64.whl", hash = "sha256:23502f431948090f597378482b4812b0caae32c22213aecf3b55325e049a6c68"},
-    {file = "PyYAML-6.0.2-cp310-cp310-win32.whl", hash = "sha256:2e99c6826ffa974fe6e27cdb5ed0021786b03fc98e5ee3c5bfe1fd5015f42b99"},
-    {file = "PyYAML-6.0.2-cp310-cp310-win_amd64.whl", hash = "sha256:a4d3091415f010369ae4ed1fc6b79def9416358877534caf6a0fdd2146c87a3e"},
-    {file = "PyYAML-6.0.2-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:cc1c1159b3d456576af7a3e4d1ba7e6924cb39de8f67111c735f6fc832082774"},
-    {file = "PyYAML-6.0.2-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:1e2120ef853f59c7419231f3bf4e7021f1b936f6ebd222406c3b60212205d2ee"},
-    {file = "PyYAML-6.0.2-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:5d225db5a45f21e78dd9358e58a98702a0302f2659a3c6cd320564b75b86f47c"},
-    {file = "PyYAML-6.0.2-cp311-cp311-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:5ac9328ec4831237bec75defaf839f7d4564be1e6b25ac710bd1a96321cc8317"},
-    {file = "PyYAML-6.0.2-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3ad2a3decf9aaba3d29c8f537ac4b243e36bef957511b4766cb0057d32b0be85"},
-    {file = "PyYAML-6.0.2-cp311-cp311-musllinux_1_1_aarch64.whl", hash = "sha256:ff3824dc5261f50c9b0dfb3be22b4567a6f938ccce4587b38952d85fd9e9afe4"},
-    {file = "PyYAML-6.0.2-cp311-cp311-musllinux_1_1_x86_64.whl", hash = "sha256:797b4f722ffa07cc8d62053e4cff1486fa6dc094105d13fea7b1de7d8bf71c9e"},
-    {file = "PyYAML-6.0.2-cp311-cp311-win32.whl", hash = "sha256:11d8f3dd2b9c1207dcaf2ee0bbbfd5991f571186ec9cc78427ba5bd32afae4b5"},
-    {file = "PyYAML-6.0.2-cp311-cp311-win_amd64.whl", hash = "sha256:e10ce637b18caea04431ce14fabcf5c64a1c61ec9c56b071a4b7ca131ca52d44"},
-    {file = "PyYAML-6.0.2-cp312-cp312-macosx_10_9_x86_64.whl", hash = "sha256:c70c95198c015b85feafc136515252a261a84561b7b1d51e3384e0655ddf25ab"},
-    {file = "PyYAML-6.0.2-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:ce826d6ef20b1bc864f0a68340c8b3287705cae2f8b4b1d932177dcc76721725"},
-    {file = "PyYAML-6.0.2-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:1f71ea527786de97d1a0cc0eacd1defc0985dcf6b3f17bb77dcfc8c34bec4dc5"},
-    {file = "PyYAML-6.0.2-cp312-cp312-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:9b22676e8097e9e22e36d6b7bda33190d0d400f345f23d4065d48f4ca7ae0425"},
-    {file = "PyYAML-6.0.2-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:80bab7bfc629882493af4aa31a4cfa43a4c57c83813253626916b8c7ada83476"},
-    {file = "PyYAML-6.0.2-cp312-cp312-musllinux_1_1_aarch64.whl", hash = "sha256:0833f8694549e586547b576dcfaba4a6b55b9e96098b36cdc7ebefe667dfed48"},
-    {file = "PyYAML-6.0.2-cp312-cp312-musllinux_1_1_x86_64.whl", hash = "sha256:8b9c7197f7cb2738065c481a0461e50ad02f18c78cd75775628afb4d7137fb3b"},
-    {file = "PyYAML-6.0.2-cp312-cp312-win32.whl", hash = "sha256:ef6107725bd54b262d6dedcc2af448a266975032bc85ef0172c5f059da6325b4"},
-    {file = "PyYAML-6.0.2-cp312-cp312-win_amd64.whl", hash = "sha256:7e7401d0de89a9a855c839bc697c079a4af81cf878373abd7dc625847d25cbd8"},
-    {file = "PyYAML-6.0.2-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:efdca5630322a10774e8e98e1af481aad470dd62c3170801852d752aa7a783ba"},
-    {file = "PyYAML-6.0.2-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:50187695423ffe49e2deacb8cd10510bc361faac997de9efef88badc3bb9e2d1"},
-    {file = "PyYAML-6.0.2-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:0ffe8360bab4910ef1b9e87fb812d8bc0a308b0d0eef8c8f44e0254ab3b07133"},
-    {file = "PyYAML-6.0.2-cp313-cp313-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:17e311b6c678207928d649faa7cb0d7b4c26a0ba73d41e99c4fff6b6c3276484"},
-    {file = "PyYAML-6.0.2-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:70b189594dbe54f75ab3a1acec5f1e3faa7e8cf2f1e08d9b561cb41b845f69d5"},
-    {file = "PyYAML-6.0.2-cp313-cp313-musllinux_1_1_aarch64.whl", hash = "sha256:41e4e3953a79407c794916fa277a82531dd93aad34e29c2a514c2c0c5fe971cc"},
-    {file = "PyYAML-6.0.2-cp313-cp313-musllinux_1_1_x86_64.whl", hash = "sha256:68ccc6023a3400877818152ad9a1033e3db8625d899c72eacb5a668902e4d652"},
-    {file = "PyYAML-6.0.2-cp313-cp313-win32.whl", hash = "sha256:bc2fa7c6b47d6bc618dd7fb02ef6fdedb1090ec036abab80d4681424b84c1183"},
-    {file = "PyYAML-6.0.2-cp313-cp313-win_amd64.whl", hash = "sha256:8388ee1976c416731879ac16da0aff3f63b286ffdd57cdeb95f3f2e085687563"},
-    {file = "PyYAML-6.0.2-cp38-cp38-macosx_10_9_x86_64.whl", hash = "sha256:24471b829b3bf607e04e88d79542a9d48bb037c2267d7927a874e6c205ca7e9a"},
-    {file = "PyYAML-6.0.2-cp38-cp38-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d7fded462629cfa4b685c5416b949ebad6cec74af5e2d42905d41e257e0869f5"},
-    {file = "PyYAML-6.0.2-cp38-cp38-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:d84a1718ee396f54f3a086ea0a66d8e552b2ab2017ef8b420e92edbc841c352d"},
-    {file = "PyYAML-6.0.2-cp38-cp38-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9056c1ecd25795207ad294bcf39f2db3d845767be0ea6e6a34d856f006006083"},
-    {file = "PyYAML-6.0.2-cp38-cp38-musllinux_1_1_x86_64.whl", hash = "sha256:82d09873e40955485746739bcb8b4586983670466c23382c19cffecbf1fd8706"},
-    {file = "PyYAML-6.0.2-cp38-cp38-win32.whl", hash = "sha256:43fa96a3ca0d6b1812e01ced1044a003533c47f6ee8aca31724f78e93ccc089a"},
-    {file = "PyYAML-6.0.2-cp38-cp38-win_amd64.whl", hash = "sha256:01179a4a8559ab5de078078f37e5c1a30d76bb88519906844fd7bdea1b7729ff"},
-    {file = "PyYAML-6.0.2-cp39-cp39-macosx_10_9_x86_64.whl", hash = "sha256:688ba32a1cffef67fd2e9398a2efebaea461578b0923624778664cc1c914db5d"},
-    {file = "PyYAML-6.0.2-cp39-cp39-macosx_11_0_arm64.whl", hash = "sha256:a8786accb172bd8afb8be14490a16625cbc387036876ab6ba70912730faf8e1f"},
-    {file = "PyYAML-6.0.2-cp39-cp39-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:d8e03406cac8513435335dbab54c0d385e4a49e4945d2909a581c83647ca0290"},
-    {file = "PyYAML-6.0.2-cp39-cp39-manylinux_2_17_s390x.manylinux2014_s390x.whl", hash = "sha256:f753120cb8181e736c57ef7636e83f31b9c0d1722c516f7e86cf15b7aa57ff12"},
-    {file = "PyYAML-6.0.2-cp39-cp39-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:3b1fdb9dc17f5a7677423d508ab4f243a726dea51fa5e70992e59a7411c89d19"},
-    {file = "PyYAML-6.0.2-cp39-cp39-musllinux_1_1_aarch64.whl", hash = "sha256:0b69e4ce7a131fe56b7e4d770c67429700908fc0752af059838b1cfb41960e4e"},
-    {file = "PyYAML-6.0.2-cp39-cp39-musllinux_1_1_x86_64.whl", hash = "sha256:a9f8c2e67970f13b16084e04f134610fd1d374bf477b17ec1599185cf611d725"},
-    {file = "PyYAML-6.0.2-cp39-cp39-win32.whl", hash = "sha256:6395c297d42274772abc367baaa79683958044e5d3835486c16da75d2a694631"},
-    {file = "PyYAML-6.0.2-cp39-cp39-win_amd64.whl", hash = "sha256:39693e1f8320ae4f43943590b49779ffb98acb81f788220ea932a6b6c51004d8"},
-    {file = "pyyaml-6.0.2.tar.gz", hash = "sha256:d584d9ec91ad65861cc08d42e834324ef890a082e591037abe114850ff7bbc3e"},
-]
-
-[[package]]
-name = "stack-data"
-version = "0.6.3"
-description = "Extract data from python stack frames and tracebacks for informative displays"
-optional = false
-python-versions = "*"
-files = [
-    {file = "stack_data-0.6.3-py3-none-any.whl", hash = "sha256:d5558e0c25a4cb0853cddad3d77da9891a08cb85dd9f9f91b9f8cd66e511e695"},
-    {file = "stack_data-0.6.3.tar.gz", hash = "sha256:836a778de4fec4dcd1dcd89ed8abff8a221f58308462e1c4aa2a3cf30148f0b9"},
-]
-
-[package.dependencies]
-asttokens = ">=2.1.0"
-executing = ">=1.2.0"
-pure-eval = "*"
-
-[package.extras]
-tests = ["cython", "littleutils", "pygments", "pytest", "typeguard"]
-
-[[package]]
-name = "tomli"
-version = "2.2.1"
-description = "A lil' TOML parser"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "tomli-2.2.1-cp311-cp311-macosx_10_9_x86_64.whl", hash = "sha256:678e4fa69e4575eb77d103de3df8a895e1591b48e740211bd1067378c69e8249"},
-    {file = "tomli-2.2.1-cp311-cp311-macosx_11_0_arm64.whl", hash = "sha256:023aa114dd824ade0100497eb2318602af309e5a55595f76b626d6d9f3b7b0a6"},
-    {file = "tomli-2.2.1-cp311-cp311-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:ece47d672db52ac607a3d9599a9d48dcb2f2f735c6c2d1f34130085bb12b112a"},
-    {file = "tomli-2.2.1-cp311-cp311-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:6972ca9c9cc9f0acaa56a8ca1ff51e7af152a9f87fb64623e31d5c83700080ee"},
-    {file = "tomli-2.2.1-cp311-cp311-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:c954d2250168d28797dd4e3ac5cf812a406cd5a92674ee4c8f123c889786aa8e"},
-    {file = "tomli-2.2.1-cp311-cp311-musllinux_1_2_aarch64.whl", hash = "sha256:8dd28b3e155b80f4d54beb40a441d366adcfe740969820caf156c019fb5c7ec4"},
-    {file = "tomli-2.2.1-cp311-cp311-musllinux_1_2_i686.whl", hash = "sha256:e59e304978767a54663af13c07b3d1af22ddee3bb2fb0618ca1593e4f593a106"},
-    {file = "tomli-2.2.1-cp311-cp311-musllinux_1_2_x86_64.whl", hash = "sha256:33580bccab0338d00994d7f16f4c4ec25b776af3ffaac1ed74e0b3fc95e885a8"},
-    {file = "tomli-2.2.1-cp311-cp311-win32.whl", hash = "sha256:465af0e0875402f1d226519c9904f37254b3045fc5084697cefb9bdde1ff99ff"},
-    {file = "tomli-2.2.1-cp311-cp311-win_amd64.whl", hash = "sha256:2d0f2fdd22b02c6d81637a3c95f8cd77f995846af7414c5c4b8d0545afa1bc4b"},
-    {file = "tomli-2.2.1-cp312-cp312-macosx_10_13_x86_64.whl", hash = "sha256:4a8f6e44de52d5e6c657c9fe83b562f5f4256d8ebbfe4ff922c495620a7f6cea"},
-    {file = "tomli-2.2.1-cp312-cp312-macosx_11_0_arm64.whl", hash = "sha256:8d57ca8095a641b8237d5b079147646153d22552f1c637fd3ba7f4b0b29167a8"},
-    {file = "tomli-2.2.1-cp312-cp312-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:4e340144ad7ae1533cb897d406382b4b6fede8890a03738ff1683af800d54192"},
-    {file = "tomli-2.2.1-cp312-cp312-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:db2b95f9de79181805df90bedc5a5ab4c165e6ec3fe99f970d0e302f384ad222"},
-    {file = "tomli-2.2.1-cp312-cp312-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:40741994320b232529c802f8bc86da4e1aa9f413db394617b9a256ae0f9a7f77"},
-    {file = "tomli-2.2.1-cp312-cp312-musllinux_1_2_aarch64.whl", hash = "sha256:400e720fe168c0f8521520190686ef8ef033fb19fc493da09779e592861b78c6"},
-    {file = "tomli-2.2.1-cp312-cp312-musllinux_1_2_i686.whl", hash = "sha256:02abe224de6ae62c19f090f68da4e27b10af2b93213d36cf44e6e1c5abd19fdd"},
-    {file = "tomli-2.2.1-cp312-cp312-musllinux_1_2_x86_64.whl", hash = "sha256:b82ebccc8c8a36f2094e969560a1b836758481f3dc360ce9a3277c65f374285e"},
-    {file = "tomli-2.2.1-cp312-cp312-win32.whl", hash = "sha256:889f80ef92701b9dbb224e49ec87c645ce5df3fa2cc548664eb8a25e03127a98"},
-    {file = "tomli-2.2.1-cp312-cp312-win_amd64.whl", hash = "sha256:7fc04e92e1d624a4a63c76474610238576942d6b8950a2d7f908a340494e67e4"},
-    {file = "tomli-2.2.1-cp313-cp313-macosx_10_13_x86_64.whl", hash = "sha256:f4039b9cbc3048b2416cc57ab3bda989a6fcf9b36cf8937f01a6e731b64f80d7"},
-    {file = "tomli-2.2.1-cp313-cp313-macosx_11_0_arm64.whl", hash = "sha256:286f0ca2ffeeb5b9bd4fcc8d6c330534323ec51b2f52da063b11c502da16f30c"},
-    {file = "tomli-2.2.1-cp313-cp313-manylinux_2_17_aarch64.manylinux2014_aarch64.whl", hash = "sha256:a92ef1a44547e894e2a17d24e7557a5e85a9e1d0048b0b5e7541f76c5032cb13"},
-    {file = "tomli-2.2.1-cp313-cp313-manylinux_2_17_x86_64.manylinux2014_x86_64.whl", hash = "sha256:9316dc65bed1684c9a98ee68759ceaed29d229e985297003e494aa825ebb0281"},
-    {file = "tomli-2.2.1-cp313-cp313-manylinux_2_5_i686.manylinux1_i686.manylinux_2_17_i686.manylinux2014_i686.whl", hash = "sha256:e85e99945e688e32d5a35c1ff38ed0b3f41f43fad8df0bdf79f72b2ba7bc5272"},
-    {file = "tomli-2.2.1-cp313-cp313-musllinux_1_2_aarch64.whl", hash = "sha256:ac065718db92ca818f8d6141b5f66369833d4a80a9d74435a268c52bdfa73140"},
-    {file = "tomli-2.2.1-cp313-cp313-musllinux_1_2_i686.whl", hash = "sha256:d920f33822747519673ee656a4b6ac33e382eca9d331c87770faa3eef562aeb2"},
-    {file = "tomli-2.2.1-cp313-cp313-musllinux_1_2_x86_64.whl", hash = "sha256:a198f10c4d1b1375d7687bc25294306e551bf1abfa4eace6650070a5c1ae2744"},
-    {file = "tomli-2.2.1-cp313-cp313-win32.whl", hash = "sha256:d3f5614314d758649ab2ab3a62d4f2004c825922f9e370b29416484086b264ec"},
-    {file = "tomli-2.2.1-cp313-cp313-win_amd64.whl", hash = "sha256:a38aa0308e754b0e3c67e344754dff64999ff9b513e691d0e786265c93583c69"},
-    {file = "tomli-2.2.1-py3-none-any.whl", hash = "sha256:cb55c73c5f4408779d0cf3eef9f762b9c9f147a77de7b258bef0a5628adc85cc"},
-    {file = "tomli-2.2.1.tar.gz", hash = "sha256:cd45e1dc79c835ce60f7404ec8119f2eb06d38b1deba146f07ced3bbc44505ff"},
-]
-
-[[package]]
-name = "traitlets"
-version = "5.14.3"
-description = "Traitlets Python configuration system"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "traitlets-5.14.3-py3-none-any.whl", hash = "sha256:b74e89e397b1ed28cc831db7aea759ba6640cb3de13090ca145426688ff1ac4f"},
-    {file = "traitlets-5.14.3.tar.gz", hash = "sha256:9ed0579d3502c94b4b3732ac120375cda96f923114522847de4b3bb98b96b6b7"},
-]
-
-[package.extras]
-docs = ["myst-parser", "pydata-sphinx-theme", "sphinx"]
-test = ["argcomplete (>=3.0.3)", "mypy (>=1.7.0)", "pre-commit", "pytest (>=7.0,<8.2)", "pytest-mock", "pytest-mypy-testing"]
-
-[[package]]
-name = "typing-extensions"
-version = "4.14.0"
-description = "Backported and Experimental Type Hints for Python 3.9+"
-optional = false
-python-versions = ">=3.9"
-files = [
-    {file = "typing_extensions-4.14.0-py3-none-any.whl", hash = "sha256:a1514509136dd0b477638fc68d6a91497af5076466ad0fa6c338e44e359944af"},
-    {file = "typing_extensions-4.14.0.tar.gz", hash = "sha256:8676b788e32f02ab42d9e7c61324048ae4c6d844a399eebace3d4979d75ceef4"},
-]
-
-[[package]]
-name = "virtualenv"
-version = "20.31.2"
-description = "Virtual Python Environment builder"
-optional = false
-python-versions = ">=3.8"
-files = [
-    {file = "virtualenv-20.31.2-py3-none-any.whl", hash = "sha256:36efd0d9650ee985f0cad72065001e66d49a6f24eb44d98980f630686243cf11"},
-    {file = "virtualenv-20.31.2.tar.gz", hash = "sha256:e10c0a9d02835e592521be48b332b6caee6887f332c111aa79a09b9e79efc2af"},
-]
-
-[package.dependencies]
-distlib = ">=0.3.7,<1"
-filelock = ">=3.12.2,<4"
-platformdirs = ">=3.9.1,<5"
-
-[package.extras]
-docs = ["furo (>=2023.7.26)", "proselint (>=0.13)", "sphinx (>=7.1.2,!=7.3)", "sphinx-argparse (>=0.4)", "sphinxcontrib-towncrier (>=0.2.1a0)", "towncrier (>=23.6)"]
-test = ["covdefaults (>=2.3)", "coverage (>=7.2.7)", "coverage-enable-subprocess (>=1)", "flaky (>=3.7)", "packaging (>=23.1)", "pytest (>=7.4)", "pytest-env (>=0.8.2)", "pytest-freezer (>=0.4.8)", "pytest-mock (>=3.11.1)", "pytest-randomly (>=3.12)", "pytest-timeout (>=2.1)", "setuptools (>=68)", "time-machine (>=2.10)"]
-
-[[package]]
-name = "wcwidth"
-version = "0.2.13"
-description = "Measures the displayed width of unicode strings in a terminal"
-optional = false
-python-versions = "*"
-files = [
-    {file = "wcwidth-0.2.13-py2.py3-none-any.whl", hash = "sha256:3da69048e4540d84af32131829ff948f1e022c1c6bdb8d6102117aac784f6859"},
-    {file = "wcwidth-0.2.13.tar.gz", hash = "sha256:72ea0c06399eb286d978fdedb6923a9eb47e1c486ce63e9b4e64fc18303972b5"},
-]
-
-[metadata]
-lock-version = "2.0"
-python-versions = ">=3.9,<3.13"
-content-hash = "b06d8f3227a29a324890821df9d59fb02a8ef9b93f3db40e2968f95b402f767f"
diff --git a/pyproject.toml b/pyproject.toml
deleted file mode 100644
index 8049116c7..000000000
--- a/pyproject.toml
+++ /dev/null
@@ -1,21 +0,0 @@
-[tool.poetry]
-name = "truss-examples"
-version = "0.1.0"
-description = ""
-authors = ["Truss Maintainers <team@trussml.com>"]
-license = "MIT"
-readme = "README.md"
-package-mode = false
-
-[tool.poetry.dependencies]
-python = ">=3.9,<3.13"
-
-[tool.poetry.group.dev.dependencies]
-black = "^23.7.0"
-ipython = "^8.14.0"
-isort = "^5.12.0"
-pre-commit = "^3.5.0"
-
-[build-system]
-requires = ["poetry-core"]
-build-backend = "poetry.core.masonry.api"
diff --git a/qwen/BEI-qwen-qwen3-embedding-0.6b-A10G/README.md b/qwen/BEI-qwen-qwen3-embedding-0.6b-A10G/README.md
deleted file mode 100644
index c22cf3181..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-0.6b-A10G/README.md
+++ /dev/null
@@ -1,157 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-0.6B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-0.6B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-0.6B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-)
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-embedding-0.6b-A10G/config.yaml b/qwen/BEI-qwen-qwen3-embedding-0.6b-A10G/config.yaml
deleted file mode 100644
index 1dd6b0171..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-0.6b-A10G/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input:
-      [
-        "Baseten Embeddings are high throughput.",
-        "Baseten Embeddings are low latency.",
-      ]
-    model: model
-  tags:
-    - openai-compatible
-model_name: Qwen3 Embedding 0.6B
-python_version: py39
-resources:
-  accelerator: A10G
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8-h100/README.md b/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8-h100/README.md
deleted file mode 100644
index c22cf3181..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8-h100/README.md
+++ /dev/null
@@ -1,157 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-0.6B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-0.6B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-0.6B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-)
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8-h100/config.yaml b/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8-h100/config.yaml
deleted file mode 100644
index 6469b634a..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8-h100/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input:
-      [
-        "Baseten Embeddings are high throughput.",
-        "Baseten Embeddings are low latency.",
-      ]
-    model: model
-  tags:
-    - openai-compatible
-model_name: Qwen3 Embedding 0.6B
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
diff --git a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md b/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md
deleted file mode 100644
index c22cf3181..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8/README.md
+++ /dev/null
@@ -1,157 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-0.6B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-0.6B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-0.6B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-0.6B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-0.6b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-)
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml b/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml
deleted file mode 100644
index 4461d6505..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-0.6b-fp8/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: ["Baseten Embeddings are high throughput.", "Baseten Embeddings are low latency."]
-    model: model
-  tags:
-    - openai-compatible
-model_name: Qwen3 Embedding 0.6B
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-0.6B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
diff --git a/qwen/BEI-qwen-qwen3-embedding-4b-fp8/README.md b/qwen/BEI-qwen-qwen3-embedding-4b-fp8/README.md
deleted file mode 100644
index 91dd4fa44..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-4b-fp8/README.md
+++ /dev/null
@@ -1,126 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-4B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-4B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-Embedding-4B](https://huggingface.co/Qwen/Qwen3-Embedding-4B).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Qwen/Qwen3-Embedding-4B  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-4b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-4b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-)
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml b/qwen/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml
deleted file mode 100644
index e2e56e9ba..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-4b-fp8/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-# this file was autogenerated by `generate_templates.py` - please do change via template only
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: ["Instruct: {task_description}\nQuery:{query}", "Instruct: Given a web search query, retrieve relevant passages that answer the query\nQuery:What is the capital of China?"]
-    model: model
-model_name: library-model-BEI-qwen3-embedding-4b
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-4B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-  version_overrides:
-    bei_version: 0.0.23
-    engine_builder_version: 0.18.1.post10.dev1
diff --git a/qwen/BEI-qwen-qwen3-embedding-8b-fp8/README.md b/qwen/BEI-qwen-qwen3-embedding-8b-fp8/README.md
deleted file mode 100644
index 4c13b4c88..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-8b-fp8/README.md
+++ /dev/null
@@ -1,157 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-8B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Embedding-8B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Embedding-8B-auto](https://huggingface.co/michaelfeil/Qwen3-Embedding-8B-auto).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-michaelfeil/Qwen3-Embedding-8B-auto  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-embedding-8b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-embedding-8b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-)
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: text string
-    model: model
-model_name: BEI-qwen-qwen3-embedding-8b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-8B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml b/qwen/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml
deleted file mode 100644
index 4fcadd06f..000000000
--- a/qwen/BEI-qwen-qwen3-embedding-8b-fp8/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-model_metadata:
-  example_model_input:
-    encoding_format: float
-    input: ["Baseten Embeddings are high throughput.", "Baseten Embeddings are low latency."]
-    model: model
-  tags:
-    - openai-compatible
-model_name: Qwen3 Embedding 8B
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Embedding-8B-auto
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /v1/embeddings
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
diff --git a/qwen/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md b/qwen/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md
deleted file mode 100644
index 89cf966fe..000000000
--- a/qwen/BEI-qwen-qwen3-reranker-0.6b-fp8/README.md
+++ /dev/null
@@ -1,142 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-0.6B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-0.6B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Reranker-0.6B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-0.6B-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/Qwen3-Reranker-0.6B-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-0.6b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-reranker-0.6b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - Baseten is a fast inference provider
-    - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-0.6b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-0.6B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 4
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml b/qwen/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml
deleted file mode 100644
index 7b5e9c718..000000000
--- a/qwen/BEI-qwen-qwen3-reranker-0.6b-fp8/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-model_metadata:
-  example_model_input:
-    inputs:
-    # advanced formatting of the string needed:
-    - "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n<Instruct>: {intruction}\n<Query>: {query}\n<Document>: {doc}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
-    - "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n<Instruct>: Given a web search query, retrieve relevant passages that answer the query\n<Query>: What is the capital of China?\n<Document>: The capital of China is Beijing.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: Qwen3 Reranker 0.6B
-python_version: py39
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-0.6B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
diff --git a/qwen/BEI-qwen-qwen3-reranker-4b-fp8/README.md b/qwen/BEI-qwen-qwen3-reranker-4b-fp8/README.md
deleted file mode 100644
index c77ddb5f7..000000000
--- a/qwen/BEI-qwen-qwen3-reranker-4b-fp8/README.md
+++ /dev/null
@@ -1,126 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-4B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-4B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [Qwen/Qwen3-Reranker-4B](https://huggingface.co/Qwen/Qwen3-Reranker-4B).
-Suitable models need to have the configurations of the `sentence-transformers` library, which are used for embeddings. Such repos contain e.g. a `sbert_config.json` or a `1_Pooling/config.json` file besides the fast-tokenizer and the safetensors file.
-
-Qwen/Qwen3-Reranker-4B  is a text-embeddings model, producing a 1D embeddings vector, given an input.
-It's frequently used for downstream tasks like clustering, used with vector databases.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-4b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-reranker-4b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings`
-```json
-{
-  "encoding_format": "float", # or base64
-  "input": "string", # can be list of strings for multiple embeddings
-  "model": "null",
-  "user": "null"
-}
-```
-
-Returns:
-```json
-{
-  "data": [
-    {
-      "embedding": [
-        0
-      ],
-      "index": 0,
-      "object": "embedding"
-    }
-  ],
-  "model": "thenlper/gte-base",
-  "object": "list",
-  "usage": {
-    "prompt_tokens": 512,
-    "total_tokens": 512
-  }
-}
-```
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/v1/embeddings`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### curl
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string", "model": "model"}'
-```
-
-### OpenAI compatible client library
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ['BASETEN_API_KEY'],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-embedding = client.embeddings.create(
-    input="Baseten Embeddings are fast",
-    model="model"
-)
-```
-### requests python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings",
-    headers={"Authorization": "Api-Key " + str(os.environ['BASETEN_API_KEY'])},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml b/qwen/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml
deleted file mode 100644
index da6222016..000000000
--- a/qwen/BEI-qwen-qwen3-reranker-4b-fp8/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-model_metadata:
-  example_model_input:
-    inputs:
-    # advanced formatting of the string needed:
-    - - "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n<Instruct>: {intruction}\n<Query>: {query}\n<Document>: {doc}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
-    - - "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n<Instruct>: Given a web search query, retrieve relevant passages that answer the query\n<Query>: What is the capital of China?\n<Document>: The capital of China is Beijing.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI Qwen3 Reranker 4b fp8
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-4B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 32768
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-  version_overrides:
-    bei_version: 0.0.23
-    engine_builder_version: 0.18.1.post10.dev1
diff --git a/qwen/BEI-qwen-qwen3-reranker-8b-fp8/README.md b/qwen/BEI-qwen-qwen3-reranker-8b-fp8/README.md
deleted file mode 100644
index 701a13667..000000000
--- a/qwen/BEI-qwen-qwen3-reranker-8b-fp8/README.md
+++ /dev/null
@@ -1,142 +0,0 @@
-# BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-8B
-
-This is a Deployment for BEI (Baseten-Embeddings-Inference) with Qwen/Qwen3-Reranker-8B. BEI is Baseten's solution for production-grade deployments via TensorRT-LLM for (text) embeddings, reranking models and prediction models.
-With BEI you get the following benefits:
-- *Lowest-latency inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama)<sup>1</sup>
-- *Highest-throughput inference* across any embedding solution (vLLM, SGlang, Infinity, TEI, Ollama) - thanks to XQA kernels, FP8 and dynamic batching.<sup>2</sup>
-- High parallelism: up to 1400 client embeddings per second
-- Cached model weights for fast vertical scaling and high availability - no Hugging Face hub dependency at runtime
-
-
-# Examples:
-This deployment is specifically designed for the Hugging Face model [michaelfeil/Qwen3-Reranker-8B-seq](https://huggingface.co/michaelfeil/Qwen3-Reranker-8B-seq).
-Suitable models can be identified by the `ForSequenceClassification` suffix in the model name. Prediction models may have one or more labels, which are returned with the prediction.
-
-michaelfeil/Qwen3-Reranker-8B-seq  is a text-classification model, used to classify a text into a category. \nIt is frequently used in sentiment analysis, spam detection, and more. It's also used for deployment of chat rating models, e.g. RLHF reward models or toxicity detection models.
-
-This model is quantized to FP8 for deployment, which is supported by Nvidia's newest GPUs e.g. H100, H100_40GB or L4. Quantization is optional, but leads to higher efficiency.
-
-## Deployment with Truss
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-
-First, clone this repository:
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd 11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8
-```
-
-With `11-embeddings-reranker-classification-tensorrt/BEI-qwen-qwen3-reranker-8b-fp8` as your working directory, you can deploy the model with the following command. Paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-# prints:
-# ✨ Model BEI-qwen-qwen3-reranker-8b-fp8-truss-example was successfully pushed ✨
-# 🪵  View logs for your deployment at https://app.baseten.co/models/yyyyyy/logs/xxxxxx
-```
-
-## Call your model
-
-### API-Schema:
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/sync/predict`
-```json
-{
-  "inputs": "Baseten is a fast inference provider",
-  "raw_scores": true,
-  "truncate": true,
-  "truncation_direction": "Right"
-}
-```
-
-```python
-import requests
-import os
-
-headers = {
-    f"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"
-}
-
-requests.post(
-    headers=headers,
-    url="https://model-xxxxxx.api.baseten.co/environments/production/sync/predict",
-    json={
-        "inputs": [["Baseten is a fast inference provider", ["classify this separately."]],
-        "raw_scores": True,
-        "truncate": True,
-        "truncation_direction": "Right"
-    }
-)
-```
-Returns:
-```json
-[
-  [
-    {
-        "label": "excitement",
-        "score": 0.99
-    }
-  ],
-  [
-    {
-        "label": "excitement",
-        "score": 0.01
-    }
-  ]
-]
-```
-Important, this is different from the `predict` route that you usually call. (https://model-xxxxxx.api.baseten.co/environments/production/predict), it contains an additional `sync` before that.
-The OpenAPI.json is available under https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json for more details.
-
-#### Advanced:
-You may also use Baseten's async jobs API, which returns a request_id, which you can use to query the status of the job and get the results.
-
-POST-Route: `https://model-xxxxxx.api.baseten.co/environments/production/async/predict`
-Read more about [Baseten's Async API here](https://docs.baseten.co/invoke/async)
-
-### OpenAI compatible client library
-OpenAI does not have a classification endpoint, therefore no client library is available.
-
-
-## Config.yaml
-By default, the following configuration is used for this deployment. This config uses `quantization_type=fp8`. This is optional, remove the `quantization_type` field or set it to `no_quant` for float16/bfloat16.
-
-```yaml
-model_metadata:
-  example_model_input:
-    inputs:
-    - Baseten is a fast inference provider
-    - Classify this separately.
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: BEI-qwen-qwen3-reranker-8b-fp8-truss-example
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-8B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
-
-```
-
-## Support
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/qwen/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml b/qwen/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml
deleted file mode 100644
index b00012f5c..000000000
--- a/qwen/BEI-qwen-qwen3-reranker-8b-fp8/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-model_metadata:
-  example_model_input:
-    inputs:
-    # advanced formatting of the string needed:
-    - "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n<Instruct>: {intruction}\n<Query>: {query}\n<Document>: {doc}<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
-    - "<|im_start|>system\nJudge whether the Document meets the requirements based on the Query and the Instruct provided. Note that the answer can only be \"yes\" or \"no\".<|im_end|>\n<|im_start|>user\n<Instruct>: Given a web search query, retrieve relevant passages that answer the query\n<Query>: What is the capital of China?\n<Document>: The capital of China is Beijing.<|im_end|>\n<|im_start|>assistant\n<think>\n\n</think>\n\n"
-    raw_scores: true
-    truncate: true
-    truncation_direction: Right
-model_name: Qwen3 Reranker 8B
-python_version: py39
-resources:
-  accelerator: H100_40GB
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    base_model: encoder
-    checkpoint_repository:
-      repo: michaelfeil/Qwen3-Reranker-8B-seq
-      revision: main
-      source: HF
-    max_num_tokens: 40960
-    num_builder_gpus: 1
-    quantization_type: fp8
-  runtime:
-    webserver_default_route: /predict
-  version_overrides:
-    bei_version: 0.0.25-b200-dev-v4
-    engine_builder_version: 0.20.0.dev1
diff --git a/qwen/engine-qwen-2-5-14b-coder-instruct/README.md b/qwen/engine-qwen-2-5-14b-coder-instruct/README.md
deleted file mode 100644
index dbea81234..000000000
--- a/qwen/engine-qwen-2-5-14b-coder-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen Coder 2.5 14B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen Coder 2.5 14B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-14b-coder-instruct/config.yaml b/qwen/engine-qwen-2-5-14b-coder-instruct/config.yaml
deleted file mode 100644
index 17395f233..000000000
--- a/qwen/engine-qwen-2-5-14b-coder-instruct/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
-          },
-          { role: "user", content: "Write a Python script for fizzbuzz." },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-Coder-14B-Instruct
-model_name: Qwen Coder 2.5 14B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-14B-Instruct
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-14b-instruct/README.md b/qwen/engine-qwen-2-5-14b-instruct/README.md
deleted file mode 100644
index b83ed1d76..000000000
--- a/qwen/engine-qwen-2-5-14b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen 2.5 14B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen 2.5 14B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-14b-instruct/config.yaml b/qwen/engine-qwen-2-5-14b-instruct/config.yaml
deleted file mode 100644
index 1bc50b8a6..000000000
--- a/qwen/engine-qwen-2-5-14b-instruct/config.yaml
+++ /dev/null
@@ -1,44 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    max_tokens: 512
-    messages:
-      - content: You are Qwen, created by Alibaba Cloud. You are a helpful assistant.
-        role: system
-      - content: What does Tongyi Qianwen mean?
-        role: user
-    stream: true
-    temperature: 0.9
-  repo_id: Qwen/Qwen2.5-14B-Instruct
-model_name: Qwen 2.5 14B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-14B-Instruct
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-32b-coder-instruct/README.md b/qwen/engine-qwen-2-5-32b-coder-instruct/README.md
deleted file mode 100644
index 962660240..000000000
--- a/qwen/engine-qwen-2-5-32b-coder-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen Coder 2.5 32B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen Coder 2.5 32B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-32b-coder-instruct/config.yaml b/qwen/engine-qwen-2-5-32b-coder-instruct/config.yaml
deleted file mode 100644
index 39f50a264..000000000
--- a/qwen/engine-qwen-2-5-32b-coder-instruct/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
-          },
-          { role: "user", content: "Write a Python script for fizzbuzz." },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-Coder-32B-Instruct
-model_name: Qwen Coder 2.5 32B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-32B-Instruct
-      source: HF
-    num_builder_gpus: 2
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-32b-instruct/README.md b/qwen/engine-qwen-2-5-32b-instruct/README.md
deleted file mode 100644
index bb04b18d1..000000000
--- a/qwen/engine-qwen-2-5-32b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen 2.5 32B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen 2.5 32B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-32b-instruct/config.yaml b/qwen/engine-qwen-2-5-32b-instruct/config.yaml
deleted file mode 100644
index 6536b7dec..000000000
--- a/qwen/engine-qwen-2-5-32b-instruct/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
-          },
-          { role: "user", content: "What does Tongyi Qianwen mean?" },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-32B-Instruct
-model_name: Qwen 2.5 32B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-32B-Instruct
-      source: HF
-    num_builder_gpus: 2
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-3b-instruct/README.md b/qwen/engine-qwen-2-5-3b-instruct/README.md
deleted file mode 100644
index 30b10008e..000000000
--- a/qwen/engine-qwen-2-5-3b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen 2.5 3B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen 2.5 3B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-3b-instruct/config.yaml b/qwen/engine-qwen-2-5-3b-instruct/config.yaml
deleted file mode 100644
index 5480b3954..000000000
--- a/qwen/engine-qwen-2-5-3b-instruct/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
-          },
-          { role: "user", content: "What does Tongyi Qianwen mean?" },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-3B-Instruct
-model_name: Qwen 2.5 3B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: A10G
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-3B-Instruct
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-72b-instruct/README.md b/qwen/engine-qwen-2-5-72b-instruct/README.md
deleted file mode 100644
index 9d5733ae4..000000000
--- a/qwen/engine-qwen-2-5-72b-instruct/README.md
+++ /dev/null
@@ -1,7 +0,0 @@
-# Qwen 2.5 72B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen 2.5 72B Instruct.
-
-Note that while other sizes of Qwen 2.5 are licensed as Apache 2.0, 72B sizes use the [qwen license](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE).
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-72b-instruct/config.yaml b/qwen/engine-qwen-2-5-72b-instruct/config.yaml
deleted file mode 100644
index a69157bd7..000000000
--- a/qwen/engine-qwen-2-5-72b-instruct/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
-          },
-          { role: "user", content: "What does Tongyi Qianwen mean?" },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-72B-Instruct
-model_name: Qwen 2.5 72B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:2
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-72B-Instruct
-      source: HF
-    num_builder_gpus: 4
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 2
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-72b-math-instruct/README.md b/qwen/engine-qwen-2-5-72b-math-instruct/README.md
deleted file mode 100644
index ce83e62b5..000000000
--- a/qwen/engine-qwen-2-5-72b-math-instruct/README.md
+++ /dev/null
@@ -1,7 +0,0 @@
-# Qwen Math 2.5 72B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen Math 2.5 72B Instruct.
-
-Note that while other sizes of Qwen 2.5 are licensed as Apache 2.0, 72B sizes use the [qwen license](https://huggingface.co/Qwen/Qwen2.5-72B-Instruct/blob/main/LICENSE).
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-72b-math-instruct/config.yaml b/qwen/engine-qwen-2-5-72b-math-instruct/config.yaml
deleted file mode 100644
index 9d7d59d95..000000000
--- a/qwen/engine-qwen-2-5-72b-math-instruct/config.yaml
+++ /dev/null
@@ -1,52 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "Please reason step by step, and put your final answer within \\boxed{}.",
-          },
-          {
-            role: "user",
-            content: "Find the value of $x$ that satisfies the equation $4x+5 = 6x+7$.",
-          },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-Math-72B-Instruct
-model_name: Qwen Math 2.5 72B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100:2
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Math-72B-Instruct
-      source: HF
-    num_builder_gpus: 4
-    quantization_type: fp8
-    max_seq_len: 32768
-    tensor_parallel_count: 2
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-7b-coder-instruct/README.md b/qwen/engine-qwen-2-5-7b-coder-instruct/README.md
deleted file mode 100644
index cac820a03..000000000
--- a/qwen/engine-qwen-2-5-7b-coder-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen Coder 2.5 7B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen Coder 2.5 7B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-7b-coder-instruct/config.yaml b/qwen/engine-qwen-2-5-7b-coder-instruct/config.yaml
deleted file mode 100644
index 19402b446..000000000
--- a/qwen/engine-qwen-2-5-7b-coder-instruct/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
-          },
-          { role: "user", content: "Write a Python script for fizzbuzz." },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-Coder-7B-Instruct
-model_name: Qwen Coder 2.5 7B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Coder-7B-Instruct
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-7b-instruct/README.md b/qwen/engine-qwen-2-5-7b-instruct/README.md
deleted file mode 100644
index 34d234473..000000000
--- a/qwen/engine-qwen-2-5-7b-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen 2.5 7B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen 2.5 7B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-7b-instruct/config.yaml b/qwen/engine-qwen-2-5-7b-instruct/config.yaml
deleted file mode 100644
index 08205932c..000000000
--- a/qwen/engine-qwen-2-5-7b-instruct/config.yaml
+++ /dev/null
@@ -1,49 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "You are Qwen, created by Alibaba Cloud. You are a helpful assistant.",
-          },
-          { role: "user", content: "What does Tongyi Qianwen mean?" },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-7B-Instruct
-model_name: Qwen 2.5 7B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-7B-Instruct
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-2-5-7b-math-instruct/README.md b/qwen/engine-qwen-2-5-7b-math-instruct/README.md
deleted file mode 100644
index 4ca182e64..000000000
--- a/qwen/engine-qwen-2-5-7b-math-instruct/README.md
+++ /dev/null
@@ -1,5 +0,0 @@
-# Qwen Math 2.5 7B Instruct Engine
-
-This example uses the [TensorRT-LLM Engine Builder for Qwen](https://docs.baseten.co/performance/examples/qwen-trt) to build and deploy an optimized inference engine for Qwen Math 2.5 7B Instruct.
-
-For advanced control over the engine building process, see [engine control in Python](https://docs.baseten.co/performance/engine-builder-customization) and [engine builder configuration](https://docs.baseten.co/performance/engine-builder-config) docs.
diff --git a/qwen/engine-qwen-2-5-7b-math-instruct/config.yaml b/qwen/engine-qwen-2-5-7b-math-instruct/config.yaml
deleted file mode 100644
index aa06ec39c..000000000
--- a/qwen/engine-qwen-2-5-7b-math-instruct/config.yaml
+++ /dev/null
@@ -1,52 +0,0 @@
-build_commands: []
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  tags:
-    - openai-compatible
-  example_model_input:
-    {
-      messages:
-        [
-          {
-            role: "system",
-            content: "Please reason step by step, and put your final answer within \\boxed{}.",
-          },
-          {
-            role: "user",
-            content: "Find the value of $x$ that satisfies the equation $4x+5 = 6x+7$.",
-          },
-        ],
-      stream: true,
-      max_tokens: 512,
-      temperature: 0.9,
-    }
-  repo_id: Qwen/Qwen2.5-Math-7B-Instruct
-model_name: Qwen Math 2.5 7B Instruct
-python_version: py39
-requirements: []
-resources:
-  accelerator: H100_40GB
-  cpu: "1"
-  memory: 24Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen2.5-Math-7B-Instruct
-      source: HF
-    num_builder_gpus: 1
-    quantization_type: no_quant
-    max_seq_len: 32768
-    tensor_parallel_count: 1
-    plugin_configuration:
-      use_paged_context_fmha: true
-      use_fp8_context_fmha: false
-      paged_kv_cache: true
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
-    request_default_max_tokens: 32768
diff --git a/qwen/engine-qwen-3-06b/config.yaml b/qwen/engine-qwen-3-06b/config.yaml
deleted file mode 100644
index f17889c21..000000000
--- a/qwen/engine-qwen-3-06b/config.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-      - content: Tell me everything you know about optimized inference.
-        role: user
-    stream: true
-    temperature: 0.5
-    chat_template_kwargs:
-      enable_thinking: true
-  tags:
-    - openai-compatible
-model_name: library-model-qwen3-06b-engine
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen3-0.6B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    num_builder_gpus: 1
-    max_seq_len: 40960
-    # plugin_configuration:
-    #   use_fp8_context_fmha: true
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 16
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/qwen/engine-qwen-3-30b-a3b-instruct-2507/config.yaml b/qwen/engine-qwen-3-30b-a3b-instruct-2507/config.yaml
deleted file mode 100644
index f014a7f84..000000000
--- a/qwen/engine-qwen-3-30b-a3b-instruct-2507/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-      - content: Tell me everything you know about optimized inference.
-        role: user
-    stream: true
-    temperature: 0.5
-    chat_template_kwargs:
-      enable_thinking: false
-  tags:
-    - openai-compatible
-model_name: qwen3-30b-a3b-instruct-2507-fp8_kv
-python_version: py39
-resources:
-  accelerator: B200
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
-trt_llm:
-  build:
-    base_model: decoder
-    checkpoint_repository:
-      repo: Qwen/Qwen3-30B-A3B-Instruct-2507
-      revision: main
-      source: HF
-    max_seq_len: 40960
-    num_builder_gpus: 4
-    plugin_configuration:
-      use_fp8_context_fmha: true
-    quantization_type: fp8_kv
-    tensor_parallel_count: 1
-  runtime:
-    batch_scheduler_policy: max_utilization
-    enable_chunked_context: true
diff --git a/qwen/engine-qwen-3-32b/config.yaml b/qwen/engine-qwen-3-32b/config.yaml
deleted file mode 100644
index 675f6ddbd..000000000
--- a/qwen/engine-qwen-3-32b/config.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-      - content: Tell me everything you know about optimized inference.
-        role: user
-    stream: true
-    temperature: 0.5
-    chat_template_kwargs:
-      enable_thinking: true
-  tags:
-    - openai-compatible
-model_name: library-model-qwen3-32B-engine
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen3-32B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    num_builder_gpus: 1
-    max_seq_len: 40960
-    # plugin_configuration:
-    #   use_fp8_context_fmha: true
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 16
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/qwen/engine-qwen-3-4b/config.yaml b/qwen/engine-qwen-3-4b/config.yaml
deleted file mode 100644
index a20f2c1db..000000000
--- a/qwen/engine-qwen-3-4b/config.yaml
+++ /dev/null
@@ -1,40 +0,0 @@
-model_metadata:
-  example_model_input:
-    max_tokens: 512
-    messages:
-      - content: Tell me everything you know about optimized inference.
-        role: user
-    stream: true
-    temperature: 0.5
-    chat_template_kwargs:
-      enable_thinking: true
-  tags:
-    - openai-compatible
-model_name: library-model-qwen3-4b-engine
-python_version: py39
-resources:
-  accelerator: H100
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: Qwen/Qwen3-4B
-      revision: main
-      source: HF
-    max_batch_size: 64
-    num_builder_gpus: 1
-    max_seq_len: 40960
-    # plugin_configuration:
-    #   use_fp8_context_fmha: true
-    quantization_type: fp8
-    speculator:
-      enable_b10_lookahead: true
-      lookahead_ngram_size: 16
-      lookahead_verification_set_size: 1
-      lookahead_windows_size: 1
-      speculative_decoding_mode: LOOKAHEAD_DECODING
-    tensor_parallel_count: 1
-  runtime:
-    enable_chunked_context: true
diff --git a/qwen/qwen-3-235B-A22B-instruct-2507-trt/config.yaml b/qwen/qwen-3-235B-A22B-instruct-2507-trt/config.yaml
deleted file mode 100644
index ef161a348..000000000
--- a/qwen/qwen-3-235B-A22B-instruct-2507-trt/config.yaml
+++ /dev/null
@@ -1,54 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What does Tongyi Qianwen mean?"
-    stream: false
-    model: "Qwen/Qwen3-235B-A22B-Instruct-2507-FP8"
-    max_tokens: 512
-    temperature: 0.6
-  tags:
-    - openai-compatible
-  repo_id: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
-model_name: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
-model_cache:
-  - repo_id: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
-    use_volume: true
-    revision: main
-    volume_folder: trt_model
-resources:
-  accelerator: H100:8
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-trt_llm:
-  build:
-    checkpoint_repository:
-      repo: michaelfeil/empty-model
-      revision: main
-      source: HF
-  inference_stack: v2
-  runtime:
-    enable_chunked_prefill: true
-    max_batch_size: 256
-    max_num_tokens: 8192
-    max_seq_len: 262144
-    served_model_name: Qwen/Qwen3-235B-A22B-Instruct-2507-FP8
-    tensor_parallel_size: 8
-    patch_kwargs:
-      disable_overlap_scheduler: True
-      model_path: /app/model_cache/trt_model
-      moe_expert_parallel_size: 8
-      cuda_graph_config:
-        enable_padding: true
-        max_batch_size: 256
-      enable_autotune: false
-      guided_decoding_backend: "xgrammar"
-      enable_iter_perf_stats: 0
-      kv_cache_config:
-        enable_block_reuse: true
-        free_gpu_memory_fraction: 0.8
-  version_overrides:
-    v2_llm_version: null
diff --git a/qwen/qwen-3-235B-sglang/config.yaml b/qwen/qwen-3-235B-sglang/config.yaml
deleted file mode 100644
index b78764125..000000000
--- a/qwen/qwen-3-235B-sglang/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What does Tongyi Qianwen mean?"
-    stream: true
-    model: "Qwen/Qwen3-235B-A22B"
-    max_tokens: 32768
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: Qwen 3 235B SGLang
-base_image:
-  image: lmsysorg/sglang:v0.4.6.post1-cu124
-model_cache:
-  - repo_id: Qwen/Qwen3-235B-A22B-FP8
-    revision: 57c8978fa7d601431cfd6750dd7355b5cdfa5a18
-    use_volume: true
-    volume_folder: "qwen3"
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-docker_server:
-  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3 --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-235B-A22B --tp 4 --reasoning-parser qwen3"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:4
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/qwen/qwen-3-30B-A3-coder/README.md b/qwen/qwen-3-30B-A3-coder/README.md
deleted file mode 100644
index f502daf2d..000000000
--- a/qwen/qwen-3-30B-A3-coder/README.md
+++ /dev/null
@@ -1,162 +0,0 @@
-# Qwen3-Coder-30B-A3B-Instruct Model
-
-This Truss serves the Qwen3-Coder-30B-A3B-Instruct model, a powerful coding-focused language model that excels at agentic coding tasks. The model is based on the [Qwen/Qwen3-Coder-30B-A3B-Instruct](https://huggingface.co/Qwen/Qwen3-Coder-30B-A3B-Instruct) model from Hugging Face and is optimized for high-performance coding assistance. It is Apache 2.0 licensed and can be used commercially without restrictions.
-
-## Model Description
-
-The Qwen3-Coder-30B-A3B-Instruct model is a specialized coding language model that features:
-
-- **Agentic Coding**: Excellent performance on agentic coding tasks and browser-use scenarios
-- **Long Context**: Native support for 256K tokens, extendable up to 1M tokens with Yarn
-- **Function Calling**: Specialized function call format for tool integration
-- **Repository-Scale Understanding**: Optimized for understanding large codebases
-- **Streaming Support**: Real-time code generation with streaming capabilities
-
-## Model Parameters
-
-The model accepts the following parameters:
-
-- `messages` (required): Array of message objects with role and content
-- `model` (optional): Model name (default: "Qwen/Qwen3-Coder-30B-A3B-Instruct")
-- `max_tokens` (optional): Maximum tokens to generate (default: 1024)
-- `temperature` (optional): Sampling temperature (default: 0.7)
-- `stream` (optional): Enable streaming response (default: true)
-- `tools` (optional): Array of function definitions for tool calling
-
-## Example Usage
-
-The model outputs structured responses compatible with OpenAI's chat completion format.
-
-```python
-import httpx
-import os
-
-# Replace with your model ID and API key
-model_id = "your-model-id"
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-
-# Example 1: Basic code generation
-basic_data = {
-    "messages": [
-        {"role": "system", "content": "You are a helpful coding assistant."},
-        {"role": "user", "content": "Write a quick sort algorithm in Python."}
-    ],
-    "max_tokens": 1024,
-    "temperature": 0.7,
-    "stream": True
-}
-
-# Example 2: Function calling for tool integration
-def square_the_number(num: float) -> dict:
-    return {"result": num ** 2}
-
-tools_data = {
-    "messages": [
-        {"role": "user", "content": "Calculate the square of 1024"}
-    ],
-    "tools": [
-        {
-            "type": "function",
-            "function": {
-                "name": "square_the_number",
-                "description": "Calculate the square of a number",
-                "parameters": {
-                    "type": "object",
-                    "required": ["num"],
-                    "properties": {
-                        "num": {
-                            "type": "number",
-                            "description": "The number to square"
-                        }
-                    }
-                }
-            }
-        }
-    ],
-    "max_tokens": 1024,
-    "temperature": 0.7
-}
-
-# Call the model
-print("Generating code...")
-response = httpx.post(
-    f"https://model-{model_id}.api.baseten.co/development/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=basic_data,
-    timeout=httpx.Timeout(60.0)
-)
-
-# Get the result
-result = response.json()
-print("Generated code:", result.get("choices", [{}])[0].get("message", {}).get("content", ""))
-```
-
-## Agentic Coding Examples
-
-The model excels at agentic coding tasks. Here are some example use cases:
-
-```python
-# Repository analysis
-repo_analysis = {
-    "messages": [
-        {"role": "user", "content": "Analyze this codebase and suggest improvements for the authentication system."}
-    ],
-    "max_tokens": 2048
-}
-
-# Code review
-code_review = {
-    "messages": [
-        {"role": "user", "content": "Review this Python function for security vulnerabilities:\n\ndef process_user_input(data):\n    return eval(data)"}
-    ],
-    "max_tokens": 1024
-}
-
-# Debugging assistance
-debugging = {
-    "messages": [
-        {"role": "user", "content": "Help me debug this error: 'TypeError: 'NoneType' object is not callable'"}
-    ],
-    "max_tokens": 1024
-}
-```
-
-## Best Practices
-
-For optimal performance, we recommend:
-
-1. **Sampling Parameters**:
-   - Temperature: 0.7
-   - Top-p: 0.8
-   - Top-k: 20
-   - Repetition penalty: 1.05
-
-2. **Context Length**: Use up to 65,536 tokens for most queries
-
-3. **Streaming**: Enable streaming for real-time code generation
-
-4. **Function Calling**: Define clear tool schemas for agentic tasks
-
-## Deployment
-
-To deploy this model:
-
-1. Clone the repository
-2. Make sure you have the Truss CLI installed (`pip install truss`)
-3. Run the deployment command:
-
-```bash
-truss push qwen/qwen-3-30B-A3-coder --publish
-```
-
-## Model Features
-
-- **OpenAI-Compatible API**: Full compatibility with OpenAI's chat completion format
-- **Streaming Support**: Real-time response streaming for better user experience
-- **Tool Calling**: Native support for function calling and tool integration
-- **Long Context**: Handles large codebases and documentation
-- **GPU Optimization**: Optimized for H100 GPUs with SGLang
-
-## License
-
-This model is licensed under Apache 2.0 and can be used commercially without restrictions.
diff --git a/qwen/qwen-3-30B-A3-coder/config.yaml b/qwen/qwen-3-30B-A3-coder/config.yaml
deleted file mode 100644
index b6c5c228a..000000000
--- a/qwen/qwen-3-30B-A3-coder/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write a quick sort algorithm."
-    stream: true
-    model: "Qwen/Qwen3-Coder-30B-A3B-Instruct"
-    max_tokens: 1024
-    temperature: 0.7
-  tags:
-    - openai-compatible
-model_name: Qwen 3 Coder
-base_image:
-  image: lmsysorg/sglang:v0.4.10.post2-cu126
-model_cache:
-  - repo_id: Qwen/Qwen3-Coder-30B-A3B-Instruct
-    revision: main
-    use_volume: true
-    volume_folder: "qwen3-coder"
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-docker_server:
-  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3-coder --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-Coder-30B-A3B-Instruct --tp 1 --reasoning-parser qwen3"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/qwen/qwen-3-30B-A3-sglang/config.yaml b/qwen/qwen-3-30B-A3-sglang/config.yaml
deleted file mode 100644
index 05cc6157c..000000000
--- a/qwen/qwen-3-30B-A3-sglang/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What does Tongyi Qianwen mean?"
-    stream: false
-    model: "Qwen/Qwen3-32B"
-    max_tokens: 512
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: Qwen 3 30B-A3 SGLang
-environment_variables:
-  hf_access_token: null
-base_image:
-  image: lmsysorg/sglang:v0.4.6.post1-cu124
-model_cache:
-  - repo_id: Qwen/Qwen3-30B-A3B-FP8
-    revision: 2daf1706ac267bae18c90a217a060817c0cebb66
-    use_volume: true
-    volume_folder: "qwen3"
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-docker_server:
-  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3 --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-30B-A3B --tp 1 --reasoning-parser qwen3"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/qwen/qwen-3-30B-A3-vllm/config.yaml b/qwen/qwen-3-30B-A3-vllm/config.yaml
deleted file mode 100644
index e7f64104b..000000000
--- a/qwen/qwen-3-30B-A3-vllm/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.8.5
-docker_server:
-  start_command: sh -c "vllm serve Qwen/Qwen3-30B-A3B --enable-reasoning --reasoning-parser deepseek_r1 --served-model-name qwen30b --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-model_metadata:
-  repo_id: Qwen/Qwen3-30B-A3B
-  example_model_input:
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What does Tongyi Qianwen mean?"
-    stream: false
-    model: "qwen30b"
-    max_tokens: 512
-    temperature: 0.7
-  tags:
-    - openai-compatible
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
-model_name: Qwen 3 30B-A3B vLLM
-environment_variables:
-  VLLM_LOGGING_LEVEL: WARNING
diff --git a/qwen/qwen-3-30b-omni-thinker/README.md b/qwen/qwen-3-30b-omni-thinker/README.md
deleted file mode 100644
index 18f0eacf0..000000000
--- a/qwen/qwen-3-30b-omni-thinker/README.md
+++ /dev/null
@@ -1,256 +0,0 @@
-# Qwen3-Omni-30B-Thinker (vLLM) on Baseten
-
-This is a [Truss](https://truss.baseten.co/) for serving the "thinker" part of **Qwen/Qwen3-Omni-30B-A3B-Intruct** with **vLLM** on Baseten. It exposes an **OpenAI-compatible /v1/chat/completions** endpoint that accepts **text, image, audio, and video** inputs and returns low-latency text responses.
-
-**Why this deployment**
-
-* *Multimodal, end-to-end*: text, images, audio, and video inputs in a single chat request; text responses today (audio TTS output is not compatible with vLLM at the moment).
-* *OpenAI-compatible API*: drop-in with `openai` libraries and many tools.
-* *Production-ready on Baseten*: autoscaling, logs, metrics, and zero-downtime publishes.
-
----
-
-# Overview
-
-**Model**: `Qwen/Qwen3-Omni-30B-A3B-Instruct`
-**Modalities**:
-
-* **Input**: text, image(s), audio, video
-* **Output**: text
-
----
-
-# Deploy with Truss
-
-Before deployment:
-
-1. Create a [Baseten account](https://app.baseten.co/signup) and an [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install Truss: `pip install --upgrade truss`
-
-Clone the examples repo (or your project) and cd into your working directory:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd qwen/qwen-3-30b-omni-thinker
-```
-
-Publish:
-
-```sh
-truss push --publish
-# ✨ Model Qwen3 Omni 30B Instruct (Thinker Only) was successfully pushed ✨
-```
-
----
-
-# Call your model
-
-Your deployment is OpenAI-compatible. Replace `model-xxxxxx` and include your Baseten API key.
-
-## API schema (Chat Completions)
-
-**POST** `https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/chat/completions`
-
-**Request (multimodal example)**:
-
-```json
-{
-  "model": "qwen3-omni",
-  "messages": [
-    {"role": "system", "content": "You are a helpful assistant."},
-    {
-      "role": "user",
-      "content": [
-        {"type": "text", "text": "Describe what you see and hear."},
-        {
-          "type": "image_url",
-          "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}
-        },
-        {
-          "type": "audio_url",
-          "audio_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}
-        }
-      ]
-    }
-  ],
-  "max_tokens": 2048,
-  "temperature": 0.7,
-  "stream": false
-}
-```
-
-**Response (truncated example)**:
-
-```json
-{
-  "id": "chatcmpl-...",
-  "object": "chat.completion",
-  "created": 1710000000,
-  "model": "qwen3-omni",
-  "choices": [
-    {
-      "index": 0,
-      "finish_reason": "stop",
-      "message": {
-        "role": "assistant",
-        "content": "I see several parked cars in front of a building and hear a short cough."
-      }
-    }
-  ],
-  "usage": {
-    "prompt_tokens": 512,
-    "completion_tokens": 24,
-    "total_tokens": 536
-  }
-}
-```
-
-### curl
-
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/chat/completions \
-  -H "Authorization: Api-Key YOUR_API_KEY" \
-  -H "Content-Type: application/json" \
-  -d '{
-    "model":"qwen3-omni",
-    "messages":[
-      {"role":"system","content":"You are a helpful assistant."},
-      {"role":"user","content":[
-        {"type":"text","text":"Describe this image and audio content."},
-        {"type":"image_url","image_url":{"url":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}},
-        {"type":"audio_url","audio_url":{"url":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}}
-      ]}
-    ],
-    "max_tokens":2048,
-    "temperature":0.7,
-    "stream":false
-  }'
-```
-
-### OpenAI Python SDK
-
-```python
-from openai import OpenAI
-import os
-
-client = OpenAI(
-    api_key=os.environ["BASETEN_API_KEY"],
-    base_url="https://model-xxxxxx.api.baseten.co/environments/production/sync/v1"
-)
-
-resp = client.chat.completions.create(
-    model="qwen3-omni",
-    messages=[
-        {"role": "system", "content": "You are a helpful assistant."},
-        {"role": "user", "content": [
-            {"type": "text", "text": "Describe this image and audio content."},
-            {"type": "image_url", "image_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}},
-            {"type": "audio_url", "audio_url": {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}}
-        ]}
-    ],
-    max_tokens=2048,
-    temperature=0.7,
-    stream=False,
-)
-print(resp.choices[0].message.content)
-```
-
-### Streaming (server-sent events)
-
-```python
-stream = client.chat.completions.create(
-    model="qwen3-omni",
-    messages=[{"role":"user","content":"Summarize this short clip."}],
-    stream=True
-)
-for event in stream:
-    if event.choices and event.choices[0].delta:
-        print(event.choices[0].delta.content or "", end="", flush=True)
-```
-
-### requests (Python)
-
-```python
-import os, requests, json
-
-url = "https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/chat/completions"
-headers = {
-    "Authorization": "Api-Key " + os.environ["BASETEN_API_KEY"],
-    "Content-Type": "application/json"
-}
-payload = {
-    "model": "qwen3-omni",
-    "messages": [{"role":"user","content":[{"type":"text","text":"Briefly describe the image."},
-        {"type":"image_url","image_url":{"url":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}}]}],
-    "max_tokens": 512
-}
-print(requests.post(url, headers=headers, data=json.dumps(payload)).json())
-```
-
-> The OpenAPI schema is available at:
-> `https://model-xxxxxx.api.baseten.co/environments/production/sync/openapi.json`
-
----
-
-# Advanced usage notes
-
-* **Images, audio, video**: Provide remote URLs via `image_url` / `audio_url` / `video_url`, or configure `--allowed-local-media-path` to allow local file ingestion.
-* **Batching/throughput**: Configure Baseten `predict_concurrency` for request-level concurrency. vLLM will also batch internally.
-* **Long contexts**: `--max-model-len 65,536` in the server command; adjust based on memory.
-* **Multi-GPU**: Add `-tp N` (tensor parallelism) in the vLLM command to shard across GPUs.
-* **Audio output**: If you require generated speech, verify your vLLM build and model variant support it; otherwise pipe text into a TTS stage.
-
----
-
-**Notes**
-
-* The container image `qwenllm/qwen3-omni:3-cu124` bundles vLLM and dependencies.
-* For **multi-GPU** boxes, add `-tp <num_gpus>` to the `vllm serve` command.
-
----
-
-# Examples
-
-### 1) Pure text
-
-```json
-{
-  "model":"qwen3-omni",
-  "messages":[{"role":"user","content":"Give me three creative app ideas for teachers."}],
-  "max_tokens":512
-}
-```
-
-### 2) Image + instruction
-
-```json
-{
-  "model":"qwen3-omni",
-  "messages":[
-    {"role":"user","content":[
-      {"type":"image_url","image_url":{"url":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}},
-      {"type":"text","text":"Describe the scene in one sentence."}
-    ]}
-  ]
-}
-```
-
-### 3) Audio + question
-
-```json
-{
-  "model":"qwen3-omni",
-  "messages":[
-    {"role":"user","content":[
-      {"type":"audio_url","audio_url":{"url":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}},
-      {"type":"text","text":"What do you hear?"}
-    ]}
-  ]
-}
-```
-
----
-
-# Support
-
-If you have questions or need help, open an issue in this repository or contact Baseten support.
diff --git a/qwen/qwen-3-30b-omni-thinker/config.yaml b/qwen/qwen-3-30b-omni-thinker/config.yaml
deleted file mode 100644
index 81cc44e0d..000000000
--- a/qwen/qwen-3-30b-omni-thinker/config.yaml
+++ /dev/null
@@ -1,41 +0,0 @@
-model_name: Qwen3 Omni 30B Instruct (Thinker Only)
-base_image:
-  image: qwenllm/qwen3-omni:3-cu124
-docker_server:
-  start_command: |
-    sh -c "vllm serve Qwen/Qwen3-Omni-30B-A3B-Instruct --dtype bfloat16 --max-model-len 65536 --served-model-name qwen3-omni"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-model_metadata:
-  example_model_input:
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content:
-          - type: text
-            text: "Describe this image and audio content."
-          - type: image_url
-            image_url:
-              url: "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"
-          - type: audio_url
-            audio_url:
-              url: "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"
-          - type: text
-            text: "What can you see and hear? Answer in one sentence."
-
-    stream: false
-    model: "qwen3-omni"
-    max_tokens: 2048
-    temperature: 0.7
-  tags:
-    - openai-compatible
-    - multimodal
-    - image-processing
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/qwen/qwen-3-30b-omni/README.md b/qwen/qwen-3-30b-omni/README.md
deleted file mode 100644
index 5d3726dfd..000000000
--- a/qwen/qwen-3-30b-omni/README.md
+++ /dev/null
@@ -1,173 +0,0 @@
-# Qwen3-Omni-30B-Thinker (vLLM) on Baseten
-
-This is a [Truss](https://truss.baseten.co/) for serving **Qwen/Qwen3-Omni-30B-A3B-Intruct** with **transformers** on Baseten. It exposes an endpoint that accepts **text, image, audio, and video** inputs and returns low-latency text and audio responses.
-
-**Why this deployment**
-
-* *Multimodal, end-to-end*: text, images, audio, and video inputs in a single chat request; text and spoken audio responses.
-* *Production-ready on Baseten*: autoscaling, logs, metrics, and zero-downtime publishes.
-
----
-
-# Overview
-
-**Model**: `Qwen/Qwen3-Omni-30B-A3B-Instruct`
-**Modalities**:
-
-* **Input**: text, image(s), audio, video
-* **Output**: text, audio
-
----
-
-# Deploy with Truss
-
-Before deployment:
-
-1. Create a [Baseten account](https://app.baseten.co/signup) and an [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install Truss: `pip install --upgrade truss`
-
-Clone the examples repo (or your project) and cd into your working directory:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd qwen/qwen-3-30b-omni-thinker
-```
-
-Publish:
-
-```sh
-truss push --publish
-# ✨ Model Qwen3 Omni 30B Instruct was successfully pushed ✨
-```
-
----
-
-# Call your model
-
-Replace `model-xxxxxx` and include your Baseten API key.
-
-## API schema (Chat Completions)
-
-**POST** `https://model-xxxxxx.api.baseten.co/development/predict`
-
-**Request example**:
-
-```json
-{
-  "speaker": "Chelsie",
-  "messages": [
-    {"role": "system", "content": "You are a helpful assistant."},
-    {
-      "role": "user",
-      "content": [
-        {"type": "text", "text": "Hello how are you?"}
-      ]
-    }
-  ]
-}
-```
-
-**Response (truncated example)**:
-
-```json
-{
-  "text": "Hello! I'm doing well, thank you. How can I assist you today?",
-  "audio": "UklGRs5hAwBXQVZFZm10IBAAAAABAAEAwF0AAIC7AAACABAAZGF0YaphAwAE..."
-}
-```
-
-### curl
-
-```bash
-curl -X POST https://model-xxxxxx.api.baseten.co/development/predict \
-  -H "Authorization: Api-Key YOUR_API_KEY" \
-  -H "Content-Type: application/json" \
-  -d '{
-    "speaker":"Aiden",
-    "messages":[
-      {"role":"system","content":"You are a helpful assistant."},
-      {"role":"user","content":[
-        {"type":"text","text":"Describe this image and audio content."},
-        {"type":"image","image":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"},
-        {"type":"audio","audio":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"}
-      ]}
-    ]
-  }'
-```
-
-### requests (Python)
-
-```python
-import os, requests, json, base64
-
-url = "https://model-xxxxxx.api.baseten.co/development/predict"
-headers = {
-    "Authorization": "Api-Key " + os.environ["BASETEN_API_KEY"],
-    "Content-Type": "application/json"
-}
-payload = {
-    "speaker": "Ethan",
-    "messages": [{"role":"user","content":[{"type":"text","text":"Briefly describe the image."},
-        {"type":"image","image":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"}]}]
-}
-result = requests.post(url, headers=headers, data=json.dumps(payload)).json()
-
-print(result["text"])
-
-with open("result.wav", "wb") as f:
-  f.write(base64.b64decode(result["audio"]))
-```
-
----
-
-**Notes**
-
-* The container image `qwenllm/qwen3-omni:3-cu124` bundles flash attention and dependencies.
-
----
-
-# Examples
-
-### 1) Pure text
-
-```json
-{
-  "model":"qwen3-omni",
-  "messages":[{"role":"user","content":"Give me three creative app ideas for teachers."}],
-  "max_tokens":512
-}
-```
-
-### 2) Image + instruction
-
-```json
-{
-  "model":"qwen3-omni",
-  "messages":[
-    {"role":"user","content":[
-      {"type":"image","image":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cars.jpg"},
-      {"type":"text","text":"Describe the scene in one sentence."}
-    ]}
-  ]
-}
-```
-
-### 3) Audio + question
-
-```json
-{
-  "model":"qwen3-omni",
-  "messages":[
-    {"role":"user","content":[
-      {"type":"audio","audio":"https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-Omni/demo/cough.wav"},
-      {"type":"text","text":"What do you hear?"}
-    ]}
-  ]
-}
-```
-
----
-
-# Support
-
-If you have questions or need help, open an issue in this repository or contact Baseten support.
diff --git a/qwen/qwen-3-30b-omni/config.yaml b/qwen/qwen-3-30b-omni/config.yaml
deleted file mode 100644
index af7b83393..000000000
--- a/qwen/qwen-3-30b-omni/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-model_name: Qwen3 Omni 30B Instruct
-base_image:
-  image: qwenllm/qwen3-omni
-model_metadata:
-  example_model_input: {
-    "speaker": "Chelsie",
-    "messages": [
-        {
-            "role": "user",
-            "content": [
-                {
-                    "type": "text",
-                    "text": "Hi, how are you?"
-                }
-            ]
-        }
-    ]
-  }
-runtime:
-  predict_concurrency : 1
-resources:
-  accelerator: H100
-  use_gpu: true
-environment_variables:
-  VLLM_LOGGING_LEVEL: INFO
diff --git a/qwen/qwen-3-32B-sglang/config.yaml b/qwen/qwen-3-32B-sglang/config.yaml
deleted file mode 100644
index 8e1e22488..000000000
--- a/qwen/qwen-3-32B-sglang/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What does Tongyi Qianwen mean?"
-    stream: true
-    model: "Qwen/Qwen3-32B"
-    max_tokens: 32768
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: Qwen 3 32B SGLang
-base_image:
-  image: lmsysorg/sglang:v0.4.6.post1-cu124
-model_cache:
-  - repo_id: Qwen/Qwen3-32B-FP8
-    revision: 37f3f67a7a82b002377985796c57f4321b85fb9a
-    use_volume: true
-    volume_folder: "qwen3"
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-docker_server:
-  start_command: sh -c "truss-transfer-cli && python3 -m sglang.launch_server --model-path /app/model_cache/qwen3 --host 0.0.0.0 --port 8000 --served-model-name Qwen/Qwen3-32B --tp 1 --reasoning-parser qwen3"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/qwen/qwen-3-asr/README.md b/qwen/qwen-3-asr/README.md
deleted file mode 100644
index af3a581d5..000000000
--- a/qwen/qwen-3-asr/README.md
+++ /dev/null
@@ -1,51 +0,0 @@
-# Qwen3 ASR 1.7B
-
-This example shows how to call a Baseten deployment using the OpenAI Python SDK to run **Qwen/Qwen3-ASR-1.7B** on an audio URL.
-
-## Prerequisites
-
-- Python 3.9+
-- OpenAI Python SDK installed:
-
-```bash
-pip install openai
-```
-
-## Example: Transcribe an audio URL
-
-```python
-from openai import OpenAI
-
-model_id = ""  # place model ID here
-
-client = OpenAI(
-    api_key="BASETEN-API-KEY",
-    base_url=f"https://model-{model_id}.api.baseten.co/environments/production/sync/v1"
-)
-
-response = client.chat.completions.create(
-    model="Qwen/Qwen3-ASR-1.7B",
-    stream=False,
-    messages=[
-        {
-            "role": "user",
-            "content": [
-                {
-                    "type": "audio_url",
-                    "audio_url":
-                        {"url": "https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav"}
-
-                }
-            ]
-        }
-    ],
-)
-
-print(response.choices[0].message.content)
-```
-
-## Sample Output
-```txt
-Response: language English<asr_text>Uh huh. Oh yeah, yeah. He wasn't even that big when I started listening to him, but and his solo music didn't do overly well, but he did very well when he started writing for other people.
-
-```
diff --git a/qwen/qwen-3-asr/config.yaml b/qwen/qwen-3-asr/config.yaml
deleted file mode 100644
index a86c79a4f..000000000
--- a/qwen/qwen-3-asr/config.yaml
+++ /dev/null
@@ -1,42 +0,0 @@
-model_metadata:
-  example_model_input:
-    stream: false
-    messages:
-      - role: user
-        content:
-          - type: audio_url
-            audio_url:
-              url: https://qianwen-res.oss-cn-beijing.aliyuncs.com/Qwen3-ASR-Repo/asr_en.wav
-  tags:
-    - openai-compatible
-model_name: Qwen3-ASR-1.7B
-secrets:
-  hf_access_token: null
-base_image:
-  image:  vllm/vllm-openai:nightly-070c811d6f74c55302557878f5982411a3346b4d
-docker_server:
-  start_command: sh -c "HF_TOKEN=$(cat /secrets/hf_access_token) vllm serve Qwen/Qwen3-ASR-1.7B --gpu-memory-utilization 0.8 --host 0.0.0.0 --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100_40GB:1
-  cpu: "1"
-  memory: 10Gi
-  use_gpu: true
-requirements:
-  - --pre --extra-index-url https://wheels.vllm.ai/nightly
-  - vllm[audio]
-  - librosa
-  - torch
-  - torchaudio
-  - pynvml
-  - ffmpeg-python
-system_packages:
-  - python3.10-venv
-  - ffmpeg
-  - openmpi-bin
-  - libopenmpi-dev
-runtime:
-  predict_concurrency: 256
diff --git a/qwen/qwen-3-next-80B-A3-instruct-sglang/config.yaml b/qwen/qwen-3-next-80B-A3-instruct-sglang/config.yaml
deleted file mode 100644
index 5c67bb964..000000000
--- a/qwen/qwen-3-next-80B-A3-instruct-sglang/config.yaml
+++ /dev/null
@@ -1,38 +0,0 @@
-base_image:
-  #image: lmsysorg/sglang@sha256:c977d3c5cf66029c8c37436a777381cb5bc861527da1b405c90ae2360417eedb
-  image: lmsysorg/sglang:v0.5.3rc1-cu126
-# build_commands:
-#   - pip install --upgrade pip
-#   - pip uninstall -y sglang
-#   - git clone https://github.com/sgl-project/sglang.git && cd sglang && pip install -e "python[all]"
-model_metadata:
-  repo_id: Qwen/Qwen3-Next-80B-A3B-Instruct
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write FizzBuzz in Python"
-    stream: true
-    model: "Qwen/Qwen3-Next-80B-A3B-Instruct"
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: sh -c "SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python3 -m sglang.launch_server --model-path Qwen/Qwen3-Next-80B-A3B-Instruct-FP8 --revision c5f5f263bdd5cc134092897864e8905d8fe7b928 --tp-size 2 --context-length 262144 --mem-fraction-static 0.8 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --served-model-name Qwen/Qwen3-Next-80B-A3B-Instruct --tool-call-parser qwen25 --host 0.0.0.0 --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-model_cache:
-  - repo_id: Qwen/Qwen3-Next-80B-A3B-Instruct-FP8
-    revision: c5f5f263bdd5cc134092897864e8905d8fe7b928
-    use_volume: true
-    volume_folder: qwen
-model_name: Qwen3-Next-80B-A3B-Instruct
diff --git a/qwen/qwen-3-next-80B-A3-thinking-sglang/config.yaml b/qwen/qwen-3-next-80B-A3-thinking-sglang/config.yaml
deleted file mode 100644
index e27892840..000000000
--- a/qwen/qwen-3-next-80B-A3-thinking-sglang/config.yaml
+++ /dev/null
@@ -1,38 +0,0 @@
-base_image:
-  #image: lmsysorg/sglang@sha256:c977d3c5cf66029c8c37436a777381cb5bc861527da1b405c90ae2360417eedb
-  image: lmsysorg/sglang:v0.5.3rc1-cu126
-# build_commands:
-#   - pip install --upgrade pip
-#   - pip uninstall -y sglang
-#   - git clone https://github.com/sgl-project/sglang.git && cd sglang && pip install -e "python[all]"
-model_metadata:
-  repo_id: Qwen/Qwen3-Next-80B-A3B-Thinking-FP8
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write FizzBuzz in Python"
-    stream: true
-    model: "Qwen/Qwen3-Next-80B-A3B-Thinking"
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: sh -c "SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1 python3 -m sglang.launch_server --model-path Qwen/Qwen3-Next-80B-A3B-Thinking-FP8 --revision 1a28d48a94e799860201879be67616b9e21c4bd2 --reasoning-parser qwen3-thinking --tp-size 2 --context-length 262144 --mem-fraction-static 0.8 --speculative-algo NEXTN --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --served-model-name Qwen/Qwen3-Next-80B-A3B-Thinking --tool-call-parser qwen25 --host 0.0.0.0 --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-model_cache:
-  - repo_id: Qwen/Qwen3-Next-80B-A3B-Thinking-FP8
-    revision: 1a28d48a94e799860201879be67616b9e21c4bd2
-    use_volume: true
-    volume_folder: qwen
-model_name: Qwen3-Next-80B-A3B-Thinking
diff --git a/qwen/qwen-3-vl-30b-a3b-instruct/config.yaml b/qwen/qwen-3-vl-30b-a3b-instruct/config.yaml
deleted file mode 100644
index 87335d34c..000000000
--- a/qwen/qwen-3-vl-30b-a3b-instruct/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    model: "Qwen/Qwen3-VL-30B-A3B-Thinking"
-    stream: false
-    max_tokens: 4096
-    messages:
-      - role: user
-        content:
-          - type: text
-            text: "What's in this image?"
-          - type: image_url
-            image_url:
-              url: "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true"
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: Qwen3-VL-30B-A3B-Instruct-FP8
-base_image:
-  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:2f7dbc9b42c51ba192e3dded515e4e07cdfdabea
-build_commands:
-  - pip install --pre --upgrade transformers
-  - pip uninstall -y vllm
-  - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git@d3d649efec8161b62e8db576f8d1d02a77d22897
-docker_server:
-  start_command: python3 -m vllm.entrypoints.openai.api_server --model Qwen/Qwen3-VL-30B-A3B-Instruct-FP8 --tool-call-parser hermes --reasoning-parser qwen3 --served-model-name Qwen/Qwen3-VL-30B-A3B-Instruct --enable-expert-parallel --enable-auto-tool-choice --tensor-parallel-size 2 --host 0.0.0.0 --port 8000
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/qwen/qwen-3-vl-30b-a3b-thinking/config.yaml b/qwen/qwen-3-vl-30b-a3b-thinking/config.yaml
deleted file mode 100644
index 8d2bf99fe..000000000
--- a/qwen/qwen-3-vl-30b-a3b-thinking/config.yaml
+++ /dev/null
@@ -1,34 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    model: "Qwen/Qwen3-VL-30B-A3B-Thinking"
-    stream: false
-    max_tokens: 4096
-    messages:
-      - role: user
-        content:
-          - type: text
-            text: "What's in this image?"
-          - type: image_url
-            image_url:
-              url: "https://github.com/sgl-project/sglang/blob/main/test/lang/example_image.png?raw=true"
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: Qwen3-VL-30B-A3B-Thinking-FP8
-base_image:
-  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:2f7dbc9b42c51ba192e3dded515e4e07cdfdabea
-build_commands:
-  - pip install --pre --upgrade transformers
-  - pip uninstall -y vllm
-  - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git@d3d649efec8161b62e8db576f8d1d02a77d22897
-docker_server:
-  start_command: python3 -m vllm.entrypoints.openai.api_server --model Qwen/Qwen3-VL-30B-A3B-Thinking-FP8 --tool-call-parser hermes --reasoning-parser qwen3 --served-model-name Qwen/Qwen3-VL-30B-A3B-Thinking --enable-expert-parallel --enable-auto-tool-choice --tensor-parallel-size 2 --host 0.0.0.0 --port 8000
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/qwen/qwen-3-vl-32b/config.yaml b/qwen/qwen-3-vl-32b/config.yaml
deleted file mode 100644
index cd7cfcd2a..000000000
--- a/qwen/qwen-3-vl-32b/config.yaml
+++ /dev/null
@@ -1,44 +0,0 @@
-base_image:
-  image: vllm/vllm-openai:v0.11.0
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    model: ""
-    messages:
-      - role: user
-        content:
-          - type: image_url
-            image_url:
-              url: "https://upload.wikimedia.org/wikipedia/commons/f/fa/Grayscale_8bits_palette_sample_image.png"
-          - type: text
-            text: "Describe this image in detail."
-    stream: true
-  tags:
-    - openai-compatible
-model_name: Qwen 3 VL 32B
-requirements:
-  - transformers>=4.55.0
-  - accelerate
-  - timm
-  - einops
-  - open-clip-torch
-  - pillow
-python_version: py312
-model_cache:
-  - repo_id: Qwen/Qwen3-VL-32B-Instruct-FP8
-    revision: main
-    use_volume: true
-    volume_folder: "qwen-3-vl-32b"
-    ignore_patterns:
-      - "original/*"
-      - "*.pth"
-docker_server:
-  start_command: vllm serve Qwen/Qwen3-VL-32B-Instruct-FP8 --tensor-parallel-size 1 --served-model-name qwen-3-vl-32b --trust-remote-code --max-model-len 16384 --gpu-memory-utilization 0.9
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:1
-  use_gpu: true
-runtime:
-  predict_concurrency: 16
diff --git a/qwen/qwen-coder-next/config.yaml b/qwen/qwen-coder-next/config.yaml
deleted file mode 100644
index 92126518b..000000000
--- a/qwen/qwen-coder-next/config.yaml
+++ /dev/null
@@ -1,39 +0,0 @@
-model_metadata:
-  example_model_input:
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What is the meaning of life?"
-    stream: true
-    model: Qwen/Qwen3-Coder-Next
-    max_tokens: 32768
-    temperature: 0.7
-  tags:
-    - openai-compatible
-
-base_image:
-  image: lmsysorg/sglang:nightly-dev-20260202-9227d4f7
-
-
-build_commands:
-  - pip uninstall -y sglang
-  - git clone https://github.com/sgl-project/sglang.git && cd sglang && pip install --upgrade pip && pip install -e "python" && pip install nvidia-cudnn-cu12>=9.16.0.29
-
-docker_server:
-  start_command: sh -c 'python -m sglang.launch_server --model Qwen/Qwen3-Coder-Next-FP8 --port 30000 --tp-size 2 --tool-call-parser qwen3_coder'
-  readiness_endpoint: /health_generate
-  liveness_endpoint: /health_generate
-  predict_endpoint: /v1/chat/completions
-  server_port: 30000
-requirements: []
-system_packages:
-- tmux
-- htop
-- nload
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency : 64
-model_name: Qwen3-Coder-Next
diff --git a/qwen/qwen-image/README.md b/qwen/qwen-image/README.md
deleted file mode 100644
index 2b6d62909..000000000
--- a/qwen/qwen-image/README.md
+++ /dev/null
@@ -1,129 +0,0 @@
-# Qwen Image Model
-
-![Qwen Image on Baseten](assets/generated_image.jpg)
-
-This Truss serves the Qwen Image model, a powerful text-to-image generation model that supports both English and Chinese prompts. The model is based on the Qwen/Qwen-Image model from Hugging Face and is optimized for high-quality image generation. It is Apache 2.0 licensed and can be used commercially without restrictions.
-
-## Model Description
-
-The Qwen Image model is a diffusion-based text-to-image model that can generate high-quality images from text prompts. It features:
-
-- **Multilingual Support**: Handles both English and Chinese prompts
-- **High Quality**: Generates 4K quality images with cinematic composition
-- **Flexible Aspect Ratios**: Supports various image dimensions
-- **Customizable Parameters**: Adjustable inference steps, guidance scale, and more
-
-## Model Parameters
-
-The model accepts the following parameters:
-
-- `prompt` (required): The text description of the image you want to generate
-- `negative_prompt` (optional): Text describing what you don't want in the image (default: "")
-- `width` (optional): Image width in pixels (default: 1024)
-- `height` (optional): Image height in pixels (default: 1024)
-- `num_inference_steps` (optional): Number of denoising steps (default: 50)
-- `true_cfg_scale` (optional): Guidance scale for generation (default: 4.0)
-- `seed` (optional): Random seed for reproducible results (default: random)
-
-## Example Usage
-
-The model outputs a base64 string which can be saved locally.
-
-```python
-import httpx
-import os
-import base64
-from PIL import Image
-from io import BytesIO
-
-# Replace with your model ID and API key
-model_id = "your-model-id"
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-
-def b64_to_pil(b64_str):
-    """Convert base64 string to PIL image"""
-    return Image.open(BytesIO(base64.b64decode(b64_str)))
-
-# Example 1: English prompt
-english_data = {
-    "prompt": "A fashionably dressed man on the streets of New York City holds a sign that says `Qwen Image on Baseten`",
-    "width": 1664,
-    "height": 928,
-    "num_inference_steps": 50,
-    "seed": 42
-}
-
-# Example 2: Chinese prompt
-chinese_data = {
-    "prompt": "一只可爱的小猫坐在花园里，阳光明媚",
-    "width": 1024,
-    "height": 1024,
-    "num_inference_steps": 50
-}
-
-# Call the model with extended timeout for image generation
-print("Generating image... This may take a moment.")
-response = httpx.post(
-    f"https://model-{model_id}.api.baseten.co/development/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=english_data,
-    timeout=httpx.Timeout(120.0)
-)
-
-# Get the result
-result = response.json()
-image_b64 = result.get("data")
-
-# Convert to image and save
-image = b64_to_pil(image_b64)
-image.save("generated_image.jpg")
-print("Image generated successfully! Saved as 'generated_image.jpg'")
-```
-
-## Aspect Ratio Examples
-
-The model supports various aspect ratios. Here are some common configurations:
-
-```python
-# Square (1:1)
-{"width": 1024, "height": 1024}
-
-# Landscape (16:9)
-{"width": 1664, "height": 928}
-
-# Portrait (9:16)
-{"width": 928, "height": 1664}
-
-# Traditional (4:3)
-{"width": 1472, "height": 1140}
-
-# Portrait Traditional (3:4)
-{"width": 1140, "height": 1472}
-```
-
-## Deployment
-
-To deploy this model:
-
-1. Clone the repository
-2. Make sure you have the Truss CLI installed (`pip install truss`)
-3. Run the deployment command:
-
-```bash
-truss push qwen/qwen-image --publish
-```
-
-## Model Features
-
-- **Automatic Language Detection**: The model automatically detects Chinese vs English prompts and applies appropriate quality enhancements
-- **Quality Enhancement**: Automatically adds "Ultra HD, 4K, cinematic composition" for English prompts or "超清，4K，电影级构图" for Chinese prompts
-- **GPU Optimization**: Uses bfloat16 precision on CUDA devices for optimal performance
-- **Base64 Output**: Returns images as base64-encoded strings for easy API integration
-
-## Requirements
-
-- CUDA-compatible GPU (recommended for optimal performance)
-- Python 3.8+
-- PyTorch
-- Diffusers library
-- Transformers library
diff --git a/qwen/qwen-image/config.yaml b/qwen/qwen-image/config.yaml
deleted file mode 100644
index 7c1f200ff..000000000
--- a/qwen/qwen-image/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-external_package_dirs: []
-model_cache:
-  - repo_id: Qwen/Qwen-Image
-    use_volume: false
-    allow_patterns:
-      - "*.json"
-      - "*.safetensors"
-      - "*.bin"
-      - "*.txt"
-      - "*.md"
-model_metadata:
-  example_model_input: {
-    "prompt": "A beautiful sunset over a mountain landscape with golden clouds, Ultra HD, 4K, cinematic composition",
-    "width": 1024,
-    "height": 1024,
-    "num_inference_steps": 50,
-    "true_cfg_scale": 4.0,
-    "seed": 42
-  }
-model_name: Qwen Image
-python_version: py311
-requirements:
-  - git+https://github.com/huggingface/diffusers
-  - torch>=2.5.1
-  - transformers>=4.48.2
-  - accelerate
-  - safetensors
-  - pillow
-  - numpy
-resources:
-  accelerator: H100
-  use_gpu: true
-secrets: {}
-system_packages:
-  - ffmpeg
-  - libsm6
-  - libxext6
diff --git a/qwen/qwen-vl/README.md b/qwen/qwen-vl/README.md
deleted file mode 100644
index f92c0e7db..000000000
--- a/qwen/qwen-vl/README.md
+++ /dev/null
@@ -1,81 +0,0 @@
-# Qwen VL Truss
-
-This is a [Truss](https://truss.baseten.co/) for [Qwen-VL](https://huggingface.co/Qwen/Qwen-VL) which is a visual language model. Qwen is a family of models developed by Alibaba Cloud. This LLM supports both English and Chinese.
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers and deploy on Baseten.
-
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd qwen-vl
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `qwen-vl` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-
-## Qwen-VL API documentation
-
-This section provides an overview of the Qwen-VL model, its parameters, and how to use it. The API consists of a single route named  `predict`, which you can invoke to generate text based on the provided prompt.
-
-### API route: `predict`
-
-The predict route is the primary method for generating text completions based on a given prompt. It takes several parameters:
-
-- __prompt__: The instruction the model will follow to extract the data from the image.
-- __image__ : The input image in the form of a URL or a base64 string.
-
-
-## Example usage
-
-```python
-import requests
-import base64
-from PIL import Image
-from io import BytesIO
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-
-data = {
-  "image": pil_to_b64(Image.open("/path/to/image/dog.jpg")),
-  "prompt": "Generate the caption in English with grounding"
-}
-
-res = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers={"Authorization": "Api-Key <BASETEN-API-KEY>"},
-    json=data,
-)
-
-print(res.json())
-```
-
-## Example Output
-
-```json
-{"output": "Picture 1: <img>/tmp/tmpw6m_zmbk.png</img>\nGenerate the caption in English with grounding<ref> A maltese dog</ref><box>(385,361),(783,934)</box> in a flower garden<|endoftext|>"}
-```
-![galu2](https://github.com/basetenlabs/truss-examples/assets/15642666/459d5a5a-37b0-49aa-830c-d933840c40a2)
diff --git a/qwen/qwen-vl/config.yaml b/qwen/qwen-vl/config.yaml
deleted file mode 100644
index 879020997..000000000
--- a/qwen/qwen-vl/config.yaml
+++ /dev/null
@@ -1,27 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_cache:
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - '*.bin'
-  - '*.tiktoken'
-  - '*.py'
-  repo_id: Qwen/Qwen-VL
-  use_volume: false
-model_name: Qwen VL
-python_version: py310
-requirements:
-- torch==2.0.1
-- accelerate==0.24.0
-- transformers==4.35.0
-- einops==0.7.0
-- torchvision==0.15.2
-- matplotlib==3.8.2
-- tiktoken==0.5.2
-- transformers_stream_generator==0.0.4
-resources:
-  accelerator: A10G
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/sana/sana_1600M/config.yaml b/sana/sana_1600M/config.yaml
deleted file mode 100644
index b3cfbf751..000000000
--- a/sana/sana_1600M/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-build_commands: []
-base_image:
-  image: alphatozeta/cuda-python:12.1.1-cudnn8-devel-ubuntu22.04
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input: {
-    "prompt": "a photo of an astronaut riding a horse on mars",
-    "height": 1024,
-    "width": 1024,
-    "guidance_scale": 5.0,
-    "pag_guidance_scale": 2.0,
-    "num_inference_steps": 18,
-    "seed": 4096,
-  }
-model_name: Sana 1600M
-python_version: py311
-requirements:
-- git+https://github.com/NVlabs/Sana.git@d7945026d8d85008aca1d1e6db5717a1069f5c84
-- huggingface-hub==0.26.3
-- hf-transfer==0.1.8
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-secrets:
-  hf_access_token: "null"
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
-- python3.10-venv
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/dpm_solver.py b/sana/sana_1600M/packages/Sana/diffusion/model/dpm_solver.py
deleted file mode 100755
index 826c373ec..000000000
--- a/sana/sana_1600M/packages/Sana/diffusion/model/dpm_solver.py
+++ /dev/null
@@ -1,1908 +0,0 @@
-# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-
-# This file is modified from https://github.com/PixArt-alpha/PixArt-sigma
-import os
-
-import torch
-from tqdm import tqdm
-
-from .nets.sana_blocks import (
-    PAGCFGIdentitySelfAttnProcessorLiteLA,
-    PAGIdentitySelfAttnProcessorLiteLA,
-    SelfAttnProcessorLiteLA,
-)
-
-
-class NoiseScheduleVP:
-    def __init__(
-        self,
-        schedule="discrete",
-        betas=None,
-        alphas_cumprod=None,
-        continuous_beta_0=0.1,
-        continuous_beta_1=20.0,
-        dtype=torch.float32,
-    ):
-        r"""Create a wrapper class for the forward SDE (VP type).
-
-        ***
-        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
-                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
-        ***
-
-        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
-        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
-        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
-
-            log_alpha_t = self.marginal_log_mean_coeff(t)
-            sigma_t = self.marginal_std(t)
-            lambda_t = self.marginal_lambda(t)
-
-        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
-
-            t = self.inverse_lambda(lambda_t)
-
-        ===============================================================
-
-        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
-
-        1. For discrete-time DPMs:
-
-            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
-                t_i = (i + 1) / N
-            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
-            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
-
-            Args:
-                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
-                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
-
-            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
-
-            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
-                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
-                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
-                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
-                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
-                and
-                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
-
-
-        2. For continuous-time DPMs:
-
-            We support the linear VPSDE for the continuous time setting. The hyperparameters for the noise
-            schedule are the default settings in Yang Song's ScoreSDE:
-
-            Args:
-                beta_min: A `float` number. The smallest beta for the linear schedule.
-                beta_max: A `float` number. The largest beta for the linear schedule.
-                T: A `float` number. The ending time of the forward process.
-
-        ===============================================================
-
-        Args:
-            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
-                    'linear' for continuous-time DPMs.
-        Returns:
-            A wrapper object of the forward SDE (VP type).
-
-        ===============================================================
-
-        Example:
-
-        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', betas=betas)
-
-        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
-
-        # For continuous-time DPMs (VPSDE), linear schedule:
-        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
-
-        """
-
-        if schedule not in ["discrete", "linear"]:
-            raise ValueError(
-                f"Unsupported noise schedule {schedule}. The schedule needs to be 'discrete' or 'linear'"
-            )
-
-        self.schedule = schedule
-        if schedule == "discrete":
-            if betas is not None:
-                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
-            else:
-                assert alphas_cumprod is not None
-                log_alphas = 0.5 * torch.log(alphas_cumprod)
-            self.T = 1.0
-            self.log_alpha_array = (
-                self.numerical_clip_alpha(log_alphas)
-                .reshape(
-                    (
-                        1,
-                        -1,
-                    )
-                )
-                .to(dtype=dtype)
-            )
-            self.total_N = self.log_alpha_array.shape[1]
-            self.t_array = (
-                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
-                .reshape((1, -1))
-                .to(dtype=dtype)
-            )
-        else:
-            self.T = 1.0
-            self.total_N = 1000
-            self.beta_0 = continuous_beta_0
-            self.beta_1 = continuous_beta_1
-
-    def numerical_clip_alpha(self, log_alphas, clipped_lambda=-5.1):
-        """
-        For some beta schedules such as cosine schedule, the log-SNR has numerical isssues.
-        We clip the log-SNR near t=T within -5.1 to ensure the stability.
-        Such a trick is very useful for diffusion models with the cosine schedule, such as i-DDPM, guided-diffusion and GLIDE.
-        """
-        log_sigmas = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_alphas))
-        lambs = log_alphas - log_sigmas
-        idx = torch.searchsorted(torch.flip(lambs, [0]), clipped_lambda)
-        if idx > 0:
-            log_alphas = log_alphas[:-idx]
-        return log_alphas
-
-    def marginal_log_mean_coeff(self, t):
-        """
-        Compute log(alpha_t) of a given continuous-time label t in [0, T].
-        """
-        if self.schedule == "discrete":
-            return interpolate_fn(
-                t.reshape((-1, 1)),
-                self.t_array.to(t.device),
-                self.log_alpha_array.to(t.device),
-            ).reshape(-1)
-        elif self.schedule == "linear":
-            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
-
-    def marginal_alpha(self, t):
-        """
-        Compute alpha_t of a given continuous-time label t in [0, T].
-        """
-        return torch.exp(self.marginal_log_mean_coeff(t))
-
-    def marginal_std(self, t):
-        """
-        Compute sigma_t of a given continuous-time label t in [0, T].
-        """
-        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
-
-    def marginal_lambda(self, t):
-        """
-        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
-        """
-        log_mean_coeff = self.marginal_log_mean_coeff(t)
-        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
-        return log_mean_coeff - log_std
-
-    def inverse_lambda(self, lamb):
-        """
-        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
-        """
-        if self.schedule == "linear":
-            tmp = (
-                2.0
-                * (self.beta_1 - self.beta_0)
-                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
-            )
-            Delta = self.beta_0**2 + tmp
-            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
-        elif self.schedule == "discrete":
-            log_alpha = -0.5 * torch.logaddexp(
-                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
-            )
-            t = interpolate_fn(
-                log_alpha.reshape((-1, 1)),
-                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
-                torch.flip(self.t_array.to(lamb.device), [1]),
-            )
-            return t.reshape((-1,))
-
-
-class NoiseScheduleFlow:
-    def __init__(
-        self,
-        schedule="discrete_flow",
-    ):
-        """Create a wrapper class for the forward SDE (EDM type)."""
-        self.T = 1
-        self.t0 = 0.001
-        self.schedule = schedule  # ['continuous', 'discrete_flow']
-        self.total_N = 1000
-
-    def marginal_log_mean_coeff(self, t):
-        """
-        Compute log(alpha_t) of a given continuous-time label t in [0, T].
-        """
-        return torch.log(self.marginal_alpha(t))
-
-    def marginal_alpha(self, t):
-        """
-        Compute alpha_t of a given continuous-time label t in [0, T].
-        """
-        return 1 - t
-
-    @staticmethod
-    def marginal_std(t):
-        """
-        Compute sigma_t of a given continuous-time label t in [0, T].
-        """
-        return t
-
-    def marginal_lambda(self, t):
-        """
-        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
-        """
-        log_mean_coeff = self.marginal_log_mean_coeff(t)
-        log_std = torch.log(self.marginal_std(t))
-        return log_mean_coeff - log_std
-
-    @staticmethod
-    def inverse_lambda(lamb):
-        """
-        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
-        """
-        return torch.exp(-lamb)
-
-    def edm_sigma(self, t):
-        return self.marginal_std(t) / self.marginal_alpha(t)
-
-    def edm_inverse_sigma(self, edmsigma):
-        sigma = edmsigma
-        lambda_t = torch.log(1 / sigma)
-        t = self.inverse_lambda(lambda_t)
-        return t
-
-
-def model_wrapper(
-    model,
-    noise_schedule,
-    model_type="noise",
-    model_kwargs={},
-    guidance_type="uncond",
-    condition=None,
-    unconditional_condition=None,
-    guidance_scale=1.0,
-    pag_scale=1.0,
-    pag_applied_layers=[],
-    interval_guidance=[0, 1.0],
-    classifier_fn=None,
-    classifier_kwargs={},
-):
-    """Create a wrapper function for the noise prediction model.
-
-    DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
-    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
-
-    We support four types of the diffusion model by setting `model_type`:
-
-        1. "noise": noise prediction model. (Trained by predicting noise).
-
-        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
-
-        3. "v": velocity prediction model. (Trained by predicting the velocity).
-            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
-
-            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
-                arXiv preprint arXiv:2202.00512 (2022).
-            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
-                arXiv preprint arXiv:2210.02303 (2022).
-
-        4. "score": marginal score function. (Trained by denoising score matching).
-            Note that the score function and the noise prediction model follows a simple relationship:
-            ```
-                noise(x_t, t) = -sigma_t * score(x_t, t)
-            ```
-
-    We support three types of guided sampling by DPMs by setting `guidance_type`:
-        1. "uncond": unconditional sampling by DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-
-        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-
-            The input `classifier_fn` has the following format:
-            ``
-                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
-            ``
-
-            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
-                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
-
-        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
-            ``
-            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
-
-            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
-                arXiv preprint arXiv:2207.12598 (2022).
-
-
-    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
-    or continuous-time labels (i.e. epsilon to T).
-
-    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
-    ``
-        def model_fn(x, t_continuous) -> noise:
-            t_input = get_model_input_time(t_continuous)
-            return noise_pred(model, x, t_input, **model_kwargs)
-    ``
-    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
-
-    ===============================================================
-
-    Args:
-        model: A diffusion model with the corresponding format described above.
-        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
-        model_type: A `str`. The parameterization type of the diffusion model.
-                    "noise" or "x_start" or "v" or "score".
-        model_kwargs: A `dict`. A dict for the other inputs of the model function.
-        guidance_type: A `str`. The type of the guidance for sampling.
-                    "uncond" or "classifier" or "classifier-free".
-        condition: A pytorch tensor. The condition for the guided sampling.
-                    Only used for "classifier" or "classifier-free" guidance type.
-        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
-                    Only used for "classifier-free" guidance type.
-        guidance_scale: A `float`. The scale for the guided sampling.
-        classifier_fn: A classifier function. Only used for the classifier guidance.
-        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
-    Returns:
-        A noise prediction model that accepts the noised data and the continuous time as the inputs.
-    """
-
-    def get_model_input_time(t_continuous):
-        """
-        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
-        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
-        For continuous-time DPMs, we just use `t_continuous`.
-        """
-        if noise_schedule.schedule == "discrete":
-            return (
-                t_continuous - 1.0 / noise_schedule.total_N
-            ) * noise_schedule.total_N
-        elif noise_schedule.schedule == "discrete_flow":
-            return t_continuous * noise_schedule.total_N
-        else:
-            return t_continuous
-
-    def noise_pred_fn(x, t_continuous, cond=None):
-        t_input = get_model_input_time(t_continuous)
-        if cond is None:
-            output = model(x, t_input, **model_kwargs)
-        else:
-            output = model(x, t_input, cond, **model_kwargs)
-        if model_type == "noise":
-            return output
-        elif model_type == "x_start":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return (x - expand_dims(alpha_t, x.dim()) * output) / expand_dims(
-                sigma_t, x.dim()
-            )
-        elif model_type == "v":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return (
-                expand_dims(alpha_t, x.dim()) * output
-                + expand_dims(sigma_t, x.dim()) * x
-            )
-        elif model_type == "score":
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            return -expand_dims(sigma_t, x.dim()) * output
-        elif model_type == "flow":
-            _, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            try:
-                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output + x
-            except:
-                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output[0] + x
-            return noise
-
-    def cond_grad_fn(x, t_input):
-        """
-        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
-        """
-        with torch.enable_grad():
-            x_in = x.detach().requires_grad_(True)
-            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
-            return torch.autograd.grad(log_prob.sum(), x_in)[0]
-
-    def model_fn(x, t_continuous):
-        """
-        The noise predicition model function that is used for DPM-Solver.
-        """
-        guidance_tp = guidance_type
-        if guidance_tp == "uncond":
-            return noise_pred_fn(x, t_continuous)
-        elif guidance_tp == "classifier":
-            assert classifier_fn is not None
-            t_input = get_model_input_time(t_continuous)
-            cond_grad = cond_grad_fn(x, t_input)
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            noise = noise_pred_fn(x, t_continuous)
-            return noise - guidance_scale * expand_dims(sigma_t, x.dim()) * cond_grad
-        elif guidance_tp == "classifier-free":
-            if (
-                guidance_scale == 1.0
-                or unconditional_condition is None
-                or not (interval_guidance[0] < t_continuous[0] < interval_guidance[1])
-            ):
-                return noise_pred_fn(x, t_continuous, cond=condition)
-            else:
-                x_in = torch.cat([x] * 2)
-                t_in = torch.cat([t_continuous] * 2)
-                c_in = torch.cat([unconditional_condition, condition])
-                try:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
-                except:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
-                        2
-                    )
-                return noise_uncond + guidance_scale * (noise - noise_uncond)
-        elif guidance_tp == "classifier-free_PAG":
-            for i in pag_applied_layers:
-                if isinstance(model, torch.nn.Module):
-                    model.blocks[i].attn.forward = (
-                        PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
-                        if guidance_scale == 1.0
-                        else PAGCFGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
-                    )
-                else:
-                    model.__self__.blocks[i].attn.forward = (
-                        PAGIdentitySelfAttnProcessorLiteLA(
-                            model.__self__.blocks[i].attn
-                        )
-                        if guidance_scale == 1.0
-                        else PAGCFGIdentitySelfAttnProcessorLiteLA(
-                            model.__self__.blocks[i].attn
-                        )
-                    )
-            num_inputs = 2 if guidance_scale == 1.0 else 3
-            x_in = torch.cat([x] * num_inputs)
-            t_in = torch.cat([t_continuous] * num_inputs)
-            c_in = torch.cat(
-                [condition, condition]
-                if guidance_scale == 1.0
-                else [unconditional_condition, condition, condition]
-            )
-
-            try:
-                chunks = noise_pred_fn(x_in, t_in, cond=c_in).chunk(num_inputs)
-            except:
-                chunks = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(num_inputs)
-
-            if guidance_scale == 1.0:
-                noise, noise_perturb = chunks
-                noise_pred = noise + pag_scale * (noise - noise_perturb)
-            else:
-                noise_uncond, noise, noise_perturb = chunks
-                noise_pred = (
-                    noise_uncond
-                    + guidance_scale * (noise - noise_uncond)
-                    + pag_scale * (noise - noise_perturb)
-                )
-            for i in pag_applied_layers:
-                if isinstance(model, torch.nn.Module):
-                    model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                        model.blocks[i].attn
-                    )
-                else:
-                    model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                        model.__self__.blocks[i].attn
-                    )
-
-            return noise_pred
-        elif guidance_tp == "classifier-free_PAG_seq":
-            num_inputs = 2
-            if t_continuous[0] < 0.5:
-                # cfg
-                if (
-                    guidance_scale == 1.0
-                    or unconditional_condition is None
-                    or not (
-                        interval_guidance[0] < t_continuous[0] < interval_guidance[1]
-                    )
-                ):
-                    return noise_pred_fn(x, t_continuous, cond=condition)
-
-                x_in = torch.cat([x] * num_inputs)
-                t_in = torch.cat([t_continuous] * num_inputs)
-                c_in = torch.cat([unconditional_condition, condition])
-
-                try:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
-                except:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
-                        num_inputs
-                    )
-                return noise_uncond + guidance_scale * (noise - noise_uncond)
-            else:
-                # pag
-                for i in pag_applied_layers:
-                    if isinstance(model, torch.nn.Module):
-                        model.blocks[i].attn.forward = (
-                            PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
-                            if guidance_scale == 1.0
-                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
-                                model.blocks[i].attn
-                            )
-                        )
-                    else:
-                        model.__self__.blocks[i].attn.forward = (
-                            PAGIdentitySelfAttnProcessorLiteLA(
-                                model.__self__.blocks[i].attn
-                            )
-                            if guidance_scale == 1.0
-                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
-                                model.__self__.blocks[i].attn
-                            )
-                        )
-                x_in = torch.cat([x] * 3)
-                t_in = torch.cat([t_continuous] * 3)
-                c_in = torch.cat([unconditional_condition, condition, condition])
-
-                try:
-                    noise_uncond, noise, noise_perturb = noise_pred_fn(
-                        x_in, t_in, cond=c_in
-                    ).chunk(3)
-                except:
-                    noise_uncond, noise, noise_perturb = noise_pred_fn(
-                        x_in, t_in, cond=c_in
-                    )[0].chunk(3)
-
-                for i in pag_applied_layers:
-                    if isinstance(model, torch.nn.Module):
-                        model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                            model.blocks[i].attn
-                        )
-                    else:
-                        model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                            model.__self__.blocks[i].attn
-                        )
-
-                return (
-                    noise_uncond
-                    + guidance_scale * (noise - noise_uncond)
-                    + pag_scale * (noise - noise_perturb)
-                )
-
-    assert model_type in ["noise", "x_start", "v", "score", "flow"]
-    assert guidance_type in [
-        "uncond",
-        "classifier",
-        "classifier-free",
-        "classifier-free_PAG",
-        "classifier-free_PAG_seq",
-    ]
-    return model_fn
-
-
-class DPM_Solver:
-    def __init__(
-        self,
-        model_fn,
-        noise_schedule,
-        algorithm_type="dpmsolver++",
-        correcting_x0_fn=None,
-        correcting_xt_fn=None,
-        thresholding_max_val=1.0,
-        dynamic_thresholding_ratio=0.995,
-    ):
-        """Construct a DPM-Solver.
-
-        We support both DPM-Solver (`algorithm_type="dpmsolver"`) and DPM-Solver++ (`algorithm_type="dpmsolver++"`).
-
-        We also support the "dynamic thresholding" method in Imagen[1]. For pixel-space diffusion models, you
-        can set both `algorithm_type="dpmsolver++"` and `correcting_x0_fn="dynamic_thresholding"` to use the
-        dynamic thresholding. The "dynamic thresholding" can greatly improve the sample quality for pixel-space
-        DPMs with large guidance scales. Note that the thresholding method is **unsuitable** for latent-space
-        DPMs (such as stable-diffusion).
-
-        To support advanced algorithms in image-to-image applications, we also support corrector functions for
-        both x0 and xt.
-
-        Args:
-            model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
-                ``
-                def model_fn(x, t_continuous):
-                    return noise
-                ``
-                The shape of `x` is `(batch_size, **shape)`, and the shape of `t_continuous` is `(batch_size,)`.
-            noise_schedule: A noise schedule object, such as NoiseScheduleVP.
-            algorithm_type: A `str`. Either "dpmsolver" or "dpmsolver++".
-            correcting_x0_fn: A `str` or a function with the following format:
-                ```
-                def correcting_x0_fn(x0, t):
-                    x0_new = ...
-                    return x0_new
-                ```
-                This function is to correct the outputs of the data prediction model at each sampling step. e.g.,
-                ```
-                x0_pred = data_pred_model(xt, t)
-                if correcting_x0_fn is not None:
-                    x0_pred = correcting_x0_fn(x0_pred, t)
-                xt_1 = update(x0_pred, xt, t)
-                ```
-                If `correcting_x0_fn="dynamic_thresholding"`, we use the dynamic thresholding proposed in Imagen[1].
-            correcting_xt_fn: A function with the following format:
-                ```
-                def correcting_xt_fn(xt, t, step):
-                    x_new = ...
-                    return x_new
-                ```
-                This function is to correct the intermediate samples xt at each sampling step. e.g.,
-                ```
-                xt = ...
-                xt = correcting_xt_fn(xt, t, step)
-                ```
-            thresholding_max_val: A `float`. The max value for thresholding.
-                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
-            dynamic_thresholding_ratio: A `float`. The ratio for dynamic thresholding (see Imagen[1] for details).
-                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
-
-        [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour,
-            Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models
-            with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
-        """
-        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
-        self.noise_schedule = noise_schedule
-        assert algorithm_type in ["dpmsolver", "dpmsolver++"]
-        self.algorithm_type = algorithm_type
-        if correcting_x0_fn == "dynamic_thresholding":
-            self.correcting_x0_fn = self.dynamic_thresholding_fn
-        else:
-            self.correcting_x0_fn = correcting_x0_fn
-        self.correcting_xt_fn = correcting_xt_fn
-        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
-        self.thresholding_max_val = thresholding_max_val
-        self.register_progress_bar()
-
-    def register_progress_bar(self, progress_fn=None):
-        """
-        Register a progress bar callback function
-
-        Args:
-            progress_fn: Callback function that takes current step and total steps as parameters
-        """
-        self.progress_fn = (
-            progress_fn if progress_fn is not None else lambda step, total: None
-        )
-
-    def update_progress(self, step, total_steps):
-        """
-        Update sampling progress
-
-        Args:
-            step: Current step number
-            total_steps: Total number of steps
-        """
-        if hasattr(self, "progress_fn"):
-            try:
-                self.progress_fn(
-                    step / total_steps, desc=f"Generating {step}/{total_steps}"
-                )
-            except:
-                self.progress_fn(step, total_steps)
-
-        else:
-            # If no progress_fn registered, use default empty function
-            pass
-
-    def dynamic_thresholding_fn(self, x0, t):
-        """
-        The dynamic thresholding method.
-        """
-        dims = x0.dim()
-        p = self.dynamic_thresholding_ratio
-        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
-        s = expand_dims(
-            torch.maximum(
-                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
-            ),
-            dims,
-        )
-        x0 = torch.clamp(x0, -s, s) / s
-        return x0
-
-    def noise_prediction_fn(self, x, t):
-        """
-        Return the noise prediction model.
-        """
-        return self.model(x, t)
-
-    def data_prediction_fn(self, x, t):
-        """
-        Return the data prediction model (with corrector).
-        """
-        noise = self.noise_prediction_fn(x, t)
-        alpha_t, sigma_t = (
-            self.noise_schedule.marginal_alpha(t),
-            self.noise_schedule.marginal_std(t),
-        )
-        x0 = (x - sigma_t * noise) / alpha_t
-        if self.correcting_x0_fn is not None:
-            x0 = self.correcting_x0_fn(x0, t)
-        return x0
-
-    def model_fn(self, x, t):
-        """
-        Convert the model to the noise prediction model or the data prediction model.
-        """
-        if self.algorithm_type == "dpmsolver++":
-            return self.data_prediction_fn(x, t)
-        else:
-            return self.noise_prediction_fn(x, t)
-
-    def get_time_steps(self, skip_type, t_T, t_0, N, device, shift=1.0):
-        """Compute the intermediate time steps for sampling.
-
-        Args:
-            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
-                - 'logSNR': uniform logSNR for the time steps.
-                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
-                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
-            t_T: A `float`. The starting time of the sampling (default is T).
-            t_0: A `float`. The ending time of the sampling (default is epsilon).
-            N: A `int`. The total number of the spacing of the time steps.
-            device: A torch device.
-        Returns:
-            A pytorch tensor of the time steps, with the shape (N + 1,).
-        """
-        if skip_type == "logSNR":
-            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
-            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
-            logSNR_steps = torch.linspace(
-                lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1
-            ).to(device)
-            return self.noise_schedule.inverse_lambda(logSNR_steps)
-        elif skip_type == "time_uniform":
-            return torch.linspace(t_T, t_0, N + 1).to(device)
-        elif skip_type == "time_quadratic":
-            t_order = 2
-            t = (
-                torch.linspace(t_T ** (1.0 / t_order), t_0 ** (1.0 / t_order), N + 1)
-                .pow(t_order)
-                .to(device)
-            )
-            return t
-        elif skip_type == "time_uniform_flow":
-            betas = torch.linspace(t_T, t_0, N + 1).to(device)
-            sigmas = 1.0 - betas
-            sigmas = (shift * sigmas / (1 + (shift - 1) * sigmas)).flip(dims=[0])
-            return sigmas
-        else:
-            raise ValueError(
-                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'"
-            )
-
-    def get_orders_and_timesteps_for_singlestep_solver(
-        self, steps, order, skip_type, t_T, t_0, device
-    ):
-        """
-        Get the order of each step for sampling by the singlestep DPM-Solver.
-
-        We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
-        Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
-            - If order == 1:
-                We take `steps` of DPM-Solver-1 (i.e. DDIM).
-            - If order == 2:
-                - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
-                - If steps % 2 == 0, we use K steps of DPM-Solver-2.
-                - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
-            - If order == 3:
-                - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
-                - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
-                - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
-                - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
-
-        ============================================
-        Args:
-            order: A `int`. The max order for the solver (2 or 3).
-            steps: A `int`. The total number of function evaluations (NFE).
-            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
-                - 'logSNR': uniform logSNR for the time steps.
-                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
-                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
-            t_T: A `float`. The starting time of the sampling (default is T).
-            t_0: A `float`. The ending time of the sampling (default is epsilon).
-            device: A torch device.
-        Returns:
-            orders: A list of the solver order of each step.
-        """
-        if order == 3:
-            K = steps // 3 + 1
-            if steps % 3 == 0:
-                orders = [
-                    3,
-                ] * (K - 2) + [2, 1]
-            elif steps % 3 == 1:
-                orders = [
-                    3,
-                ] * (K - 1) + [1]
-            else:
-                orders = [
-                    3,
-                ] * (K - 1) + [2]
-        elif order == 2:
-            if steps % 2 == 0:
-                K = steps // 2
-                orders = [
-                    2,
-                ] * K
-            else:
-                K = steps // 2 + 1
-                orders = [
-                    2,
-                ] * (K - 1) + [1]
-        elif order == 1:
-            K = 1
-            orders = [
-                1,
-            ] * steps
-        else:
-            raise ValueError("'order' must be '1' or '2' or '3'.")
-        if skip_type == "logSNR":
-            # To reproduce the results in DPM-Solver paper
-            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
-        else:
-            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
-                torch.cumsum(
-                    torch.tensor(
-                        [
-                            0,
-                        ]
-                        + orders
-                    ),
-                    0,
-                ).to(device)
-            ]
-        return timesteps_outer, orders
-
-    def denoise_to_zero_fn(self, x, s):
-        """
-        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
-        """
-        return self.data_prediction_fn(x, s)
-
-    def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
-        """
-        DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            model_s: A pytorch tensor. The model function evaluated at time `s`.
-                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
-            return_intermediate: A `bool`. If true, also return the model value at time `s`.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        ns = self.noise_schedule
-        dims = x.dim()
-        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
-        h = lambda_t - lambda_s
-        log_alpha_s, log_alpha_t = (
-            ns.marginal_log_mean_coeff(s),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
-        alpha_t = torch.exp(log_alpha_t)
-
-        if self.algorithm_type == "dpmsolver++":
-            phi_1 = torch.expm1(-h)
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_t = sigma_t / sigma_s * x - alpha_t * phi_1 * model_s
-            if return_intermediate:
-                return x_t, {"model_s": model_s}
-            else:
-                return x_t
-        else:
-            phi_1 = torch.expm1(h)
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_t = torch.exp(log_alpha_t - log_alpha_s) * x - (sigma_t * phi_1) * model_s
-            if return_intermediate:
-                return x_t, {"model_s": model_s}
-            else:
-                return x_t
-
-    def singlestep_dpm_solver_second_update(
-        self,
-        x,
-        s,
-        t,
-        r1=0.5,
-        model_s=None,
-        return_intermediate=False,
-        solver_type="dpmsolver",
-    ):
-        """
-        Singlestep solver DPM-Solver-2 from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            r1: A `float`. The hyperparameter of the second-order solver.
-            model_s: A pytorch tensor. The model function evaluated at time `s`.
-                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
-            return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if solver_type not in ["dpmsolver", "taylor"]:
-            raise ValueError(
-                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
-            )
-        if r1 is None:
-            r1 = 0.5
-        ns = self.noise_schedule
-        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
-        h = lambda_t - lambda_s
-        lambda_s1 = lambda_s + r1 * h
-        s1 = ns.inverse_lambda(lambda_s1)
-        log_alpha_s, log_alpha_s1, log_alpha_t = (
-            ns.marginal_log_mean_coeff(s),
-            ns.marginal_log_mean_coeff(s1),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_s, sigma_s1, sigma_t = (
-            ns.marginal_std(s),
-            ns.marginal_std(s1),
-            ns.marginal_std(t),
-        )
-        alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
-
-        if self.algorithm_type == "dpmsolver++":
-            phi_11 = torch.expm1(-r1 * h)
-            phi_1 = torch.expm1(-h)
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
-            model_s1 = self.model_fn(x_s1, s1)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    - (0.5 / r1) * (alpha_t * phi_1) * (model_s1 - model_s)
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    + (1.0 / r1) * (alpha_t * (phi_1 / h + 1.0)) * (model_s1 - model_s)
-                )
-        else:
-            phi_11 = torch.expm1(r1 * h)
-            phi_1 = torch.expm1(h)
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_s1 = (
-                torch.exp(log_alpha_s1 - log_alpha_s) * x
-                - (sigma_s1 * phi_11) * model_s
-            )
-            model_s1 = self.model_fn(x_s1, s1)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    torch.exp(log_alpha_t - log_alpha_s) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (0.5 / r1) * (sigma_t * phi_1) * (model_s1 - model_s)
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    torch.exp(log_alpha_t - log_alpha_s) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (1.0 / r1) * (sigma_t * (phi_1 / h - 1.0)) * (model_s1 - model_s)
-                )
-        if return_intermediate:
-            return x_t, {"model_s": model_s, "model_s1": model_s1}
-        else:
-            return x_t
-
-    def singlestep_dpm_solver_third_update(
-        self,
-        x,
-        s,
-        t,
-        r1=1.0 / 3.0,
-        r2=2.0 / 3.0,
-        model_s=None,
-        model_s1=None,
-        return_intermediate=False,
-        solver_type="dpmsolver",
-    ):
-        """
-        Singlestep solver DPM-Solver-3 from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            r1: A `float`. The hyperparameter of the third-order solver.
-            r2: A `float`. The hyperparameter of the third-order solver.
-            model_s: A pytorch tensor. The model function evaluated at time `s`.
-                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
-            model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
-                If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
-            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if solver_type not in ["dpmsolver", "taylor"]:
-            raise ValueError(
-                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
-            )
-        if r1 is None:
-            r1 = 1.0 / 3.0
-        if r2 is None:
-            r2 = 2.0 / 3.0
-        ns = self.noise_schedule
-        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
-        h = lambda_t - lambda_s
-        lambda_s1 = lambda_s + r1 * h
-        lambda_s2 = lambda_s + r2 * h
-        s1 = ns.inverse_lambda(lambda_s1)
-        s2 = ns.inverse_lambda(lambda_s2)
-        log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = (
-            ns.marginal_log_mean_coeff(s),
-            ns.marginal_log_mean_coeff(s1),
-            ns.marginal_log_mean_coeff(s2),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_s, sigma_s1, sigma_s2, sigma_t = (
-            ns.marginal_std(s),
-            ns.marginal_std(s1),
-            ns.marginal_std(s2),
-            ns.marginal_std(t),
-        )
-        alpha_s1, alpha_s2, alpha_t = (
-            torch.exp(log_alpha_s1),
-            torch.exp(log_alpha_s2),
-            torch.exp(log_alpha_t),
-        )
-
-        if self.algorithm_type == "dpmsolver++":
-            phi_11 = torch.expm1(-r1 * h)
-            phi_12 = torch.expm1(-r2 * h)
-            phi_1 = torch.expm1(-h)
-            phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.0
-            phi_2 = phi_1 / h + 1.0
-            phi_3 = phi_2 / h - 0.5
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            if model_s1 is None:
-                x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
-                model_s1 = self.model_fn(x_s1, s1)
-            x_s2 = (
-                (sigma_s2 / sigma_s) * x
-                - (alpha_s2 * phi_12) * model_s
-                + r2 / r1 * (alpha_s2 * phi_22) * (model_s1 - model_s)
-            )
-            model_s2 = self.model_fn(x_s2, s2)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    + (1.0 / r2) * (alpha_t * phi_2) * (model_s2 - model_s)
-                )
-            elif solver_type == "taylor":
-                D1_0 = (1.0 / r1) * (model_s1 - model_s)
-                D1_1 = (1.0 / r2) * (model_s2 - model_s)
-                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
-                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    + (alpha_t * phi_2) * D1
-                    - (alpha_t * phi_3) * D2
-                )
-        else:
-            phi_11 = torch.expm1(r1 * h)
-            phi_12 = torch.expm1(r2 * h)
-            phi_1 = torch.expm1(h)
-            phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.0
-            phi_2 = phi_1 / h - 1.0
-            phi_3 = phi_2 / h - 0.5
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            if model_s1 is None:
-                x_s1 = (torch.exp(log_alpha_s1 - log_alpha_s)) * x - (
-                    sigma_s1 * phi_11
-                ) * model_s
-                model_s1 = self.model_fn(x_s1, s1)
-            x_s2 = (
-                (torch.exp(log_alpha_s2 - log_alpha_s)) * x
-                - (sigma_s2 * phi_12) * model_s
-                - r2 / r1 * (sigma_s2 * phi_22) * (model_s1 - model_s)
-            )
-            model_s2 = self.model_fn(x_s2, s2)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_s)) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (1.0 / r2) * (sigma_t * phi_2) * (model_s2 - model_s)
-                )
-            elif solver_type == "taylor":
-                D1_0 = (1.0 / r1) * (model_s1 - model_s)
-                D1_1 = (1.0 / r2) * (model_s2 - model_s)
-                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
-                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_s)) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (sigma_t * phi_2) * D1
-                    - (sigma_t * phi_3) * D2
-                )
-
-        if return_intermediate:
-            return x_t, {"model_s": model_s, "model_s1": model_s1, "model_s2": model_s2}
-        else:
-            return x_t
-
-    def multistep_dpm_solver_second_update(
-        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
-    ):
-        """
-        Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            model_prev_list: A list of pytorch tensor. The previous computed model values.
-            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if solver_type not in ["dpmsolver", "taylor"]:
-            raise ValueError(
-                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
-            )
-        ns = self.noise_schedule
-        model_prev_1, model_prev_0 = model_prev_list[-2], model_prev_list[-1]
-        t_prev_1, t_prev_0 = t_prev_list[-2], t_prev_list[-1]
-        lambda_prev_1, lambda_prev_0, lambda_t = (
-            ns.marginal_lambda(t_prev_1),
-            ns.marginal_lambda(t_prev_0),
-            ns.marginal_lambda(t),
-        )
-        log_alpha_prev_0, log_alpha_t = (
-            ns.marginal_log_mean_coeff(t_prev_0),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
-        alpha_t = torch.exp(log_alpha_t)
-
-        h_0 = lambda_prev_0 - lambda_prev_1
-        h = lambda_t - lambda_prev_0
-        r0 = h_0 / h
-        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
-        if self.algorithm_type == "dpmsolver++":
-            phi_1 = torch.expm1(-h)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (sigma_t / sigma_prev_0) * x
-                    - (alpha_t * phi_1) * model_prev_0
-                    - 0.5 * (alpha_t * phi_1) * D1_0
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    (sigma_t / sigma_prev_0) * x
-                    - (alpha_t * phi_1) * model_prev_0
-                    + (alpha_t * (phi_1 / h + 1.0)) * D1_0
-                )
-        else:
-            phi_1 = torch.expm1(h)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
-                    - (sigma_t * phi_1) * model_prev_0
-                    - 0.5 * (sigma_t * phi_1) * D1_0
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
-                    - (sigma_t * phi_1) * model_prev_0
-                    - (sigma_t * (phi_1 / h - 1.0)) * D1_0
-                )
-        return x_t
-
-    def multistep_dpm_solver_third_update(
-        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
-    ):
-        """
-        Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            model_prev_list: A list of pytorch tensor. The previous computed model values.
-            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        ns = self.noise_schedule
-        model_prev_2, model_prev_1, model_prev_0 = model_prev_list
-        t_prev_2, t_prev_1, t_prev_0 = t_prev_list
-        lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = (
-            ns.marginal_lambda(t_prev_2),
-            ns.marginal_lambda(t_prev_1),
-            ns.marginal_lambda(t_prev_0),
-            ns.marginal_lambda(t),
-        )
-        log_alpha_prev_0, log_alpha_t = (
-            ns.marginal_log_mean_coeff(t_prev_0),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
-        alpha_t = torch.exp(log_alpha_t)
-
-        h_1 = lambda_prev_1 - lambda_prev_2
-        h_0 = lambda_prev_0 - lambda_prev_1
-        h = lambda_t - lambda_prev_0
-        r0, r1 = h_0 / h, h_1 / h
-        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
-        D1_1 = (1.0 / r1) * (model_prev_1 - model_prev_2)
-        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
-        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
-        if self.algorithm_type == "dpmsolver++":
-            phi_1 = torch.expm1(-h)
-            phi_2 = phi_1 / h + 1.0
-            phi_3 = phi_2 / h - 0.5
-            x_t = (
-                (sigma_t / sigma_prev_0) * x
-                - (alpha_t * phi_1) * model_prev_0
-                + (alpha_t * phi_2) * D1
-                - (alpha_t * phi_3) * D2
-            )
-        else:
-            phi_1 = torch.expm1(h)
-            phi_2 = phi_1 / h - 1.0
-            phi_3 = phi_2 / h - 0.5
-            x_t = (
-                (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
-                - (sigma_t * phi_1) * model_prev_0
-                - (sigma_t * phi_2) * D1
-                - (sigma_t * phi_3) * D2
-            )
-        return x_t
-
-    def singlestep_dpm_solver_update(
-        self,
-        x,
-        s,
-        t,
-        order,
-        return_intermediate=False,
-        solver_type="dpmsolver",
-        r1=None,
-        r2=None,
-    ):
-        """
-        Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
-            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-            r1: A `float`. The hyperparameter of the second-order or third-order solver.
-            r2: A `float`. The hyperparameter of the third-order solver.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if order == 1:
-            return self.dpm_solver_first_update(
-                x, s, t, return_intermediate=return_intermediate
-            )
-        elif order == 2:
-            return self.singlestep_dpm_solver_second_update(
-                x,
-                s,
-                t,
-                return_intermediate=return_intermediate,
-                solver_type=solver_type,
-                r1=r1,
-            )
-        elif order == 3:
-            return self.singlestep_dpm_solver_third_update(
-                x,
-                s,
-                t,
-                return_intermediate=return_intermediate,
-                solver_type=solver_type,
-                r1=r1,
-                r2=r2,
-            )
-        else:
-            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
-
-    def multistep_dpm_solver_update(
-        self, x, model_prev_list, t_prev_list, t, order, solver_type="dpmsolver"
-    ):
-        """
-        Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            model_prev_list: A list of pytorch tensor. The previous computed model values.
-            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if order == 1:
-            return self.dpm_solver_first_update(
-                x, t_prev_list[-1], t, model_s=model_prev_list[-1]
-            )
-        elif order == 2:
-            return self.multistep_dpm_solver_second_update(
-                x, model_prev_list, t_prev_list, t, solver_type=solver_type
-            )
-        elif order == 3:
-            return self.multistep_dpm_solver_third_update(
-                x, model_prev_list, t_prev_list, t, solver_type=solver_type
-            )
-        else:
-            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
-
-    def dpm_solver_adaptive(
-        self,
-        x,
-        order,
-        t_T,
-        t_0,
-        h_init=0.05,
-        atol=0.0078,
-        rtol=0.05,
-        theta=0.9,
-        t_err=1e-5,
-        solver_type="dpmsolver",
-    ):
-        """
-        The adaptive step size solver based on singlestep DPM-Solver.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `t_T`.
-            order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
-            t_T: A `float`. The starting time of the sampling (default is T).
-            t_0: A `float`. The ending time of the sampling (default is epsilon).
-            h_init: A `float`. The initial step size (for logSNR).
-            atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
-            rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
-            theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
-            t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
-                current time and `t_0` is less than `t_err`. The default setting is 1e-5.
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_0: A pytorch tensor. The approximated solution at time `t_0`.
-
-        [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
-        """
-        ns = self.noise_schedule
-        s = t_T * torch.ones((1,)).to(x)
-        lambda_s = ns.marginal_lambda(s)
-        lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
-        h = h_init * torch.ones_like(s).to(x)
-        x_prev = x
-        nfe = 0
-        if order == 2:
-            r1 = 0.5
-            lower_update = lambda x, s, t: self.dpm_solver_first_update(
-                x, s, t, return_intermediate=True
-            )
-            higher_update = (
-                lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(
-                    x, s, t, r1=r1, solver_type=solver_type, **kwargs
-                )
-            )
-        elif order == 3:
-            r1, r2 = 1.0 / 3.0, 2.0 / 3.0
-            lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(
-                x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type
-            )
-            higher_update = (
-                lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(
-                    x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs
-                )
-            )
-        else:
-            raise ValueError(
-                f"For adaptive step size solver, order must be 2 or 3, got {order}"
-            )
-        while torch.abs(s - t_0).mean() > t_err:
-            t = ns.inverse_lambda(lambda_s + h)
-            x_lower, lower_noise_kwargs = lower_update(x, s, t)
-            x_higher = higher_update(x, s, t, **lower_noise_kwargs)
-            delta = torch.max(
-                torch.ones_like(x).to(x) * atol,
-                rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)),
-            )
-            norm_fn = lambda v: torch.sqrt(
-                torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)
-            )
-            E = norm_fn((x_higher - x_lower) / delta).max()
-            if torch.all(E <= 1.0):
-                x = x_higher
-                s = t
-                x_prev = x_lower
-                lambda_s = ns.marginal_lambda(s)
-            h = torch.min(
-                theta * h * torch.float_power(E, -1.0 / order).float(),
-                lambda_0 - lambda_s,
-            )
-            nfe += order
-        print("adaptive solver nfe", nfe)
-        return x
-
-    def add_noise(self, x, t, noise=None):
-        """
-        Compute the noised input xt = alpha_t * x + sigma_t * noise.
-
-        Args:
-            x: A `torch.Tensor` with shape `(batch_size, *shape)`.
-            t: A `torch.Tensor` with shape `(t_size,)`.
-        Returns:
-            xt with shape `(t_size, batch_size, *shape)`.
-        """
-        alpha_t, sigma_t = (
-            self.noise_schedule.marginal_alpha(t),
-            self.noise_schedule.marginal_std(t),
-        )
-        if noise is None:
-            noise = torch.randn((t.shape[0], *x.shape), device=x.device)
-        x = x.reshape((-1, *x.shape))
-        xt = expand_dims(alpha_t, x.dim()) * x + expand_dims(sigma_t, x.dim()) * noise
-        if t.shape[0] == 1:
-            return xt.squeeze(0)
-        else:
-            return xt
-
-    def inverse(
-        self,
-        x,
-        steps=20,
-        t_start=None,
-        t_end=None,
-        order=2,
-        skip_type="time_uniform",
-        method="multistep",
-        lower_order_final=True,
-        denoise_to_zero=False,
-        solver_type="dpmsolver",
-        atol=0.0078,
-        rtol=0.05,
-        return_intermediate=False,
-    ):
-        """
-        Inverse the sample `x` from time `t_start` to `t_end` by DPM-Solver.
-        For discrete-time DPMs, we use `t_start=1/N`, where `N` is the total time steps during training.
-        """
-        t_0 = 1.0 / self.noise_schedule.total_N if t_start is None else t_start
-        t_T = self.noise_schedule.T if t_end is None else t_end
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-        return self.sample(
-            x,
-            steps=steps,
-            t_start=t_0,
-            t_end=t_T,
-            order=order,
-            skip_type=skip_type,
-            method=method,
-            lower_order_final=lower_order_final,
-            denoise_to_zero=denoise_to_zero,
-            solver_type=solver_type,
-            atol=atol,
-            rtol=rtol,
-            return_intermediate=return_intermediate,
-        )
-
-    def sample(
-        self,
-        x,
-        steps=20,
-        t_start=None,
-        t_end=None,
-        order=2,
-        skip_type="time_uniform",
-        method="multistep",
-        lower_order_final=True,
-        denoise_to_zero=False,
-        solver_type="dpmsolver",
-        atol=0.0078,
-        rtol=0.05,
-        return_intermediate=False,
-        flow_shift=1.0,
-    ):
-        """
-        Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
-
-        =====================================================
-
-        We support the following algorithms for both noise prediction model and data prediction model:
-            - 'singlestep':
-                Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
-                We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
-                The total number of function evaluations (NFE) == `steps`.
-                Given a fixed NFE == `steps`, the sampling procedure is:
-                    - If `order` == 1:
-                        - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
-                    - If `order` == 2:
-                        - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
-                        - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
-                        - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
-                    - If `order` == 3:
-                        - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
-                        - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
-                        - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
-                        - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
-            - 'multistep':
-                Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
-                We initialize the first `order` values by lower order multistep solvers.
-                Given a fixed NFE == `steps`, the sampling procedure is:
-                    Denote K = steps.
-                    - If `order` == 1:
-                        - We use K steps of DPM-Solver-1 (i.e. DDIM).
-                    - If `order` == 2:
-                        - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
-                    - If `order` == 3:
-                        - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
-            - 'singlestep_fixed':
-                Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
-                We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
-            - 'adaptive':
-                Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
-                We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
-                You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
-                (NFE) and the sample quality.
-                    - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
-                    - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
-
-        =====================================================
-
-        Some advices for choosing the algorithm:
-            - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
-                Use singlestep DPM-Solver or DPM-Solver++ ("DPM-Solver-fast" in the paper) with `order = 3`.
-                e.g., DPM-Solver:
-                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver")
-                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
-                            skip_type='time_uniform', method='singlestep')
-                e.g., DPM-Solver++:
-                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
-                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
-                            skip_type='time_uniform', method='singlestep')
-            - For **guided sampling with large guidance scale** by DPMs:
-                Use multistep DPM-Solver with `algorithm_type="dpmsolver++"` and `order = 2`.
-                e.g.
-                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
-                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
-                            skip_type='time_uniform', method='multistep')
-
-        We support three types of `skip_type`:
-            - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
-            - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
-            - 'time_quadratic': quadratic time for the time steps.
-
-        =====================================================
-        Args:
-            x: A pytorch tensor. The initial value at time `t_start`
-                e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
-            steps: A `int`. The total number of function evaluations (NFE).
-            t_start: A `float`. The starting time of the sampling.
-                If `T` is None, we use self.noise_schedule.T (default is 1.0).
-            t_end: A `float`. The ending time of the sampling.
-                If `t_end` is None, we use 1. / self.noise_schedule.total_N.
-                e.g. if total_N == 1000, we have `t_end` == 1e-3.
-                For discrete-time DPMs:
-                    - We recommend `t_end` == 1. / self.noise_schedule.total_N.
-                For continuous-time DPMs:
-                    - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
-            order: A `int`. The order of DPM-Solver.
-            skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
-            method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
-            denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
-                Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
-
-                This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
-                score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
-                for diffusion models sampling by diffusion SDEs for low-resolutional images
-                (such as CIFAR-10). However, we observed that such trick does not matter for
-                high-resolutional images. As it needs an additional NFE, we do not recommend
-                it for high-resolutional images.
-            lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
-                Only valid for `method=multistep` and `steps < 15`. We empirically find that
-                this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
-                (especially for steps <= 10). So we recommend to set it to be `True`.
-            solver_type: A `str`. The taylor expansion type for the solver. `dpmsolver` or `taylor`. We recommend `dpmsolver`.
-            atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
-            rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
-            return_intermediate: A `bool`. Whether to save the xt at each step.
-                When set to `True`, method returns a tuple (x0, intermediates); when set to False, method returns only x0.
-        Returns:
-            x_end: A pytorch tensor. The approximated solution at time `t_end`.
-
-        """
-        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
-        t_T = self.noise_schedule.T if t_start is None else t_start
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-        if return_intermediate:
-            assert method in [
-                "multistep",
-                "singlestep",
-                "singlestep_fixed",
-            ], "Cannot use adaptive solver when saving intermediate values"
-        if self.correcting_xt_fn is not None:
-            assert method in [
-                "multistep",
-                "singlestep",
-                "singlestep_fixed",
-            ], "Cannot use adaptive solver when correcting_xt_fn is not None"
-        device = x.device
-        intermediates = []
-        with torch.no_grad():
-            if method == "adaptive":
-                x = self.dpm_solver_adaptive(
-                    x,
-                    order=order,
-                    t_T=t_T,
-                    t_0=t_0,
-                    atol=atol,
-                    rtol=rtol,
-                    solver_type=solver_type,
-                )
-            elif method == "multistep":
-                assert steps >= order
-                timesteps = self.get_time_steps(
-                    skip_type=skip_type,
-                    t_T=t_T,
-                    t_0=t_0,
-                    N=steps,
-                    device=device,
-                    shift=flow_shift,
-                )
-                assert timesteps.shape[0] - 1 == steps
-                # Init the initial values.
-                step = 0
-                t = timesteps[step]
-                t_prev_list = [t]
-                model_prev_list = [self.model_fn(x, t)]
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-                self.update_progress(step + 1, len(timesteps))
-                # Init the first `order` values by lower order multistep DPM-Solver.
-                for step in range(1, order):
-                    t = timesteps[step]
-                    x = self.multistep_dpm_solver_update(
-                        x,
-                        model_prev_list,
-                        t_prev_list,
-                        t,
-                        step,
-                        solver_type=solver_type,
-                    )
-                    if self.correcting_xt_fn is not None:
-                        x = self.correcting_xt_fn(x, t, step)
-                    if return_intermediate:
-                        intermediates.append(x)
-                    t_prev_list.append(t)
-                    model_prev_list.append(self.model_fn(x, t))
-                    # update progress bar
-                    self.update_progress(step + 1, len(timesteps))
-                # Compute the remaining values by `order`-th order multistep DPM-Solver.
-                for step in tqdm(
-                    range(order, steps + 1),
-                    disable=os.getenv("DPM_TQDM", "False") == "True",
-                ):
-                    t = timesteps[step]
-                    # We only use lower order for steps < 10
-                    # if lower_order_final and steps < 10:
-                    if lower_order_final:  # recommended by Shuchen Xue
-                        step_order = min(order, steps + 1 - step)
-                    else:
-                        step_order = order
-                    x = self.multistep_dpm_solver_update(
-                        x,
-                        model_prev_list,
-                        t_prev_list,
-                        t,
-                        step_order,
-                        solver_type=solver_type,
-                    )
-                    if self.correcting_xt_fn is not None:
-                        x = self.correcting_xt_fn(x, t, step)
-                    if return_intermediate:
-                        intermediates.append(x)
-                    for i in range(order - 1):
-                        t_prev_list[i] = t_prev_list[i + 1]
-                        model_prev_list[i] = model_prev_list[i + 1]
-                    t_prev_list[-1] = t
-                    # We do not need to evaluate the final model value.
-                    if step < steps:
-                        model_prev_list[-1] = self.model_fn(x, t)
-                    # update progress bar
-                    self.update_progress(step + 1, len(timesteps))
-            elif method in ["singlestep", "singlestep_fixed"]:
-                if method == "singlestep":
-                    timesteps_outer, orders = (
-                        self.get_orders_and_timesteps_for_singlestep_solver(
-                            steps=steps,
-                            order=order,
-                            skip_type=skip_type,
-                            t_T=t_T,
-                            t_0=t_0,
-                            device=device,
-                        )
-                    )
-                elif method == "singlestep_fixed":
-                    K = steps // order
-                    orders = [
-                        order,
-                    ] * K
-                    timesteps_outer = self.get_time_steps(
-                        skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device
-                    )
-                for step, order in enumerate(orders):
-                    s, t = timesteps_outer[step], timesteps_outer[step + 1]
-                    timesteps_inner = self.get_time_steps(
-                        skip_type=skip_type,
-                        t_T=s.item(),
-                        t_0=t.item(),
-                        N=order,
-                        device=device,
-                    )
-                    lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
-                    h = lambda_inner[-1] - lambda_inner[0]
-                    r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
-                    r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
-                    x = self.singlestep_dpm_solver_update(
-                        x, s, t, order, solver_type=solver_type, r1=r1, r2=r2
-                    )
-                    if self.correcting_xt_fn is not None:
-                        x = self.correcting_xt_fn(x, t, step)
-                    if return_intermediate:
-                        intermediates.append(x)
-                    self.update_progress(step + 1, len(timesteps_outer))
-            else:
-                raise ValueError(f"Got wrong method {method}")
-            if denoise_to_zero:
-                t = torch.ones((1,)).to(device) * t_0
-                x = self.denoise_to_zero_fn(x, t)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step + 1)
-                if return_intermediate:
-                    intermediates.append(x)
-        if return_intermediate:
-            return x, intermediates
-        else:
-            return x
-
-
-#############################################################
-# other utility functions
-#############################################################
-
-
-def interpolate_fn(x, xp, yp):
-    """
-    A piecewise linear function y = f(x), using xp and yp as keypoints.
-    We implement f(x) in a differentiable way (i.e. applicable for autograd).
-    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
-
-    Args:
-        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
-        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
-        yp: PyTorch tensor with shape [C, K].
-    Returns:
-        The function values f(x), with shape [N, C].
-    """
-    N, K = x.shape[0], xp.shape[1]
-    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
-    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
-    x_idx = torch.argmin(x_indices, dim=2)
-    cand_start_idx = x_idx - 1
-    start_idx = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(1, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    end_idx = torch.where(
-        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
-    )
-    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
-    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
-    start_idx2 = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(0, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
-    start_y = torch.gather(
-        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
-    ).squeeze(2)
-    end_y = torch.gather(
-        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
-    ).squeeze(2)
-    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
-    return cand
-
-
-def expand_dims(v, dims):
-    """
-    Expand the tensor `v` to the dim `dims`.
-
-    Args:
-        `v`: a PyTorch tensor with shape [N].
-        `dim`: a `int`.
-    Returns:
-        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
-    """
-    return v[(...,) + (None,) * (dims - 1)]
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/edm_sample.py b/sana/sana_1600M/packages/Sana/diffusion/model/edm_sample.py
deleted file mode 100755
index 9e99301eb..000000000
--- a/sana/sana_1600M/packages/Sana/diffusion/model/edm_sample.py
+++ /dev/null
@@ -1,281 +0,0 @@
-# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-
-# Modified from OpenAI's diffusion repos
-#     GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py
-#     ADM:   https://github.com/openai/guided-diffusion/blob/main/guided_diffusion
-#     IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
-
-
-import numpy as np
-from diffusion.model.utils import *
-from tqdm import tqdm
-
-# ----------------------------------------------------------------------------
-# Proposed EDM sampler (Algorithm 2).
-
-
-def edm_sampler(
-    net,
-    latents,
-    class_labels=None,
-    cfg_scale=None,
-    randn_like=torch.randn_like,
-    num_steps=18,
-    sigma_min=0.002,
-    sigma_max=80,
-    rho=7,
-    S_churn=0,
-    S_min=0,
-    S_max=float("inf"),
-    S_noise=1,
-    **kwargs,
-):
-    # Adjust noise levels based on what's supported by the network.
-    sigma_min = max(sigma_min, net.sigma_min)
-    sigma_max = min(sigma_max, net.sigma_max)
-
-    # Time step discretization.
-    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
-    t_steps = (
-        sigma_max ** (1 / rho)
-        + step_indices
-        / (num_steps - 1)
-        * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
-    ) ** rho
-    t_steps = torch.cat(
-        [net.round_sigma(t_steps), torch.zeros_like(t_steps[:1])]
-    )  # t_N = 0
-
-    # Main sampling loop.
-    x_next = latents.to(torch.float64) * t_steps[0]
-    for i, (t_cur, t_next) in tqdm(
-        list(enumerate(zip(t_steps[:-1], t_steps[1:])))
-    ):  # 0, ..., N-1
-        x_cur = x_next
-
-        # Increase noise temporarily.
-        gamma = (
-            min(S_churn / num_steps, np.sqrt(2) - 1) if S_min <= t_cur <= S_max else 0
-        )
-        t_hat = net.round_sigma(t_cur + gamma * t_cur)
-        x_hat = x_cur + (t_hat**2 - t_cur**2).sqrt() * S_noise * randn_like(x_cur)
-
-        # Euler step.
-        denoised = net(x_hat.float(), t_hat, class_labels, cfg_scale, **kwargs)["x"].to(
-            torch.float64
-        )
-        d_cur = (x_hat - denoised) / t_hat
-        x_next = x_hat + (t_next - t_hat) * d_cur
-
-        # Apply 2nd order correction.
-        if i < num_steps - 1:
-            denoised = net(x_next.float(), t_next, class_labels, cfg_scale, **kwargs)[
-                "x"
-            ].to(torch.float64)
-            d_prime = (x_next - denoised) / t_next
-            x_next = x_hat + (t_next - t_hat) * (0.5 * d_cur + 0.5 * d_prime)
-
-    return x_next
-
-
-# ----------------------------------------------------------------------------
-# Generalized ablation sampler, representing the superset of all sampling
-# methods discussed in the paper.
-
-
-def ablation_sampler(
-    net,
-    latents,
-    class_labels=None,
-    cfg_scale=None,
-    feat=None,
-    randn_like=torch.randn_like,
-    num_steps=18,
-    sigma_min=None,
-    sigma_max=None,
-    rho=7,
-    solver="heun",
-    discretization="edm",
-    schedule="linear",
-    scaling="none",
-    epsilon_s=1e-3,
-    C_1=0.001,
-    C_2=0.008,
-    M=1000,
-    alpha=1,
-    S_churn=0,
-    S_min=0,
-    S_max=float("inf"),
-    S_noise=1,
-):
-    assert solver in ["euler", "heun"]
-    assert discretization in ["vp", "ve", "iddpm", "edm"]
-    assert schedule in ["vp", "ve", "linear"]
-    assert scaling in ["vp", "none"]
-
-    # Helper functions for VP & VE noise level schedules.
-    vp_sigma = (
-        lambda beta_d, beta_min: lambda t: (
-            np.e ** (0.5 * beta_d * (t**2) + beta_min * t) - 1
-        )
-        ** 0.5
-    )
-    vp_sigma_deriv = (
-        lambda beta_d, beta_min: lambda t: 0.5
-        * (beta_min + beta_d * t)
-        * (sigma(t) + 1 / sigma(t))
-    )
-    vp_sigma_inv = (
-        lambda beta_d, beta_min: lambda sigma: (
-            (beta_min**2 + 2 * beta_d * (sigma**2 + 1).log()).sqrt() - beta_min
-        )
-        / beta_d
-    )
-    ve_sigma = lambda t: t.sqrt()
-    ve_sigma_deriv = lambda t: 0.5 / t.sqrt()
-    ve_sigma_inv = lambda sigma: sigma**2
-
-    # Select default noise level range based on the specified time step discretization.
-    if sigma_min is None:
-        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=epsilon_s)
-        sigma_min = {"vp": vp_def, "ve": 0.02, "iddpm": 0.002, "edm": 0.002}[
-            discretization
-        ]
-    if sigma_max is None:
-        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=1)
-        sigma_max = {"vp": vp_def, "ve": 100, "iddpm": 81, "edm": 80}[discretization]
-
-    # Adjust noise levels based on what's supported by the network.
-    sigma_min = max(sigma_min, net.sigma_min)
-    sigma_max = min(sigma_max, net.sigma_max)
-
-    # Compute corresponding betas for VP.
-    vp_beta_d = (
-        2
-        * (np.log(sigma_min**2 + 1) / epsilon_s - np.log(sigma_max**2 + 1))
-        / (epsilon_s - 1)
-    )
-    vp_beta_min = np.log(sigma_max**2 + 1) - 0.5 * vp_beta_d
-
-    # Define time steps in terms of noise level.
-    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
-    if discretization == "vp":
-        orig_t_steps = 1 + step_indices / (num_steps - 1) * (epsilon_s - 1)
-        sigma_steps = vp_sigma(vp_beta_d, vp_beta_min)(orig_t_steps)
-    elif discretization == "ve":
-        orig_t_steps = (sigma_max**2) * (
-            (sigma_min**2 / sigma_max**2) ** (step_indices / (num_steps - 1))
-        )
-        sigma_steps = ve_sigma(orig_t_steps)
-    elif discretization == "iddpm":
-        u = torch.zeros(M + 1, dtype=torch.float64, device=latents.device)
-        alpha_bar = lambda j: (0.5 * np.pi * j / M / (C_2 + 1)).sin() ** 2
-        for j in torch.arange(M, 0, -1, device=latents.device):  # M, ..., 1
-            u[j - 1] = (
-                (u[j] ** 2 + 1) / (alpha_bar(j - 1) / alpha_bar(j)).clip(min=C_1) - 1
-            ).sqrt()
-        u_filtered = u[torch.logical_and(u >= sigma_min, u <= sigma_max)]
-        sigma_steps = u_filtered[
-            ((len(u_filtered) - 1) / (num_steps - 1) * step_indices)
-            .round()
-            .to(torch.int64)
-        ]
-    else:
-        assert discretization == "edm"
-        sigma_steps = (
-            sigma_max ** (1 / rho)
-            + step_indices
-            / (num_steps - 1)
-            * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
-        ) ** rho
-
-    # Define noise level schedule.
-    if schedule == "vp":
-        sigma = vp_sigma(vp_beta_d, vp_beta_min)
-        sigma_deriv = vp_sigma_deriv(vp_beta_d, vp_beta_min)
-        sigma_inv = vp_sigma_inv(vp_beta_d, vp_beta_min)
-    elif schedule == "ve":
-        sigma = ve_sigma
-        sigma_deriv = ve_sigma_deriv
-        sigma_inv = ve_sigma_inv
-    else:
-        assert schedule == "linear"
-        sigma = lambda t: t
-        sigma_deriv = lambda t: 1
-        sigma_inv = lambda sigma: sigma
-
-    # Define scaling schedule.
-    if scaling == "vp":
-        s = lambda t: 1 / (1 + sigma(t) ** 2).sqrt()
-        s_deriv = lambda t: -sigma(t) * sigma_deriv(t) * (s(t) ** 3)
-    else:
-        assert scaling == "none"
-        s = lambda t: 1
-        s_deriv = lambda t: 0
-
-    # Compute final time steps based on the corresponding noise levels.
-    t_steps = sigma_inv(net.round_sigma(sigma_steps))
-    t_steps = torch.cat([t_steps, torch.zeros_like(t_steps[:1])])  # t_N = 0
-
-    # Main sampling loop.
-    t_next = t_steps[0]
-    x_next = latents.to(torch.float64) * (sigma(t_next) * s(t_next))
-    for i, (t_cur, t_next) in enumerate(zip(t_steps[:-1], t_steps[1:])):  # 0, ..., N-1
-        x_cur = x_next
-
-        # Increase noise temporarily.
-        gamma = (
-            min(S_churn / num_steps, np.sqrt(2) - 1)
-            if S_min <= sigma(t_cur) <= S_max
-            else 0
-        )
-        t_hat = sigma_inv(net.round_sigma(sigma(t_cur) + gamma * sigma(t_cur)))
-        x_hat = s(t_hat) / s(t_cur) * x_cur + (
-            sigma(t_hat) ** 2 - sigma(t_cur) ** 2
-        ).clip(min=0).sqrt() * s(t_hat) * S_noise * randn_like(x_cur)
-
-        # Euler step.
-        h = t_next - t_hat
-        denoised = net(
-            x_hat.float() / s(t_hat), sigma(t_hat), class_labels, cfg_scale, feat=feat
-        )["x"].to(torch.float64)
-        d_cur = (
-            sigma_deriv(t_hat) / sigma(t_hat) + s_deriv(t_hat) / s(t_hat)
-        ) * x_hat - sigma_deriv(t_hat) * s(t_hat) / sigma(t_hat) * denoised
-        x_prime = x_hat + alpha * h * d_cur
-        t_prime = t_hat + alpha * h
-
-        # Apply 2nd order correction.
-        if solver == "euler" or i == num_steps - 1:
-            x_next = x_hat + h * d_cur
-        else:
-            assert solver == "heun"
-            denoised = net(
-                x_prime.float() / s(t_prime),
-                sigma(t_prime),
-                class_labels,
-                cfg_scale,
-                feat=feat,
-            )["x"].to(torch.float64)
-            d_prime = (
-                sigma_deriv(t_prime) / sigma(t_prime) + s_deriv(t_prime) / s(t_prime)
-            ) * x_prime - sigma_deriv(t_prime) * s(t_prime) / sigma(t_prime) * denoised
-            x_next = x_hat + h * (
-                (1 - 1 / (2 * alpha)) * d_cur + 1 / (2 * alpha) * d_prime
-            )
-
-    return x_next
diff --git a/sana/sana_1600M/packages/Sana/diffusion/model/sa_solver.py b/sana/sana_1600M/packages/Sana/diffusion/model/sa_solver.py
deleted file mode 100755
index b7566cef5..000000000
--- a/sana/sana_1600M/packages/Sana/diffusion/model/sa_solver.py
+++ /dev/null
@@ -1,1618 +0,0 @@
-# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-
-import math
-
-import torch
-from tqdm import tqdm
-
-
-class NoiseScheduleVP:
-    def __init__(
-        self,
-        schedule="discrete",
-        betas=None,
-        alphas_cumprod=None,
-        continuous_beta_0=0.1,
-        continuous_beta_1=20.0,
-        dtype=torch.float32,
-    ):
-        """Thanks to DPM-Solver for their code base"""
-        r"""Create a wrapper class for the forward SDE (VP type).
-        ***
-        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
-                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
-        ***
-        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
-        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
-        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
-            log_alpha_t = self.marginal_log_mean_coeff(t)
-            sigma_t = self.marginal_std(t)
-            lambda_t = self.marginal_lambda(t)
-        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
-            t = self.inverse_lambda(lambda_t)
-        ===============================================================
-        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
-        1. For discrete-time DPMs:
-            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
-                t_i = (i + 1) / N
-            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
-            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
-            Args:
-                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
-                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
-            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
-            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
-                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
-                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
-                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
-                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
-                and
-                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
-        2. For continuous-time DPMs:
-            We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
-            schedule are the default settings in DDPM and improved-DDPM:
-            Args:
-                beta_min: A `float` number. The smallest beta for the linear schedule.
-                beta_max: A `float` number. The largest beta for the linear schedule.
-                cosine_s: A `float` number. The hyperparameter in the cosine schedule.
-                cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
-                T: A `float` number. The ending time of the forward process.
-        ===============================================================
-        Args:
-            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
-                    'linear' or 'cosine' for continuous-time DPMs.
-        Returns:
-            A wrapper object of the forward SDE (VP type).
-
-        ===============================================================
-        Example:
-        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', betas=betas)
-        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
-        # For continuous-time DPMs (VPSDE), linear schedule:
-        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
-        """
-
-        if schedule not in ["discrete", "linear", "cosine"]:
-            raise ValueError(
-                "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(
-                    schedule
-                )
-            )
-
-        self.schedule = schedule
-        if schedule == "discrete":
-            if betas is not None:
-                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
-            else:
-                assert alphas_cumprod is not None
-                log_alphas = 0.5 * torch.log(alphas_cumprod)
-            self.total_N = len(log_alphas)
-            self.T = 1.0
-            self.t_array = (
-                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
-                .reshape((1, -1))
-                .to(dtype=dtype)
-            )
-            self.log_alpha_array = log_alphas.reshape(
-                (
-                    1,
-                    -1,
-                )
-            ).to(dtype=dtype)
-        else:
-            self.total_N = 1000
-            self.beta_0 = continuous_beta_0
-            self.beta_1 = continuous_beta_1
-            self.cosine_s = 0.008
-            self.cosine_beta_max = 999.0
-            self.cosine_t_max = (
-                math.atan(self.cosine_beta_max * (1.0 + self.cosine_s) / math.pi)
-                * 2.0
-                * (1.0 + self.cosine_s)
-                / math.pi
-                - self.cosine_s
-            )
-            self.cosine_log_alpha_0 = math.log(
-                math.cos(self.cosine_s / (1.0 + self.cosine_s) * math.pi / 2.0)
-            )
-            self.schedule = schedule
-            if schedule == "cosine":
-                # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
-                # Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
-                self.T = 0.9946
-            else:
-                self.T = 1.0
-
-    def marginal_log_mean_coeff(self, t):
-        """
-        Compute log(alpha_t) of a given continuous-time label t in [0, T].
-        """
-        if self.schedule == "discrete":
-            return interpolate_fn(
-                t.reshape((-1, 1)),
-                self.t_array.to(t.device),
-                self.log_alpha_array.to(t.device),
-            ).reshape(-1)
-        elif self.schedule == "linear":
-            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
-        elif self.schedule == "cosine":
-            log_alpha_fn = lambda s: torch.log(
-                torch.cos((s + self.cosine_s) / (1.0 + self.cosine_s) * math.pi / 2.0)
-            )
-            log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
-            return log_alpha_t
-
-    def marginal_alpha(self, t):
-        """
-        Compute alpha_t of a given continuous-time label t in [0, T].
-        """
-        return torch.exp(self.marginal_log_mean_coeff(t))
-
-    def marginal_std(self, t):
-        """
-        Compute sigma_t of a given continuous-time label t in [0, T].
-        """
-        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
-
-    def marginal_lambda(self, t):
-        """
-        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
-        """
-        log_mean_coeff = self.marginal_log_mean_coeff(t)
-        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
-        return log_mean_coeff - log_std
-
-    def inverse_lambda(self, lamb):
-        """
-        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
-        """
-        if self.schedule == "linear":
-            tmp = (
-                2.0
-                * (self.beta_1 - self.beta_0)
-                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
-            )
-            Delta = self.beta_0**2 + tmp
-            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
-        elif self.schedule == "discrete":
-            log_alpha = -0.5 * torch.logaddexp(
-                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
-            )
-            t = interpolate_fn(
-                log_alpha.reshape((-1, 1)),
-                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
-                torch.flip(self.t_array.to(lamb.device), [1]),
-            )
-            return t.reshape((-1,))
-        else:
-            log_alpha = -0.5 * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
-            t_fn = (
-                lambda log_alpha_t: torch.arccos(
-                    torch.exp(log_alpha_t + self.cosine_log_alpha_0)
-                )
-                * 2.0
-                * (1.0 + self.cosine_s)
-                / math.pi
-                - self.cosine_s
-            )
-            t = t_fn(log_alpha)
-            return t
-
-    def edm_sigma(self, t):
-        return self.marginal_std(t) / self.marginal_alpha(t)
-
-    def edm_inverse_sigma(self, edmsigma):
-        alpha = 1 / (edmsigma**2 + 1).sqrt()
-        sigma = alpha * edmsigma
-        lambda_t = torch.log(alpha / sigma)
-        t = self.inverse_lambda(lambda_t)
-        return t
-
-
-def model_wrapper(
-    model,
-    noise_schedule,
-    model_type="noise",
-    model_kwargs={},
-    guidance_type="uncond",
-    condition=None,
-    unconditional_condition=None,
-    guidance_scale=1.0,
-    classifier_fn=None,
-    classifier_kwargs={},
-):
-    """Thanks to DPM-Solver for their code base"""
-    """Create a wrapper function for the noise prediction model.
-    SA-Solver needs to solve the continuous-time diffusion SDEs. For DPMs trained on discrete-time labels, we need to
-    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
-    We support four types of the diffusion model by setting `model_type`:
-        1. "noise": noise prediction model. (Trained by predicting noise).
-        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
-        3. "v": velocity prediction model. (Trained by predicting the velocity).
-            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
-            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
-                arXiv preprint arXiv:2202.00512 (2022).
-            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
-                arXiv preprint arXiv:2210.02303 (2022).
-
-        4. "score": marginal score function. (Trained by denoising score matching).
-            Note that the score function and the noise prediction model follows a simple relationship:
-            ```
-                noise(x_t, t) = -sigma_t * score(x_t, t)
-            ```
-    We support three types of guided sampling by DPMs by setting `guidance_type`:
-        1. "uncond": unconditional sampling by DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-            The input `classifier_fn` has the following format:
-            ``
-                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
-            ``
-            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
-                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
-        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
-            ``
-            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
-            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
-                arXiv preprint arXiv:2207.12598 (2022).
-
-    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
-    or continuous-time labels (i.e. epsilon to T).
-    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
-    ``
-        def model_fn(x, t_continuous) -> noise:
-            t_input = get_model_input_time(t_continuous)
-            return noise_pred(model, x, t_input, **model_kwargs)
-    ``
-    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for SA-Solver.
-    ===============================================================
-    Args:
-        model: A diffusion model with the corresponding format described above.
-        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
-        model_type: A `str`. The parameterization type of the diffusion model.
-                    "noise" or "x_start" or "v" or "score".
-        model_kwargs: A `dict`. A dict for the other inputs of the model function.
-        guidance_type: A `str`. The type of the guidance for sampling.
-                    "uncond" or "classifier" or "classifier-free".
-        condition: A pytorch tensor. The condition for the guided sampling.
-                    Only used for "classifier" or "classifier-free" guidance type.
-        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
-                    Only used for "classifier-free" guidance type.
-        guidance_scale: A `float`. The scale for the guided sampling.
-        classifier_fn: A classifier function. Only used for the classifier guidance.
-        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
-    Returns:
-        A noise prediction model that accepts the noised data and the continuous time as the inputs.
-    """
-
-    def get_model_input_time(t_continuous):
-        """
-        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
-        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
-        For continuous-time DPMs, we just use `t_continuous`.
-        """
-        if noise_schedule.schedule == "discrete":
-            return (t_continuous - 1.0 / noise_schedule.total_N) * 1000.0
-        else:
-            return t_continuous
-
-    def noise_pred_fn(x, t_continuous, cond=None):
-        t_input = get_model_input_time(t_continuous)
-        if cond is None:
-            output = model(x, t_input, **model_kwargs)
-        else:
-            output = model(x, t_input, cond, **model_kwargs)
-        if model_type == "noise":
-            return output
-        elif model_type == "x_start":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return (x - alpha_t[0] * output) / sigma_t[0]
-        elif model_type == "v":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return alpha_t[0] * output + sigma_t[0] * x
-        elif model_type == "score":
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            return -sigma_t[0] * output
-
-    def cond_grad_fn(x, t_input):
-        """
-        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
-        """
-        with torch.enable_grad():
-            x_in = x.detach().requires_grad_(True)
-            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
-            return torch.autograd.grad(log_prob.sum(), x_in)[0]
-
-    def model_fn(x, t_continuous):
-        """
-        The noise predicition model function that is used for DPM-Solver.
-        """
-        if guidance_type == "uncond":
-            return noise_pred_fn(x, t_continuous)
-        elif guidance_type == "classifier":
-            assert classifier_fn is not None
-            t_input = get_model_input_time(t_continuous)
-            cond_grad = cond_grad_fn(x, t_input)
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            noise = noise_pred_fn(x, t_continuous)
-            return noise - guidance_scale * sigma_t * cond_grad
-        elif guidance_type == "classifier-free":
-            if guidance_scale == 1.0 or unconditional_condition is None:
-                return noise_pred_fn(x, t_continuous, cond=condition)
-            else:
-                x_in = torch.cat([x] * 2)
-                t_in = torch.cat([t_continuous] * 2)
-                c_in = torch.cat([unconditional_condition, condition])
-                noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
-                return noise_uncond + guidance_scale * (noise - noise_uncond)
-
-    assert model_type in ["noise", "x_start", "v", "score"]
-    assert guidance_type in ["uncond", "classifier", "classifier-free"]
-    return model_fn
-
-
-class SASolver:
-    def __init__(
-        self,
-        model_fn,
-        noise_schedule,
-        algorithm_type="data_prediction",
-        correcting_x0_fn=None,
-        correcting_xt_fn=None,
-        thresholding_max_val=1.0,
-        dynamic_thresholding_ratio=0.995,
-    ):
-        """
-        Construct a SA-Solver
-        The default value for algorithm_type is "data_prediction" and we recommend not to change it to
-        "noise_prediction". For details, please see Appendix A.2.4 in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
-        self.noise_schedule = noise_schedule
-        assert algorithm_type in ["data_prediction", "noise_prediction"]
-
-        if correcting_x0_fn == "dynamic_thresholding":
-            self.correcting_x0_fn = self.dynamic_thresholding_fn
-        else:
-            self.correcting_x0_fn = correcting_x0_fn
-
-        self.correcting_xt_fn = correcting_xt_fn
-        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
-        self.thresholding_max_val = thresholding_max_val
-
-        self.predict_x0 = algorithm_type == "data_prediction"
-
-        self.sigma_min = float(self.noise_schedule.edm_sigma(torch.tensor([1e-3])))
-        self.sigma_max = float(self.noise_schedule.edm_sigma(torch.tensor([1])))
-
-    def dynamic_thresholding_fn(self, x0, t=None):
-        """
-        The dynamic thresholding method.
-        """
-        dims = x0.dim()
-        p = self.dynamic_thresholding_ratio
-        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
-        s = expand_dims(
-            torch.maximum(
-                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
-            ),
-            dims,
-        )
-        x0 = torch.clamp(x0, -s, s) / s
-        return x0
-
-    def noise_prediction_fn(self, x, t):
-        """
-        Return the noise prediction model.
-        """
-        return self.model(x, t)
-
-    def data_prediction_fn(self, x, t):
-        """
-        Return the data prediction model (with corrector).
-        """
-        noise = self.noise_prediction_fn(x, t)
-        alpha_t, sigma_t = (
-            self.noise_schedule.marginal_alpha(t),
-            self.noise_schedule.marginal_std(t),
-        )
-        x0 = (x - sigma_t * noise) / alpha_t
-        if self.correcting_x0_fn is not None:
-            x0 = self.correcting_x0_fn(x0)
-        return x0
-
-    def model_fn(self, x, t):
-        """
-        Convert the model to the noise prediction model or the data prediction model.
-        """
-
-        if self.predict_x0:
-            return self.data_prediction_fn(x, t)
-        else:
-            return self.noise_prediction_fn(x, t)
-
-    def get_time_steps(self, skip_type, t_T, t_0, N, order, device):
-        """Compute the intermediate time steps for sampling."""
-        if skip_type == "logSNR":
-            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
-            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
-            logSNR_steps = lambda_T + torch.linspace(
-                torch.tensor(0.0).cpu().item(),
-                (lambda_0 - lambda_T).cpu().item() ** (1.0 / order),
-                N + 1,
-            ).pow(order).to(device)
-            return self.noise_schedule.inverse_lambda(logSNR_steps)
-        elif skip_type == "time":
-            t = (
-                torch.linspace(t_T ** (1.0 / order), t_0 ** (1.0 / order), N + 1)
-                .pow(order)
-                .to(device)
-            )
-            return t
-        elif skip_type == "karras":
-            sigma_min = max(0.002, self.sigma_min)
-            sigma_max = min(80, self.sigma_max)
-            sigma_steps = (
-                torch.linspace(sigma_max ** (1.0 / 7), sigma_min ** (1.0 / 7), N + 1)
-                .pow(7)
-                .to(device)
-            )
-            t = self.noise_schedule.edm_inverse_sigma(sigma_steps)
-            return t
-        else:
-            raise ValueError(
-                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time' or 'karras'"
-            )
-
-    def denoise_to_zero_fn(self, x, s):
-        """
-        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
-        """
-        return self.data_prediction_fn(x, s)
-
-    def get_coefficients_exponential_negative(
-        self, order, interval_start, interval_end
-    ):
-        """
-        Calculate the integral of exp(-x) * x^order dx from interval_start to interval_end
-        For calculating the coefficient of gradient terms after the lagrange interpolation,
-        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        For noise_prediction formula.
-        """
-        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
-
-        if order == 0:
-            return torch.exp(-interval_end) * (
-                torch.exp(interval_end - interval_start) - 1
-            )
-        elif order == 1:
-            return torch.exp(-interval_end) * (
-                (interval_start + 1) * torch.exp(interval_end - interval_start)
-                - (interval_end + 1)
-            )
-        elif order == 2:
-            return torch.exp(-interval_end) * (
-                (interval_start**2 + 2 * interval_start + 2)
-                * torch.exp(interval_end - interval_start)
-                - (interval_end**2 + 2 * interval_end + 2)
-            )
-        elif order == 3:
-            return torch.exp(-interval_end) * (
-                (interval_start**3 + 3 * interval_start**2 + 6 * interval_start + 6)
-                * torch.exp(interval_end - interval_start)
-                - (interval_end**3 + 3 * interval_end**2 + 6 * interval_end + 6)
-            )
-
-    def get_coefficients_exponential_positive(
-        self, order, interval_start, interval_end, tau
-    ):
-        """
-        Calculate the integral of exp(x(1+tau^2)) * x^order dx from interval_start to interval_end
-        For calculating the coefficient of gradient terms after the lagrange interpolation,
-        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        For data_prediction formula.
-        """
-        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
-
-        # after change of variable(cov)
-        interval_end_cov = (1 + tau**2) * interval_end
-        interval_start_cov = (1 + tau**2) * interval_start
-
-        if order == 0:
-            return (
-                torch.exp(interval_end_cov)
-                * (1 - torch.exp(-(interval_end_cov - interval_start_cov)))
-                / (1 + tau**2)
-            )
-        elif order == 1:
-            return (
-                torch.exp(interval_end_cov)
-                * (
-                    (interval_end_cov - 1)
-                    - (interval_start_cov - 1)
-                    * torch.exp(-(interval_end_cov - interval_start_cov))
-                )
-                / ((1 + tau**2) ** 2)
-            )
-        elif order == 2:
-            return (
-                torch.exp(interval_end_cov)
-                * (
-                    (interval_end_cov**2 - 2 * interval_end_cov + 2)
-                    - (interval_start_cov**2 - 2 * interval_start_cov + 2)
-                    * torch.exp(-(interval_end_cov - interval_start_cov))
-                )
-                / ((1 + tau**2) ** 3)
-            )
-        elif order == 3:
-            return (
-                torch.exp(interval_end_cov)
-                * (
-                    (
-                        interval_end_cov**3
-                        - 3 * interval_end_cov**2
-                        + 6 * interval_end_cov
-                        - 6
-                    )
-                    - (
-                        interval_start_cov**3
-                        - 3 * interval_start_cov**2
-                        + 6 * interval_start_cov
-                        - 6
-                    )
-                    * torch.exp(-(interval_end_cov - interval_start_cov))
-                )
-                / ((1 + tau**2) ** 4)
-            )
-
-    def lagrange_polynomial_coefficient(self, order, lambda_list):
-        """
-        Calculate the coefficient of lagrange polynomial
-        For lagrange interpolation
-        """
-        assert order in [0, 1, 2, 3]
-        assert order == len(lambda_list) - 1
-        if order == 0:
-            return [[1]]
-        elif order == 1:
-            return [
-                [
-                    1 / (lambda_list[0] - lambda_list[1]),
-                    -lambda_list[1] / (lambda_list[0] - lambda_list[1]),
-                ],
-                [
-                    1 / (lambda_list[1] - lambda_list[0]),
-                    -lambda_list[0] / (lambda_list[1] - lambda_list[0]),
-                ],
-            ]
-        elif order == 2:
-            denominator1 = (lambda_list[0] - lambda_list[1]) * (
-                lambda_list[0] - lambda_list[2]
-            )
-            denominator2 = (lambda_list[1] - lambda_list[0]) * (
-                lambda_list[1] - lambda_list[2]
-            )
-            denominator3 = (lambda_list[2] - lambda_list[0]) * (
-                lambda_list[2] - lambda_list[1]
-            )
-            return [
-                [
-                    1 / denominator1,
-                    (-lambda_list[1] - lambda_list[2]) / denominator1,
-                    lambda_list[1] * lambda_list[2] / denominator1,
-                ],
-                [
-                    1 / denominator2,
-                    (-lambda_list[0] - lambda_list[2]) / denominator2,
-                    lambda_list[0] * lambda_list[2] / denominator2,
-                ],
-                [
-                    1 / denominator3,
-                    (-lambda_list[0] - lambda_list[1]) / denominator3,
-                    lambda_list[0] * lambda_list[1] / denominator3,
-                ],
-            ]
-        elif order == 3:
-            denominator1 = (
-                (lambda_list[0] - lambda_list[1])
-                * (lambda_list[0] - lambda_list[2])
-                * (lambda_list[0] - lambda_list[3])
-            )
-            denominator2 = (
-                (lambda_list[1] - lambda_list[0])
-                * (lambda_list[1] - lambda_list[2])
-                * (lambda_list[1] - lambda_list[3])
-            )
-            denominator3 = (
-                (lambda_list[2] - lambda_list[0])
-                * (lambda_list[2] - lambda_list[1])
-                * (lambda_list[2] - lambda_list[3])
-            )
-            denominator4 = (
-                (lambda_list[3] - lambda_list[0])
-                * (lambda_list[3] - lambda_list[1])
-                * (lambda_list[3] - lambda_list[2])
-            )
-            return [
-                [
-                    1 / denominator1,
-                    (-lambda_list[1] - lambda_list[2] - lambda_list[3]) / denominator1,
-                    (
-                        lambda_list[1] * lambda_list[2]
-                        + lambda_list[1] * lambda_list[3]
-                        + lambda_list[2] * lambda_list[3]
-                    )
-                    / denominator1,
-                    (-lambda_list[1] * lambda_list[2] * lambda_list[3]) / denominator1,
-                ],
-                [
-                    1 / denominator2,
-                    (-lambda_list[0] - lambda_list[2] - lambda_list[3]) / denominator2,
-                    (
-                        lambda_list[0] * lambda_list[2]
-                        + lambda_list[0] * lambda_list[3]
-                        + lambda_list[2] * lambda_list[3]
-                    )
-                    / denominator2,
-                    (-lambda_list[0] * lambda_list[2] * lambda_list[3]) / denominator2,
-                ],
-                [
-                    1 / denominator3,
-                    (-lambda_list[0] - lambda_list[1] - lambda_list[3]) / denominator3,
-                    (
-                        lambda_list[0] * lambda_list[1]
-                        + lambda_list[0] * lambda_list[3]
-                        + lambda_list[1] * lambda_list[3]
-                    )
-                    / denominator3,
-                    (-lambda_list[0] * lambda_list[1] * lambda_list[3]) / denominator3,
-                ],
-                [
-                    1 / denominator4,
-                    (-lambda_list[0] - lambda_list[1] - lambda_list[2]) / denominator4,
-                    (
-                        lambda_list[0] * lambda_list[1]
-                        + lambda_list[0] * lambda_list[2]
-                        + lambda_list[1] * lambda_list[2]
-                    )
-                    / denominator4,
-                    (-lambda_list[0] * lambda_list[1] * lambda_list[2]) / denominator4,
-                ],
-            ]
-
-    def get_coefficients_fn(
-        self, order, interval_start, interval_end, lambda_list, tau
-    ):
-        """
-        Calculate the coefficient of gradients.
-        """
-        assert order in [1, 2, 3, 4]
-        assert order == len(lambda_list), (
-            "the length of lambda list must be equal to the order"
-        )
-        coefficients = []
-        lagrange_coefficient = self.lagrange_polynomial_coefficient(
-            order - 1, lambda_list
-        )
-        for i in range(order):
-            coefficient = 0
-            for j in range(order):
-                if self.predict_x0:
-                    coefficient += lagrange_coefficient[i][
-                        j
-                    ] * self.get_coefficients_exponential_positive(
-                        order - 1 - j, interval_start, interval_end, tau
-                    )
-                else:
-                    coefficient += lagrange_coefficient[i][
-                        j
-                    ] * self.get_coefficients_exponential_negative(
-                        order - 1 - j, interval_start, interval_end
-                    )
-            coefficients.append(coefficient)
-        assert len(coefficients) == order, (
-            "the length of coefficients does not match the order"
-        )
-        return coefficients
-
-    def adams_bashforth_update(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Predictor, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def adams_moulton_update(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        t_list = t_prev_list + [t]
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def adams_bashforth_update_few_steps(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Predictor, with the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        if self.predict_x0:
-            if (
-                order == 2
-            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to unipc. Note: This is used only for few steps sampling.
-                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
-                # ODE case
-                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
-                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
-                gradient_coefficients[0] += (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h**2 / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2)
-                    )
-                    / (
-                        ns.marginal_lambda(t_prev_list[-1])
-                        - ns.marginal_lambda(t_prev_list[-2])
-                    )
-                )
-                gradient_coefficients[1] -= (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h**2 / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2)
-                    )
-                    / (
-                        ns.marginal_lambda(t_prev_list[-1])
-                        - ns.marginal_lambda(t_prev_list[-2])
-                    )
-                )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def adams_moulton_update_few_steps(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        t_list = t_prev_list + [t]
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        if self.predict_x0:
-            if (
-                order == 2
-            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to UniPC. Note: This is used only for few steps sampling.
-                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
-                # ODE case
-                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
-                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
-                gradient_coefficients[0] += (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2 * h)
-                    )
-                )
-                gradient_coefficients[1] -= (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2 * h)
-                    )
-                )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def sample_few_steps(
-        self,
-        x,
-        tau,
-        steps=5,
-        t_start=None,
-        t_end=None,
-        skip_type="time",
-        skip_order=1,
-        predictor_order=3,
-        corrector_order=4,
-        pc_mode="PEC",
-        return_intermediate=False,
-    ):
-        """
-        For the PC-mode, please refer to the wiki page
-        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
-        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
-        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
-        """
-
-        skip_first_step = False
-        skip_final_step = True
-        lower_order_final = True
-        denoise_to_zero = False
-
-        assert pc_mode in [
-            "PEC",
-            "PECE",
-        ], "Predictor-corrector mode only supports PEC and PECE"
-        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
-        t_T = self.noise_schedule.T if t_start is None else t_start
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-
-        device = x.device
-        intermediates = []
-        with torch.no_grad():
-            assert steps >= max(predictor_order, corrector_order - 1)
-            timesteps = self.get_time_steps(
-                skip_type=skip_type,
-                t_T=t_T,
-                t_0=t_0,
-                N=steps,
-                order=skip_order,
-                device=device,
-            )
-            assert timesteps.shape[0] - 1 == steps
-            # Init the initial values.
-            step = 0
-            t = timesteps[step]
-            noise = torch.randn_like(x)
-            t_prev_list = [t]
-            # do not evaluate if skip_first_step
-            if skip_first_step:
-                if self.predict_x0:
-                    alpha_t = self.noise_schedule.marginal_alpha(t)
-                    sigma_t = self.noise_schedule.marginal_std(t)
-                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
-                else:
-                    model_prev_list = [x]
-            else:
-                model_prev_list = [self.model_fn(x, t)]
-
-            if self.correcting_xt_fn is not None:
-                x = self.correcting_xt_fn(x, t, step)
-            if return_intermediate:
-                intermediates.append(x)
-
-            # determine the first several values
-            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
-                t = timesteps[step]
-                predictor_order_used = min(predictor_order, step)
-                corrector_order_used = min(corrector_order, step + 1)
-                noise = torch.randn_like(x)
-                # predictor step
-                x_p = self.adams_bashforth_update_few_steps(
-                    order=predictor_order_used,
-                    x=x,
-                    tau=tau(t),
-                    model_prev_list=model_prev_list,
-                    t_prev_list=t_prev_list,
-                    noise=noise,
-                    t=t,
-                )
-                # evaluation step
-                model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                model_prev_list.append(model_x)
-                # corrector step
-                if corrector_order > 0:
-                    x = self.adams_moulton_update_few_steps(
-                        order=corrector_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x = x_p
-
-                # evaluation step if correction and mode = pece
-                if corrector_order > 0:
-                    if pc_mode == "PECE":
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-
-            for step in tqdm(
-                range(max(predictor_order, corrector_order - 1), steps + 1)
-            ):
-                if lower_order_final:
-                    predictor_order_used = min(predictor_order, steps - step + 1)
-                    corrector_order_used = min(corrector_order, steps - step + 2)
-
-                else:
-                    predictor_order_used = predictor_order
-                    corrector_order_used = corrector_order
-                t = timesteps[step]
-                noise = torch.randn_like(x)
-
-                # predictor step
-                if skip_final_step and step == steps and not denoise_to_zero:
-                    x_p = self.adams_bashforth_update_few_steps(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=0,
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x_p = self.adams_bashforth_update_few_steps(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-
-                # evaluation step
-                # do not evaluate if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                # do not update if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_prev_list.append(model_x)
-
-                # corrector step
-                # do not correct if skip_final_step and step = steps
-                if corrector_order > 0:
-                    if not skip_final_step or step < steps:
-                        x = self.adams_moulton_update_few_steps(
-                            order=corrector_order_used,
-                            x=x,
-                            tau=tau(t),
-                            model_prev_list=model_prev_list,
-                            t_prev_list=t_prev_list,
-                            noise=noise,
-                            t=t,
-                        )
-                    else:
-                        x = x_p
-                else:
-                    x = x_p
-
-                # evaluation step if mode = pece and step != steps
-                if corrector_order > 0:
-                    if pc_mode == "PECE" and step < steps:
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-                del model_prev_list[0]
-
-            if denoise_to_zero:
-                t = torch.ones((1,)).to(device) * t_0
-                x = self.denoise_to_zero_fn(x, t)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step + 1)
-                if return_intermediate:
-                    intermediates.append(x)
-        if return_intermediate:
-            return x, intermediates
-        else:
-            return x
-
-    def sample_more_steps(
-        self,
-        x,
-        tau,
-        steps=20,
-        t_start=None,
-        t_end=None,
-        skip_type="time",
-        skip_order=1,
-        predictor_order=3,
-        corrector_order=4,
-        pc_mode="PEC",
-        return_intermediate=False,
-    ):
-        """
-        For the PC-mode, please refer to the wiki page
-        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
-        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
-        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
-        """
-
-        skip_first_step = False
-        skip_final_step = False
-        lower_order_final = True
-        denoise_to_zero = True
-
-        assert pc_mode in [
-            "PEC",
-            "PECE",
-        ], "Predictor-corrector mode only supports PEC and PECE"
-        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
-        t_T = self.noise_schedule.T if t_start is None else t_start
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-
-        device = x.device
-        intermediates = []
-        with torch.no_grad():
-            assert steps >= max(predictor_order, corrector_order - 1)
-            timesteps = self.get_time_steps(
-                skip_type=skip_type,
-                t_T=t_T,
-                t_0=t_0,
-                N=steps,
-                order=skip_order,
-                device=device,
-            )
-            assert timesteps.shape[0] - 1 == steps
-            # Init the initial values.
-            step = 0
-            t = timesteps[step]
-            noise = torch.randn_like(x)
-            t_prev_list = [t]
-            # do not evaluate if skip_first_step
-            if skip_first_step:
-                if self.predict_x0:
-                    alpha_t = self.noise_schedule.marginal_alpha(t)
-                    sigma_t = self.noise_schedule.marginal_std(t)
-                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
-                else:
-                    model_prev_list = [x]
-            else:
-                model_prev_list = [self.model_fn(x, t)]
-
-            if self.correcting_xt_fn is not None:
-                x = self.correcting_xt_fn(x, t, step)
-            if return_intermediate:
-                intermediates.append(x)
-
-            # determine the first several values
-            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
-                t = timesteps[step]
-                predictor_order_used = min(predictor_order, step)
-                corrector_order_used = min(corrector_order, step + 1)
-                noise = torch.randn_like(x)
-                # predictor step
-                x_p = self.adams_bashforth_update(
-                    order=predictor_order_used,
-                    x=x,
-                    tau=tau(t),
-                    model_prev_list=model_prev_list,
-                    t_prev_list=t_prev_list,
-                    noise=noise,
-                    t=t,
-                )
-                # evaluation step
-                model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                model_prev_list.append(model_x)
-                # corrector step
-                if corrector_order > 0:
-                    x = self.adams_moulton_update(
-                        order=corrector_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x = x_p
-
-                # evaluation step if mode = pece
-                if corrector_order > 0:
-                    if pc_mode == "PECE":
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-
-            for step in tqdm(
-                range(max(predictor_order, corrector_order - 1), steps + 1)
-            ):
-                if lower_order_final:
-                    predictor_order_used = min(predictor_order, steps - step + 1)
-                    corrector_order_used = min(corrector_order, steps - step + 2)
-
-                else:
-                    predictor_order_used = predictor_order
-                    corrector_order_used = corrector_order
-                t = timesteps[step]
-                noise = torch.randn_like(x)
-
-                # predictor step
-                if skip_final_step and step == steps and not denoise_to_zero:
-                    x_p = self.adams_bashforth_update(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=0,
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x_p = self.adams_bashforth_update(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-
-                # evaluation step
-                # do not evaluate if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                # do not update if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_prev_list.append(model_x)
-
-                # corrector step
-                # do not correct if skip_final_step and step = steps
-                if corrector_order > 0:
-                    if not skip_final_step or step < steps:
-                        x = self.adams_moulton_update(
-                            order=corrector_order_used,
-                            x=x,
-                            tau=tau(t),
-                            model_prev_list=model_prev_list,
-                            t_prev_list=t_prev_list,
-                            noise=noise,
-                            t=t,
-                        )
-                    else:
-                        x = x_p
-                else:
-                    x = x_p
-
-                # evaluation step if mode = pece and step != steps
-                if corrector_order > 0:
-                    if pc_mode == "PECE" and step < steps:
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-                del model_prev_list[0]
-
-            if denoise_to_zero:
-                t = torch.ones((1,)).to(device) * t_0
-                x = self.denoise_to_zero_fn(x, t)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step + 1)
-                if return_intermediate:
-                    intermediates.append(x)
-        if return_intermediate:
-            return x, intermediates
-        else:
-            return x
-
-    def sample(
-        self,
-        mode,
-        x,
-        tau,
-        steps,
-        t_start=None,
-        t_end=None,
-        skip_type="time",
-        skip_order=1,
-        predictor_order=3,
-        corrector_order=4,
-        pc_mode="PEC",
-        return_intermediate=False,
-    ):
-        """
-        For the PC-mode, please refer to the wiki page
-        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
-        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
-        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
-
-        'few_steps' mode is recommended. The differences between 'few_steps' and 'more_steps' are as below:
-        1) 'few_steps' do not correct at final step and do not denoise to zero, while 'more_steps' do these two.
-        Thus the NFEs for 'few_steps' = steps, NFEs for 'more_steps' = steps + 2
-        For most of the experiments and tasks, we find these two operations do not have much help to sample quality.
-        2) 'few_steps' use a rescaling trick as in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        We find it will slightly improve the sample quality especially in few steps.
-        """
-        assert mode in [
-            "few_steps",
-            "more_steps",
-        ], "mode must be either 'few_steps' or 'more_steps'"
-        if mode == "few_steps":
-            return self.sample_few_steps(
-                x=x,
-                tau=tau,
-                steps=steps,
-                t_start=t_start,
-                t_end=t_end,
-                skip_type=skip_type,
-                skip_order=skip_order,
-                predictor_order=predictor_order,
-                corrector_order=corrector_order,
-                pc_mode=pc_mode,
-                return_intermediate=return_intermediate,
-            )
-        else:
-            return self.sample_more_steps(
-                x=x,
-                tau=tau,
-                steps=steps,
-                t_start=t_start,
-                t_end=t_end,
-                skip_type=skip_type,
-                skip_order=skip_order,
-                predictor_order=predictor_order,
-                corrector_order=corrector_order,
-                pc_mode=pc_mode,
-                return_intermediate=return_intermediate,
-            )
-
-
-#############################################################
-# other utility functions
-#############################################################
-
-
-def interpolate_fn(x, xp, yp):
-    """
-    A piecewise linear function y = f(x), using xp and yp as keypoints.
-    We implement f(x) in a differentiable way (i.e. applicable for autograd).
-    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
-    Args:
-        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
-        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
-        yp: PyTorch tensor with shape [C, K].
-    Returns:
-        The function values f(x), with shape [N, C].
-    """
-    N, K = x.shape[0], xp.shape[1]
-    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
-    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
-    x_idx = torch.argmin(x_indices, dim=2)
-    cand_start_idx = x_idx - 1
-    start_idx = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(1, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    end_idx = torch.where(
-        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
-    )
-    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
-    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
-    start_idx2 = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(0, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
-    start_y = torch.gather(
-        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
-    ).squeeze(2)
-    end_y = torch.gather(
-        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
-    ).squeeze(2)
-    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
-    return cand
-
-
-def expand_dims(v, dims):
-    """
-    Expand the tensor `v` to the dim `dims`.
-    Args:
-        `v`: a PyTorch tensor with shape [N].
-        `dim`: a `int`.
-    Returns:
-        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
-    """
-    return v[(...,) + (None,) * (dims - 1)]
diff --git a/sana/sana_600M/config.yaml b/sana/sana_600M/config.yaml
deleted file mode 100644
index fb9afda13..000000000
--- a/sana/sana_600M/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-build_commands: []
-base_image:
-  image: alphatozeta/cuda-python:12.1.1-cudnn8-devel-ubuntu22.04
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input: {
-    "prompt": "a photo of an astronaut riding a horse on mars",
-    "height": 1024,
-    "width": 1024,
-    "guidance_scale": 5.0,
-    "pag_guidance_scale": 2.0,
-    "num_inference_steps": 18,
-    "seed": 4096,
-  }
-model_name: Sana 600M
-python_version: py311
-requirements:
-- git+https://github.com/NVlabs/Sana.git@d7945026d8d85008aca1d1e6db5717a1069f5c84
-- huggingface-hub==0.26.3
-- hf-transfer==0.1.8
-resources:
-  accelerator: H100_40GB
-  use_gpu: true
-secrets:
-  hf_access_token: "null"
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
-- python3.10-venv
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/dpm_solver.py b/sana/sana_600M/packages/Sana/diffusion/model/dpm_solver.py
deleted file mode 100755
index 826c373ec..000000000
--- a/sana/sana_600M/packages/Sana/diffusion/model/dpm_solver.py
+++ /dev/null
@@ -1,1908 +0,0 @@
-# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-
-# This file is modified from https://github.com/PixArt-alpha/PixArt-sigma
-import os
-
-import torch
-from tqdm import tqdm
-
-from .nets.sana_blocks import (
-    PAGCFGIdentitySelfAttnProcessorLiteLA,
-    PAGIdentitySelfAttnProcessorLiteLA,
-    SelfAttnProcessorLiteLA,
-)
-
-
-class NoiseScheduleVP:
-    def __init__(
-        self,
-        schedule="discrete",
-        betas=None,
-        alphas_cumprod=None,
-        continuous_beta_0=0.1,
-        continuous_beta_1=20.0,
-        dtype=torch.float32,
-    ):
-        r"""Create a wrapper class for the forward SDE (VP type).
-
-        ***
-        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
-                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
-        ***
-
-        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
-        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
-        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
-
-            log_alpha_t = self.marginal_log_mean_coeff(t)
-            sigma_t = self.marginal_std(t)
-            lambda_t = self.marginal_lambda(t)
-
-        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
-
-            t = self.inverse_lambda(lambda_t)
-
-        ===============================================================
-
-        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
-
-        1. For discrete-time DPMs:
-
-            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
-                t_i = (i + 1) / N
-            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
-            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
-
-            Args:
-                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
-                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
-
-            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
-
-            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
-                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
-                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
-                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
-                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
-                and
-                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
-
-
-        2. For continuous-time DPMs:
-
-            We support the linear VPSDE for the continuous time setting. The hyperparameters for the noise
-            schedule are the default settings in Yang Song's ScoreSDE:
-
-            Args:
-                beta_min: A `float` number. The smallest beta for the linear schedule.
-                beta_max: A `float` number. The largest beta for the linear schedule.
-                T: A `float` number. The ending time of the forward process.
-
-        ===============================================================
-
-        Args:
-            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
-                    'linear' for continuous-time DPMs.
-        Returns:
-            A wrapper object of the forward SDE (VP type).
-
-        ===============================================================
-
-        Example:
-
-        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', betas=betas)
-
-        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
-
-        # For continuous-time DPMs (VPSDE), linear schedule:
-        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
-
-        """
-
-        if schedule not in ["discrete", "linear"]:
-            raise ValueError(
-                f"Unsupported noise schedule {schedule}. The schedule needs to be 'discrete' or 'linear'"
-            )
-
-        self.schedule = schedule
-        if schedule == "discrete":
-            if betas is not None:
-                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
-            else:
-                assert alphas_cumprod is not None
-                log_alphas = 0.5 * torch.log(alphas_cumprod)
-            self.T = 1.0
-            self.log_alpha_array = (
-                self.numerical_clip_alpha(log_alphas)
-                .reshape(
-                    (
-                        1,
-                        -1,
-                    )
-                )
-                .to(dtype=dtype)
-            )
-            self.total_N = self.log_alpha_array.shape[1]
-            self.t_array = (
-                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
-                .reshape((1, -1))
-                .to(dtype=dtype)
-            )
-        else:
-            self.T = 1.0
-            self.total_N = 1000
-            self.beta_0 = continuous_beta_0
-            self.beta_1 = continuous_beta_1
-
-    def numerical_clip_alpha(self, log_alphas, clipped_lambda=-5.1):
-        """
-        For some beta schedules such as cosine schedule, the log-SNR has numerical isssues.
-        We clip the log-SNR near t=T within -5.1 to ensure the stability.
-        Such a trick is very useful for diffusion models with the cosine schedule, such as i-DDPM, guided-diffusion and GLIDE.
-        """
-        log_sigmas = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_alphas))
-        lambs = log_alphas - log_sigmas
-        idx = torch.searchsorted(torch.flip(lambs, [0]), clipped_lambda)
-        if idx > 0:
-            log_alphas = log_alphas[:-idx]
-        return log_alphas
-
-    def marginal_log_mean_coeff(self, t):
-        """
-        Compute log(alpha_t) of a given continuous-time label t in [0, T].
-        """
-        if self.schedule == "discrete":
-            return interpolate_fn(
-                t.reshape((-1, 1)),
-                self.t_array.to(t.device),
-                self.log_alpha_array.to(t.device),
-            ).reshape(-1)
-        elif self.schedule == "linear":
-            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
-
-    def marginal_alpha(self, t):
-        """
-        Compute alpha_t of a given continuous-time label t in [0, T].
-        """
-        return torch.exp(self.marginal_log_mean_coeff(t))
-
-    def marginal_std(self, t):
-        """
-        Compute sigma_t of a given continuous-time label t in [0, T].
-        """
-        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
-
-    def marginal_lambda(self, t):
-        """
-        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
-        """
-        log_mean_coeff = self.marginal_log_mean_coeff(t)
-        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
-        return log_mean_coeff - log_std
-
-    def inverse_lambda(self, lamb):
-        """
-        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
-        """
-        if self.schedule == "linear":
-            tmp = (
-                2.0
-                * (self.beta_1 - self.beta_0)
-                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
-            )
-            Delta = self.beta_0**2 + tmp
-            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
-        elif self.schedule == "discrete":
-            log_alpha = -0.5 * torch.logaddexp(
-                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
-            )
-            t = interpolate_fn(
-                log_alpha.reshape((-1, 1)),
-                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
-                torch.flip(self.t_array.to(lamb.device), [1]),
-            )
-            return t.reshape((-1,))
-
-
-class NoiseScheduleFlow:
-    def __init__(
-        self,
-        schedule="discrete_flow",
-    ):
-        """Create a wrapper class for the forward SDE (EDM type)."""
-        self.T = 1
-        self.t0 = 0.001
-        self.schedule = schedule  # ['continuous', 'discrete_flow']
-        self.total_N = 1000
-
-    def marginal_log_mean_coeff(self, t):
-        """
-        Compute log(alpha_t) of a given continuous-time label t in [0, T].
-        """
-        return torch.log(self.marginal_alpha(t))
-
-    def marginal_alpha(self, t):
-        """
-        Compute alpha_t of a given continuous-time label t in [0, T].
-        """
-        return 1 - t
-
-    @staticmethod
-    def marginal_std(t):
-        """
-        Compute sigma_t of a given continuous-time label t in [0, T].
-        """
-        return t
-
-    def marginal_lambda(self, t):
-        """
-        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
-        """
-        log_mean_coeff = self.marginal_log_mean_coeff(t)
-        log_std = torch.log(self.marginal_std(t))
-        return log_mean_coeff - log_std
-
-    @staticmethod
-    def inverse_lambda(lamb):
-        """
-        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
-        """
-        return torch.exp(-lamb)
-
-    def edm_sigma(self, t):
-        return self.marginal_std(t) / self.marginal_alpha(t)
-
-    def edm_inverse_sigma(self, edmsigma):
-        sigma = edmsigma
-        lambda_t = torch.log(1 / sigma)
-        t = self.inverse_lambda(lambda_t)
-        return t
-
-
-def model_wrapper(
-    model,
-    noise_schedule,
-    model_type="noise",
-    model_kwargs={},
-    guidance_type="uncond",
-    condition=None,
-    unconditional_condition=None,
-    guidance_scale=1.0,
-    pag_scale=1.0,
-    pag_applied_layers=[],
-    interval_guidance=[0, 1.0],
-    classifier_fn=None,
-    classifier_kwargs={},
-):
-    """Create a wrapper function for the noise prediction model.
-
-    DPM-Solver needs to solve the continuous-time diffusion ODEs. For DPMs trained on discrete-time labels, we need to
-    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
-
-    We support four types of the diffusion model by setting `model_type`:
-
-        1. "noise": noise prediction model. (Trained by predicting noise).
-
-        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
-
-        3. "v": velocity prediction model. (Trained by predicting the velocity).
-            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
-
-            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
-                arXiv preprint arXiv:2202.00512 (2022).
-            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
-                arXiv preprint arXiv:2210.02303 (2022).
-
-        4. "score": marginal score function. (Trained by denoising score matching).
-            Note that the score function and the noise prediction model follows a simple relationship:
-            ```
-                noise(x_t, t) = -sigma_t * score(x_t, t)
-            ```
-
-    We support three types of guided sampling by DPMs by setting `guidance_type`:
-        1. "uncond": unconditional sampling by DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-
-        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-
-            The input `classifier_fn` has the following format:
-            ``
-                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
-            ``
-
-            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
-                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
-
-        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
-            ``
-            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
-
-            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
-                arXiv preprint arXiv:2207.12598 (2022).
-
-
-    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
-    or continuous-time labels (i.e. epsilon to T).
-
-    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
-    ``
-        def model_fn(x, t_continuous) -> noise:
-            t_input = get_model_input_time(t_continuous)
-            return noise_pred(model, x, t_input, **model_kwargs)
-    ``
-    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for DPM-Solver.
-
-    ===============================================================
-
-    Args:
-        model: A diffusion model with the corresponding format described above.
-        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
-        model_type: A `str`. The parameterization type of the diffusion model.
-                    "noise" or "x_start" or "v" or "score".
-        model_kwargs: A `dict`. A dict for the other inputs of the model function.
-        guidance_type: A `str`. The type of the guidance for sampling.
-                    "uncond" or "classifier" or "classifier-free".
-        condition: A pytorch tensor. The condition for the guided sampling.
-                    Only used for "classifier" or "classifier-free" guidance type.
-        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
-                    Only used for "classifier-free" guidance type.
-        guidance_scale: A `float`. The scale for the guided sampling.
-        classifier_fn: A classifier function. Only used for the classifier guidance.
-        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
-    Returns:
-        A noise prediction model that accepts the noised data and the continuous time as the inputs.
-    """
-
-    def get_model_input_time(t_continuous):
-        """
-        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
-        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
-        For continuous-time DPMs, we just use `t_continuous`.
-        """
-        if noise_schedule.schedule == "discrete":
-            return (
-                t_continuous - 1.0 / noise_schedule.total_N
-            ) * noise_schedule.total_N
-        elif noise_schedule.schedule == "discrete_flow":
-            return t_continuous * noise_schedule.total_N
-        else:
-            return t_continuous
-
-    def noise_pred_fn(x, t_continuous, cond=None):
-        t_input = get_model_input_time(t_continuous)
-        if cond is None:
-            output = model(x, t_input, **model_kwargs)
-        else:
-            output = model(x, t_input, cond, **model_kwargs)
-        if model_type == "noise":
-            return output
-        elif model_type == "x_start":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return (x - expand_dims(alpha_t, x.dim()) * output) / expand_dims(
-                sigma_t, x.dim()
-            )
-        elif model_type == "v":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return (
-                expand_dims(alpha_t, x.dim()) * output
-                + expand_dims(sigma_t, x.dim()) * x
-            )
-        elif model_type == "score":
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            return -expand_dims(sigma_t, x.dim()) * output
-        elif model_type == "flow":
-            _, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            try:
-                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output + x
-            except:
-                noise = (1 - expand_dims(sigma_t, x.dim()).to(x)) * output[0] + x
-            return noise
-
-    def cond_grad_fn(x, t_input):
-        """
-        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
-        """
-        with torch.enable_grad():
-            x_in = x.detach().requires_grad_(True)
-            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
-            return torch.autograd.grad(log_prob.sum(), x_in)[0]
-
-    def model_fn(x, t_continuous):
-        """
-        The noise predicition model function that is used for DPM-Solver.
-        """
-        guidance_tp = guidance_type
-        if guidance_tp == "uncond":
-            return noise_pred_fn(x, t_continuous)
-        elif guidance_tp == "classifier":
-            assert classifier_fn is not None
-            t_input = get_model_input_time(t_continuous)
-            cond_grad = cond_grad_fn(x, t_input)
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            noise = noise_pred_fn(x, t_continuous)
-            return noise - guidance_scale * expand_dims(sigma_t, x.dim()) * cond_grad
-        elif guidance_tp == "classifier-free":
-            if (
-                guidance_scale == 1.0
-                or unconditional_condition is None
-                or not (interval_guidance[0] < t_continuous[0] < interval_guidance[1])
-            ):
-                return noise_pred_fn(x, t_continuous, cond=condition)
-            else:
-                x_in = torch.cat([x] * 2)
-                t_in = torch.cat([t_continuous] * 2)
-                c_in = torch.cat([unconditional_condition, condition])
-                try:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
-                except:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
-                        2
-                    )
-                return noise_uncond + guidance_scale * (noise - noise_uncond)
-        elif guidance_tp == "classifier-free_PAG":
-            for i in pag_applied_layers:
-                if isinstance(model, torch.nn.Module):
-                    model.blocks[i].attn.forward = (
-                        PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
-                        if guidance_scale == 1.0
-                        else PAGCFGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
-                    )
-                else:
-                    model.__self__.blocks[i].attn.forward = (
-                        PAGIdentitySelfAttnProcessorLiteLA(
-                            model.__self__.blocks[i].attn
-                        )
-                        if guidance_scale == 1.0
-                        else PAGCFGIdentitySelfAttnProcessorLiteLA(
-                            model.__self__.blocks[i].attn
-                        )
-                    )
-            num_inputs = 2 if guidance_scale == 1.0 else 3
-            x_in = torch.cat([x] * num_inputs)
-            t_in = torch.cat([t_continuous] * num_inputs)
-            c_in = torch.cat(
-                [condition, condition]
-                if guidance_scale == 1.0
-                else [unconditional_condition, condition, condition]
-            )
-
-            try:
-                chunks = noise_pred_fn(x_in, t_in, cond=c_in).chunk(num_inputs)
-            except:
-                chunks = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(num_inputs)
-
-            if guidance_scale == 1.0:
-                noise, noise_perturb = chunks
-                noise_pred = noise + pag_scale * (noise - noise_perturb)
-            else:
-                noise_uncond, noise, noise_perturb = chunks
-                noise_pred = (
-                    noise_uncond
-                    + guidance_scale * (noise - noise_uncond)
-                    + pag_scale * (noise - noise_perturb)
-                )
-            for i in pag_applied_layers:
-                if isinstance(model, torch.nn.Module):
-                    model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                        model.blocks[i].attn
-                    )
-                else:
-                    model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                        model.__self__.blocks[i].attn
-                    )
-
-            return noise_pred
-        elif guidance_tp == "classifier-free_PAG_seq":
-            num_inputs = 2
-            if t_continuous[0] < 0.5:
-                # cfg
-                if (
-                    guidance_scale == 1.0
-                    or unconditional_condition is None
-                    or not (
-                        interval_guidance[0] < t_continuous[0] < interval_guidance[1]
-                    )
-                ):
-                    return noise_pred_fn(x, t_continuous, cond=condition)
-
-                x_in = torch.cat([x] * num_inputs)
-                t_in = torch.cat([t_continuous] * num_inputs)
-                c_in = torch.cat([unconditional_condition, condition])
-
-                try:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
-                except:
-                    noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in)[0].chunk(
-                        num_inputs
-                    )
-                return noise_uncond + guidance_scale * (noise - noise_uncond)
-            else:
-                # pag
-                for i in pag_applied_layers:
-                    if isinstance(model, torch.nn.Module):
-                        model.blocks[i].attn.forward = (
-                            PAGIdentitySelfAttnProcessorLiteLA(model.blocks[i].attn)
-                            if guidance_scale == 1.0
-                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
-                                model.blocks[i].attn
-                            )
-                        )
-                    else:
-                        model.__self__.blocks[i].attn.forward = (
-                            PAGIdentitySelfAttnProcessorLiteLA(
-                                model.__self__.blocks[i].attn
-                            )
-                            if guidance_scale == 1.0
-                            else PAGCFGIdentitySelfAttnProcessorLiteLA(
-                                model.__self__.blocks[i].attn
-                            )
-                        )
-                x_in = torch.cat([x] * 3)
-                t_in = torch.cat([t_continuous] * 3)
-                c_in = torch.cat([unconditional_condition, condition, condition])
-
-                try:
-                    noise_uncond, noise, noise_perturb = noise_pred_fn(
-                        x_in, t_in, cond=c_in
-                    ).chunk(3)
-                except:
-                    noise_uncond, noise, noise_perturb = noise_pred_fn(
-                        x_in, t_in, cond=c_in
-                    )[0].chunk(3)
-
-                for i in pag_applied_layers:
-                    if isinstance(model, torch.nn.Module):
-                        model.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                            model.blocks[i].attn
-                        )
-                    else:
-                        model.__self__.blocks[i].attn.forward = SelfAttnProcessorLiteLA(
-                            model.__self__.blocks[i].attn
-                        )
-
-                return (
-                    noise_uncond
-                    + guidance_scale * (noise - noise_uncond)
-                    + pag_scale * (noise - noise_perturb)
-                )
-
-    assert model_type in ["noise", "x_start", "v", "score", "flow"]
-    assert guidance_type in [
-        "uncond",
-        "classifier",
-        "classifier-free",
-        "classifier-free_PAG",
-        "classifier-free_PAG_seq",
-    ]
-    return model_fn
-
-
-class DPM_Solver:
-    def __init__(
-        self,
-        model_fn,
-        noise_schedule,
-        algorithm_type="dpmsolver++",
-        correcting_x0_fn=None,
-        correcting_xt_fn=None,
-        thresholding_max_val=1.0,
-        dynamic_thresholding_ratio=0.995,
-    ):
-        """Construct a DPM-Solver.
-
-        We support both DPM-Solver (`algorithm_type="dpmsolver"`) and DPM-Solver++ (`algorithm_type="dpmsolver++"`).
-
-        We also support the "dynamic thresholding" method in Imagen[1]. For pixel-space diffusion models, you
-        can set both `algorithm_type="dpmsolver++"` and `correcting_x0_fn="dynamic_thresholding"` to use the
-        dynamic thresholding. The "dynamic thresholding" can greatly improve the sample quality for pixel-space
-        DPMs with large guidance scales. Note that the thresholding method is **unsuitable** for latent-space
-        DPMs (such as stable-diffusion).
-
-        To support advanced algorithms in image-to-image applications, we also support corrector functions for
-        both x0 and xt.
-
-        Args:
-            model_fn: A noise prediction model function which accepts the continuous-time input (t in [epsilon, T]):
-                ``
-                def model_fn(x, t_continuous):
-                    return noise
-                ``
-                The shape of `x` is `(batch_size, **shape)`, and the shape of `t_continuous` is `(batch_size,)`.
-            noise_schedule: A noise schedule object, such as NoiseScheduleVP.
-            algorithm_type: A `str`. Either "dpmsolver" or "dpmsolver++".
-            correcting_x0_fn: A `str` or a function with the following format:
-                ```
-                def correcting_x0_fn(x0, t):
-                    x0_new = ...
-                    return x0_new
-                ```
-                This function is to correct the outputs of the data prediction model at each sampling step. e.g.,
-                ```
-                x0_pred = data_pred_model(xt, t)
-                if correcting_x0_fn is not None:
-                    x0_pred = correcting_x0_fn(x0_pred, t)
-                xt_1 = update(x0_pred, xt, t)
-                ```
-                If `correcting_x0_fn="dynamic_thresholding"`, we use the dynamic thresholding proposed in Imagen[1].
-            correcting_xt_fn: A function with the following format:
-                ```
-                def correcting_xt_fn(xt, t, step):
-                    x_new = ...
-                    return x_new
-                ```
-                This function is to correct the intermediate samples xt at each sampling step. e.g.,
-                ```
-                xt = ...
-                xt = correcting_xt_fn(xt, t, step)
-                ```
-            thresholding_max_val: A `float`. The max value for thresholding.
-                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
-            dynamic_thresholding_ratio: A `float`. The ratio for dynamic thresholding (see Imagen[1] for details).
-                Valid only when use `dpmsolver++` and `correcting_x0_fn="dynamic_thresholding"`.
-
-        [1] Chitwan Saharia, William Chan, Saurabh Saxena, Lala Li, Jay Whang, Emily Denton, Seyed Kamyar Seyed Ghasemipour,
-            Burcu Karagol Ayan, S Sara Mahdavi, Rapha Gontijo Lopes, et al. Photorealistic text-to-image diffusion models
-            with deep language understanding. arXiv preprint arXiv:2205.11487, 2022b.
-        """
-        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
-        self.noise_schedule = noise_schedule
-        assert algorithm_type in ["dpmsolver", "dpmsolver++"]
-        self.algorithm_type = algorithm_type
-        if correcting_x0_fn == "dynamic_thresholding":
-            self.correcting_x0_fn = self.dynamic_thresholding_fn
-        else:
-            self.correcting_x0_fn = correcting_x0_fn
-        self.correcting_xt_fn = correcting_xt_fn
-        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
-        self.thresholding_max_val = thresholding_max_val
-        self.register_progress_bar()
-
-    def register_progress_bar(self, progress_fn=None):
-        """
-        Register a progress bar callback function
-
-        Args:
-            progress_fn: Callback function that takes current step and total steps as parameters
-        """
-        self.progress_fn = (
-            progress_fn if progress_fn is not None else lambda step, total: None
-        )
-
-    def update_progress(self, step, total_steps):
-        """
-        Update sampling progress
-
-        Args:
-            step: Current step number
-            total_steps: Total number of steps
-        """
-        if hasattr(self, "progress_fn"):
-            try:
-                self.progress_fn(
-                    step / total_steps, desc=f"Generating {step}/{total_steps}"
-                )
-            except:
-                self.progress_fn(step, total_steps)
-
-        else:
-            # If no progress_fn registered, use default empty function
-            pass
-
-    def dynamic_thresholding_fn(self, x0, t):
-        """
-        The dynamic thresholding method.
-        """
-        dims = x0.dim()
-        p = self.dynamic_thresholding_ratio
-        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
-        s = expand_dims(
-            torch.maximum(
-                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
-            ),
-            dims,
-        )
-        x0 = torch.clamp(x0, -s, s) / s
-        return x0
-
-    def noise_prediction_fn(self, x, t):
-        """
-        Return the noise prediction model.
-        """
-        return self.model(x, t)
-
-    def data_prediction_fn(self, x, t):
-        """
-        Return the data prediction model (with corrector).
-        """
-        noise = self.noise_prediction_fn(x, t)
-        alpha_t, sigma_t = (
-            self.noise_schedule.marginal_alpha(t),
-            self.noise_schedule.marginal_std(t),
-        )
-        x0 = (x - sigma_t * noise) / alpha_t
-        if self.correcting_x0_fn is not None:
-            x0 = self.correcting_x0_fn(x0, t)
-        return x0
-
-    def model_fn(self, x, t):
-        """
-        Convert the model to the noise prediction model or the data prediction model.
-        """
-        if self.algorithm_type == "dpmsolver++":
-            return self.data_prediction_fn(x, t)
-        else:
-            return self.noise_prediction_fn(x, t)
-
-    def get_time_steps(self, skip_type, t_T, t_0, N, device, shift=1.0):
-        """Compute the intermediate time steps for sampling.
-
-        Args:
-            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
-                - 'logSNR': uniform logSNR for the time steps.
-                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
-                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
-            t_T: A `float`. The starting time of the sampling (default is T).
-            t_0: A `float`. The ending time of the sampling (default is epsilon).
-            N: A `int`. The total number of the spacing of the time steps.
-            device: A torch device.
-        Returns:
-            A pytorch tensor of the time steps, with the shape (N + 1,).
-        """
-        if skip_type == "logSNR":
-            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
-            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
-            logSNR_steps = torch.linspace(
-                lambda_T.cpu().item(), lambda_0.cpu().item(), N + 1
-            ).to(device)
-            return self.noise_schedule.inverse_lambda(logSNR_steps)
-        elif skip_type == "time_uniform":
-            return torch.linspace(t_T, t_0, N + 1).to(device)
-        elif skip_type == "time_quadratic":
-            t_order = 2
-            t = (
-                torch.linspace(t_T ** (1.0 / t_order), t_0 ** (1.0 / t_order), N + 1)
-                .pow(t_order)
-                .to(device)
-            )
-            return t
-        elif skip_type == "time_uniform_flow":
-            betas = torch.linspace(t_T, t_0, N + 1).to(device)
-            sigmas = 1.0 - betas
-            sigmas = (shift * sigmas / (1 + (shift - 1) * sigmas)).flip(dims=[0])
-            return sigmas
-        else:
-            raise ValueError(
-                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time_uniform' or 'time_quadratic'"
-            )
-
-    def get_orders_and_timesteps_for_singlestep_solver(
-        self, steps, order, skip_type, t_T, t_0, device
-    ):
-        """
-        Get the order of each step for sampling by the singlestep DPM-Solver.
-
-        We combine both DPM-Solver-1,2,3 to use all the function evaluations, which is named as "DPM-Solver-fast".
-        Given a fixed number of function evaluations by `steps`, the sampling procedure by DPM-Solver-fast is:
-            - If order == 1:
-                We take `steps` of DPM-Solver-1 (i.e. DDIM).
-            - If order == 2:
-                - Denote K = (steps // 2). We take K or (K + 1) intermediate time steps for sampling.
-                - If steps % 2 == 0, we use K steps of DPM-Solver-2.
-                - If steps % 2 == 1, we use K steps of DPM-Solver-2 and 1 step of DPM-Solver-1.
-            - If order == 3:
-                - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
-                - If steps % 3 == 0, we use (K - 2) steps of DPM-Solver-3, and 1 step of DPM-Solver-2 and 1 step of DPM-Solver-1.
-                - If steps % 3 == 1, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-1.
-                - If steps % 3 == 2, we use (K - 1) steps of DPM-Solver-3 and 1 step of DPM-Solver-2.
-
-        ============================================
-        Args:
-            order: A `int`. The max order for the solver (2 or 3).
-            steps: A `int`. The total number of function evaluations (NFE).
-            skip_type: A `str`. The type for the spacing of the time steps. We support three types:
-                - 'logSNR': uniform logSNR for the time steps.
-                - 'time_uniform': uniform time for the time steps. (**Recommended for high-resolutional data**.)
-                - 'time_quadratic': quadratic time for the time steps. (Used in DDIM for low-resolutional data.)
-            t_T: A `float`. The starting time of the sampling (default is T).
-            t_0: A `float`. The ending time of the sampling (default is epsilon).
-            device: A torch device.
-        Returns:
-            orders: A list of the solver order of each step.
-        """
-        if order == 3:
-            K = steps // 3 + 1
-            if steps % 3 == 0:
-                orders = [
-                    3,
-                ] * (K - 2) + [2, 1]
-            elif steps % 3 == 1:
-                orders = [
-                    3,
-                ] * (K - 1) + [1]
-            else:
-                orders = [
-                    3,
-                ] * (K - 1) + [2]
-        elif order == 2:
-            if steps % 2 == 0:
-                K = steps // 2
-                orders = [
-                    2,
-                ] * K
-            else:
-                K = steps // 2 + 1
-                orders = [
-                    2,
-                ] * (K - 1) + [1]
-        elif order == 1:
-            K = 1
-            orders = [
-                1,
-            ] * steps
-        else:
-            raise ValueError("'order' must be '1' or '2' or '3'.")
-        if skip_type == "logSNR":
-            # To reproduce the results in DPM-Solver paper
-            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, K, device)
-        else:
-            timesteps_outer = self.get_time_steps(skip_type, t_T, t_0, steps, device)[
-                torch.cumsum(
-                    torch.tensor(
-                        [
-                            0,
-                        ]
-                        + orders
-                    ),
-                    0,
-                ).to(device)
-            ]
-        return timesteps_outer, orders
-
-    def denoise_to_zero_fn(self, x, s):
-        """
-        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
-        """
-        return self.data_prediction_fn(x, s)
-
-    def dpm_solver_first_update(self, x, s, t, model_s=None, return_intermediate=False):
-        """
-        DPM-Solver-1 (equivalent to DDIM) from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            model_s: A pytorch tensor. The model function evaluated at time `s`.
-                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
-            return_intermediate: A `bool`. If true, also return the model value at time `s`.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        ns = self.noise_schedule
-        dims = x.dim()
-        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
-        h = lambda_t - lambda_s
-        log_alpha_s, log_alpha_t = (
-            ns.marginal_log_mean_coeff(s),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_s, sigma_t = ns.marginal_std(s), ns.marginal_std(t)
-        alpha_t = torch.exp(log_alpha_t)
-
-        if self.algorithm_type == "dpmsolver++":
-            phi_1 = torch.expm1(-h)
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_t = sigma_t / sigma_s * x - alpha_t * phi_1 * model_s
-            if return_intermediate:
-                return x_t, {"model_s": model_s}
-            else:
-                return x_t
-        else:
-            phi_1 = torch.expm1(h)
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_t = torch.exp(log_alpha_t - log_alpha_s) * x - (sigma_t * phi_1) * model_s
-            if return_intermediate:
-                return x_t, {"model_s": model_s}
-            else:
-                return x_t
-
-    def singlestep_dpm_solver_second_update(
-        self,
-        x,
-        s,
-        t,
-        r1=0.5,
-        model_s=None,
-        return_intermediate=False,
-        solver_type="dpmsolver",
-    ):
-        """
-        Singlestep solver DPM-Solver-2 from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            r1: A `float`. The hyperparameter of the second-order solver.
-            model_s: A pytorch tensor. The model function evaluated at time `s`.
-                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
-            return_intermediate: A `bool`. If true, also return the model value at time `s` and `s1` (the intermediate time).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if solver_type not in ["dpmsolver", "taylor"]:
-            raise ValueError(
-                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
-            )
-        if r1 is None:
-            r1 = 0.5
-        ns = self.noise_schedule
-        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
-        h = lambda_t - lambda_s
-        lambda_s1 = lambda_s + r1 * h
-        s1 = ns.inverse_lambda(lambda_s1)
-        log_alpha_s, log_alpha_s1, log_alpha_t = (
-            ns.marginal_log_mean_coeff(s),
-            ns.marginal_log_mean_coeff(s1),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_s, sigma_s1, sigma_t = (
-            ns.marginal_std(s),
-            ns.marginal_std(s1),
-            ns.marginal_std(t),
-        )
-        alpha_s1, alpha_t = torch.exp(log_alpha_s1), torch.exp(log_alpha_t)
-
-        if self.algorithm_type == "dpmsolver++":
-            phi_11 = torch.expm1(-r1 * h)
-            phi_1 = torch.expm1(-h)
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
-            model_s1 = self.model_fn(x_s1, s1)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    - (0.5 / r1) * (alpha_t * phi_1) * (model_s1 - model_s)
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    + (1.0 / r1) * (alpha_t * (phi_1 / h + 1.0)) * (model_s1 - model_s)
-                )
-        else:
-            phi_11 = torch.expm1(r1 * h)
-            phi_1 = torch.expm1(h)
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            x_s1 = (
-                torch.exp(log_alpha_s1 - log_alpha_s) * x
-                - (sigma_s1 * phi_11) * model_s
-            )
-            model_s1 = self.model_fn(x_s1, s1)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    torch.exp(log_alpha_t - log_alpha_s) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (0.5 / r1) * (sigma_t * phi_1) * (model_s1 - model_s)
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    torch.exp(log_alpha_t - log_alpha_s) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (1.0 / r1) * (sigma_t * (phi_1 / h - 1.0)) * (model_s1 - model_s)
-                )
-        if return_intermediate:
-            return x_t, {"model_s": model_s, "model_s1": model_s1}
-        else:
-            return x_t
-
-    def singlestep_dpm_solver_third_update(
-        self,
-        x,
-        s,
-        t,
-        r1=1.0 / 3.0,
-        r2=2.0 / 3.0,
-        model_s=None,
-        model_s1=None,
-        return_intermediate=False,
-        solver_type="dpmsolver",
-    ):
-        """
-        Singlestep solver DPM-Solver-3 from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            r1: A `float`. The hyperparameter of the third-order solver.
-            r2: A `float`. The hyperparameter of the third-order solver.
-            model_s: A pytorch tensor. The model function evaluated at time `s`.
-                If `model_s` is None, we evaluate the model by `x` and `s`; otherwise we directly use it.
-            model_s1: A pytorch tensor. The model function evaluated at time `s1` (the intermediate time given by `r1`).
-                If `model_s1` is None, we evaluate the model at `s1`; otherwise we directly use it.
-            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if solver_type not in ["dpmsolver", "taylor"]:
-            raise ValueError(
-                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
-            )
-        if r1 is None:
-            r1 = 1.0 / 3.0
-        if r2 is None:
-            r2 = 2.0 / 3.0
-        ns = self.noise_schedule
-        lambda_s, lambda_t = ns.marginal_lambda(s), ns.marginal_lambda(t)
-        h = lambda_t - lambda_s
-        lambda_s1 = lambda_s + r1 * h
-        lambda_s2 = lambda_s + r2 * h
-        s1 = ns.inverse_lambda(lambda_s1)
-        s2 = ns.inverse_lambda(lambda_s2)
-        log_alpha_s, log_alpha_s1, log_alpha_s2, log_alpha_t = (
-            ns.marginal_log_mean_coeff(s),
-            ns.marginal_log_mean_coeff(s1),
-            ns.marginal_log_mean_coeff(s2),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_s, sigma_s1, sigma_s2, sigma_t = (
-            ns.marginal_std(s),
-            ns.marginal_std(s1),
-            ns.marginal_std(s2),
-            ns.marginal_std(t),
-        )
-        alpha_s1, alpha_s2, alpha_t = (
-            torch.exp(log_alpha_s1),
-            torch.exp(log_alpha_s2),
-            torch.exp(log_alpha_t),
-        )
-
-        if self.algorithm_type == "dpmsolver++":
-            phi_11 = torch.expm1(-r1 * h)
-            phi_12 = torch.expm1(-r2 * h)
-            phi_1 = torch.expm1(-h)
-            phi_22 = torch.expm1(-r2 * h) / (r2 * h) + 1.0
-            phi_2 = phi_1 / h + 1.0
-            phi_3 = phi_2 / h - 0.5
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            if model_s1 is None:
-                x_s1 = (sigma_s1 / sigma_s) * x - (alpha_s1 * phi_11) * model_s
-                model_s1 = self.model_fn(x_s1, s1)
-            x_s2 = (
-                (sigma_s2 / sigma_s) * x
-                - (alpha_s2 * phi_12) * model_s
-                + r2 / r1 * (alpha_s2 * phi_22) * (model_s1 - model_s)
-            )
-            model_s2 = self.model_fn(x_s2, s2)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    + (1.0 / r2) * (alpha_t * phi_2) * (model_s2 - model_s)
-                )
-            elif solver_type == "taylor":
-                D1_0 = (1.0 / r1) * (model_s1 - model_s)
-                D1_1 = (1.0 / r2) * (model_s2 - model_s)
-                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
-                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
-                x_t = (
-                    (sigma_t / sigma_s) * x
-                    - (alpha_t * phi_1) * model_s
-                    + (alpha_t * phi_2) * D1
-                    - (alpha_t * phi_3) * D2
-                )
-        else:
-            phi_11 = torch.expm1(r1 * h)
-            phi_12 = torch.expm1(r2 * h)
-            phi_1 = torch.expm1(h)
-            phi_22 = torch.expm1(r2 * h) / (r2 * h) - 1.0
-            phi_2 = phi_1 / h - 1.0
-            phi_3 = phi_2 / h - 0.5
-
-            if model_s is None:
-                model_s = self.model_fn(x, s)
-            if model_s1 is None:
-                x_s1 = (torch.exp(log_alpha_s1 - log_alpha_s)) * x - (
-                    sigma_s1 * phi_11
-                ) * model_s
-                model_s1 = self.model_fn(x_s1, s1)
-            x_s2 = (
-                (torch.exp(log_alpha_s2 - log_alpha_s)) * x
-                - (sigma_s2 * phi_12) * model_s
-                - r2 / r1 * (sigma_s2 * phi_22) * (model_s1 - model_s)
-            )
-            model_s2 = self.model_fn(x_s2, s2)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_s)) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (1.0 / r2) * (sigma_t * phi_2) * (model_s2 - model_s)
-                )
-            elif solver_type == "taylor":
-                D1_0 = (1.0 / r1) * (model_s1 - model_s)
-                D1_1 = (1.0 / r2) * (model_s2 - model_s)
-                D1 = (r2 * D1_0 - r1 * D1_1) / (r2 - r1)
-                D2 = 2.0 * (D1_1 - D1_0) / (r2 - r1)
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_s)) * x
-                    - (sigma_t * phi_1) * model_s
-                    - (sigma_t * phi_2) * D1
-                    - (sigma_t * phi_3) * D2
-                )
-
-        if return_intermediate:
-            return x_t, {"model_s": model_s, "model_s1": model_s1, "model_s2": model_s2}
-        else:
-            return x_t
-
-    def multistep_dpm_solver_second_update(
-        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
-    ):
-        """
-        Multistep solver DPM-Solver-2 from time `t_prev_list[-1]` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            model_prev_list: A list of pytorch tensor. The previous computed model values.
-            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if solver_type not in ["dpmsolver", "taylor"]:
-            raise ValueError(
-                f"'solver_type' must be either 'dpmsolver' or 'taylor', got {solver_type}"
-            )
-        ns = self.noise_schedule
-        model_prev_1, model_prev_0 = model_prev_list[-2], model_prev_list[-1]
-        t_prev_1, t_prev_0 = t_prev_list[-2], t_prev_list[-1]
-        lambda_prev_1, lambda_prev_0, lambda_t = (
-            ns.marginal_lambda(t_prev_1),
-            ns.marginal_lambda(t_prev_0),
-            ns.marginal_lambda(t),
-        )
-        log_alpha_prev_0, log_alpha_t = (
-            ns.marginal_log_mean_coeff(t_prev_0),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
-        alpha_t = torch.exp(log_alpha_t)
-
-        h_0 = lambda_prev_0 - lambda_prev_1
-        h = lambda_t - lambda_prev_0
-        r0 = h_0 / h
-        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
-        if self.algorithm_type == "dpmsolver++":
-            phi_1 = torch.expm1(-h)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (sigma_t / sigma_prev_0) * x
-                    - (alpha_t * phi_1) * model_prev_0
-                    - 0.5 * (alpha_t * phi_1) * D1_0
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    (sigma_t / sigma_prev_0) * x
-                    - (alpha_t * phi_1) * model_prev_0
-                    + (alpha_t * (phi_1 / h + 1.0)) * D1_0
-                )
-        else:
-            phi_1 = torch.expm1(h)
-            if solver_type == "dpmsolver":
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
-                    - (sigma_t * phi_1) * model_prev_0
-                    - 0.5 * (sigma_t * phi_1) * D1_0
-                )
-            elif solver_type == "taylor":
-                x_t = (
-                    (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
-                    - (sigma_t * phi_1) * model_prev_0
-                    - (sigma_t * (phi_1 / h - 1.0)) * D1_0
-                )
-        return x_t
-
-    def multistep_dpm_solver_third_update(
-        self, x, model_prev_list, t_prev_list, t, solver_type="dpmsolver"
-    ):
-        """
-        Multistep solver DPM-Solver-3 from time `t_prev_list[-1]` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            model_prev_list: A list of pytorch tensor. The previous computed model values.
-            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        ns = self.noise_schedule
-        model_prev_2, model_prev_1, model_prev_0 = model_prev_list
-        t_prev_2, t_prev_1, t_prev_0 = t_prev_list
-        lambda_prev_2, lambda_prev_1, lambda_prev_0, lambda_t = (
-            ns.marginal_lambda(t_prev_2),
-            ns.marginal_lambda(t_prev_1),
-            ns.marginal_lambda(t_prev_0),
-            ns.marginal_lambda(t),
-        )
-        log_alpha_prev_0, log_alpha_t = (
-            ns.marginal_log_mean_coeff(t_prev_0),
-            ns.marginal_log_mean_coeff(t),
-        )
-        sigma_prev_0, sigma_t = ns.marginal_std(t_prev_0), ns.marginal_std(t)
-        alpha_t = torch.exp(log_alpha_t)
-
-        h_1 = lambda_prev_1 - lambda_prev_2
-        h_0 = lambda_prev_0 - lambda_prev_1
-        h = lambda_t - lambda_prev_0
-        r0, r1 = h_0 / h, h_1 / h
-        D1_0 = (1.0 / r0) * (model_prev_0 - model_prev_1)
-        D1_1 = (1.0 / r1) * (model_prev_1 - model_prev_2)
-        D1 = D1_0 + (r0 / (r0 + r1)) * (D1_0 - D1_1)
-        D2 = (1.0 / (r0 + r1)) * (D1_0 - D1_1)
-        if self.algorithm_type == "dpmsolver++":
-            phi_1 = torch.expm1(-h)
-            phi_2 = phi_1 / h + 1.0
-            phi_3 = phi_2 / h - 0.5
-            x_t = (
-                (sigma_t / sigma_prev_0) * x
-                - (alpha_t * phi_1) * model_prev_0
-                + (alpha_t * phi_2) * D1
-                - (alpha_t * phi_3) * D2
-            )
-        else:
-            phi_1 = torch.expm1(h)
-            phi_2 = phi_1 / h - 1.0
-            phi_3 = phi_2 / h - 0.5
-            x_t = (
-                (torch.exp(log_alpha_t - log_alpha_prev_0)) * x
-                - (sigma_t * phi_1) * model_prev_0
-                - (sigma_t * phi_2) * D1
-                - (sigma_t * phi_3) * D2
-            )
-        return x_t
-
-    def singlestep_dpm_solver_update(
-        self,
-        x,
-        s,
-        t,
-        order,
-        return_intermediate=False,
-        solver_type="dpmsolver",
-        r1=None,
-        r2=None,
-    ):
-        """
-        Singlestep DPM-Solver with the order `order` from time `s` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            s: A pytorch tensor. The starting time, with the shape (1,).
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
-            return_intermediate: A `bool`. If true, also return the model value at time `s`, `s1` and `s2` (the intermediate times).
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-            r1: A `float`. The hyperparameter of the second-order or third-order solver.
-            r2: A `float`. The hyperparameter of the third-order solver.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if order == 1:
-            return self.dpm_solver_first_update(
-                x, s, t, return_intermediate=return_intermediate
-            )
-        elif order == 2:
-            return self.singlestep_dpm_solver_second_update(
-                x,
-                s,
-                t,
-                return_intermediate=return_intermediate,
-                solver_type=solver_type,
-                r1=r1,
-            )
-        elif order == 3:
-            return self.singlestep_dpm_solver_third_update(
-                x,
-                s,
-                t,
-                return_intermediate=return_intermediate,
-                solver_type=solver_type,
-                r1=r1,
-                r2=r2,
-            )
-        else:
-            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
-
-    def multistep_dpm_solver_update(
-        self, x, model_prev_list, t_prev_list, t, order, solver_type="dpmsolver"
-    ):
-        """
-        Multistep DPM-Solver with the order `order` from time `t_prev_list[-1]` to time `t`.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `s`.
-            model_prev_list: A list of pytorch tensor. The previous computed model values.
-            t_prev_list: A list of pytorch tensor. The previous times, each time has the shape (1,)
-            t: A pytorch tensor. The ending time, with the shape (1,).
-            order: A `int`. The order of DPM-Solver. We only support order == 1 or 2 or 3.
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_t: A pytorch tensor. The approximated solution at time `t`.
-        """
-        if order == 1:
-            return self.dpm_solver_first_update(
-                x, t_prev_list[-1], t, model_s=model_prev_list[-1]
-            )
-        elif order == 2:
-            return self.multistep_dpm_solver_second_update(
-                x, model_prev_list, t_prev_list, t, solver_type=solver_type
-            )
-        elif order == 3:
-            return self.multistep_dpm_solver_third_update(
-                x, model_prev_list, t_prev_list, t, solver_type=solver_type
-            )
-        else:
-            raise ValueError(f"Solver order must be 1 or 2 or 3, got {order}")
-
-    def dpm_solver_adaptive(
-        self,
-        x,
-        order,
-        t_T,
-        t_0,
-        h_init=0.05,
-        atol=0.0078,
-        rtol=0.05,
-        theta=0.9,
-        t_err=1e-5,
-        solver_type="dpmsolver",
-    ):
-        """
-        The adaptive step size solver based on singlestep DPM-Solver.
-
-        Args:
-            x: A pytorch tensor. The initial value at time `t_T`.
-            order: A `int`. The (higher) order of the solver. We only support order == 2 or 3.
-            t_T: A `float`. The starting time of the sampling (default is T).
-            t_0: A `float`. The ending time of the sampling (default is epsilon).
-            h_init: A `float`. The initial step size (for logSNR).
-            atol: A `float`. The absolute tolerance of the solver. For image data, the default setting is 0.0078, followed [1].
-            rtol: A `float`. The relative tolerance of the solver. The default setting is 0.05.
-            theta: A `float`. The safety hyperparameter for adapting the step size. The default setting is 0.9, followed [1].
-            t_err: A `float`. The tolerance for the time. We solve the diffusion ODE until the absolute error between the
-                current time and `t_0` is less than `t_err`. The default setting is 1e-5.
-            solver_type: either 'dpmsolver' or 'taylor'. The type for the high-order solvers.
-                The type slightly impacts the performance. We recommend to use 'dpmsolver' type.
-        Returns:
-            x_0: A pytorch tensor. The approximated solution at time `t_0`.
-
-        [1] A. Jolicoeur-Martineau, K. Li, R. Piché-Taillefer, T. Kachman, and I. Mitliagkas, "Gotta go fast when generating data with score-based models," arXiv preprint arXiv:2105.14080, 2021.
-        """
-        ns = self.noise_schedule
-        s = t_T * torch.ones((1,)).to(x)
-        lambda_s = ns.marginal_lambda(s)
-        lambda_0 = ns.marginal_lambda(t_0 * torch.ones_like(s).to(x))
-        h = h_init * torch.ones_like(s).to(x)
-        x_prev = x
-        nfe = 0
-        if order == 2:
-            r1 = 0.5
-            lower_update = lambda x, s, t: self.dpm_solver_first_update(
-                x, s, t, return_intermediate=True
-            )
-            higher_update = (
-                lambda x, s, t, **kwargs: self.singlestep_dpm_solver_second_update(
-                    x, s, t, r1=r1, solver_type=solver_type, **kwargs
-                )
-            )
-        elif order == 3:
-            r1, r2 = 1.0 / 3.0, 2.0 / 3.0
-            lower_update = lambda x, s, t: self.singlestep_dpm_solver_second_update(
-                x, s, t, r1=r1, return_intermediate=True, solver_type=solver_type
-            )
-            higher_update = (
-                lambda x, s, t, **kwargs: self.singlestep_dpm_solver_third_update(
-                    x, s, t, r1=r1, r2=r2, solver_type=solver_type, **kwargs
-                )
-            )
-        else:
-            raise ValueError(
-                f"For adaptive step size solver, order must be 2 or 3, got {order}"
-            )
-        while torch.abs(s - t_0).mean() > t_err:
-            t = ns.inverse_lambda(lambda_s + h)
-            x_lower, lower_noise_kwargs = lower_update(x, s, t)
-            x_higher = higher_update(x, s, t, **lower_noise_kwargs)
-            delta = torch.max(
-                torch.ones_like(x).to(x) * atol,
-                rtol * torch.max(torch.abs(x_lower), torch.abs(x_prev)),
-            )
-            norm_fn = lambda v: torch.sqrt(
-                torch.square(v.reshape((v.shape[0], -1))).mean(dim=-1, keepdim=True)
-            )
-            E = norm_fn((x_higher - x_lower) / delta).max()
-            if torch.all(E <= 1.0):
-                x = x_higher
-                s = t
-                x_prev = x_lower
-                lambda_s = ns.marginal_lambda(s)
-            h = torch.min(
-                theta * h * torch.float_power(E, -1.0 / order).float(),
-                lambda_0 - lambda_s,
-            )
-            nfe += order
-        print("adaptive solver nfe", nfe)
-        return x
-
-    def add_noise(self, x, t, noise=None):
-        """
-        Compute the noised input xt = alpha_t * x + sigma_t * noise.
-
-        Args:
-            x: A `torch.Tensor` with shape `(batch_size, *shape)`.
-            t: A `torch.Tensor` with shape `(t_size,)`.
-        Returns:
-            xt with shape `(t_size, batch_size, *shape)`.
-        """
-        alpha_t, sigma_t = (
-            self.noise_schedule.marginal_alpha(t),
-            self.noise_schedule.marginal_std(t),
-        )
-        if noise is None:
-            noise = torch.randn((t.shape[0], *x.shape), device=x.device)
-        x = x.reshape((-1, *x.shape))
-        xt = expand_dims(alpha_t, x.dim()) * x + expand_dims(sigma_t, x.dim()) * noise
-        if t.shape[0] == 1:
-            return xt.squeeze(0)
-        else:
-            return xt
-
-    def inverse(
-        self,
-        x,
-        steps=20,
-        t_start=None,
-        t_end=None,
-        order=2,
-        skip_type="time_uniform",
-        method="multistep",
-        lower_order_final=True,
-        denoise_to_zero=False,
-        solver_type="dpmsolver",
-        atol=0.0078,
-        rtol=0.05,
-        return_intermediate=False,
-    ):
-        """
-        Inverse the sample `x` from time `t_start` to `t_end` by DPM-Solver.
-        For discrete-time DPMs, we use `t_start=1/N`, where `N` is the total time steps during training.
-        """
-        t_0 = 1.0 / self.noise_schedule.total_N if t_start is None else t_start
-        t_T = self.noise_schedule.T if t_end is None else t_end
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-        return self.sample(
-            x,
-            steps=steps,
-            t_start=t_0,
-            t_end=t_T,
-            order=order,
-            skip_type=skip_type,
-            method=method,
-            lower_order_final=lower_order_final,
-            denoise_to_zero=denoise_to_zero,
-            solver_type=solver_type,
-            atol=atol,
-            rtol=rtol,
-            return_intermediate=return_intermediate,
-        )
-
-    def sample(
-        self,
-        x,
-        steps=20,
-        t_start=None,
-        t_end=None,
-        order=2,
-        skip_type="time_uniform",
-        method="multistep",
-        lower_order_final=True,
-        denoise_to_zero=False,
-        solver_type="dpmsolver",
-        atol=0.0078,
-        rtol=0.05,
-        return_intermediate=False,
-        flow_shift=1.0,
-    ):
-        """
-        Compute the sample at time `t_end` by DPM-Solver, given the initial `x` at time `t_start`.
-
-        =====================================================
-
-        We support the following algorithms for both noise prediction model and data prediction model:
-            - 'singlestep':
-                Singlestep DPM-Solver (i.e. "DPM-Solver-fast" in the paper), which combines different orders of singlestep DPM-Solver.
-                We combine all the singlestep solvers with order <= `order` to use up all the function evaluations (steps).
-                The total number of function evaluations (NFE) == `steps`.
-                Given a fixed NFE == `steps`, the sampling procedure is:
-                    - If `order` == 1:
-                        - Denote K = steps. We use K steps of DPM-Solver-1 (i.e. DDIM).
-                    - If `order` == 2:
-                        - Denote K = (steps // 2) + (steps % 2). We take K intermediate time steps for sampling.
-                        - If steps % 2 == 0, we use K steps of singlestep DPM-Solver-2.
-                        - If steps % 2 == 1, we use (K - 1) steps of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
-                    - If `order` == 3:
-                        - Denote K = (steps // 3 + 1). We take K intermediate time steps for sampling.
-                        - If steps % 3 == 0, we use (K - 2) steps of singlestep DPM-Solver-3, and 1 step of singlestep DPM-Solver-2 and 1 step of DPM-Solver-1.
-                        - If steps % 3 == 1, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of DPM-Solver-1.
-                        - If steps % 3 == 2, we use (K - 1) steps of singlestep DPM-Solver-3 and 1 step of singlestep DPM-Solver-2.
-            - 'multistep':
-                Multistep DPM-Solver with the order of `order`. The total number of function evaluations (NFE) == `steps`.
-                We initialize the first `order` values by lower order multistep solvers.
-                Given a fixed NFE == `steps`, the sampling procedure is:
-                    Denote K = steps.
-                    - If `order` == 1:
-                        - We use K steps of DPM-Solver-1 (i.e. DDIM).
-                    - If `order` == 2:
-                        - We firstly use 1 step of DPM-Solver-1, then use (K - 1) step of multistep DPM-Solver-2.
-                    - If `order` == 3:
-                        - We firstly use 1 step of DPM-Solver-1, then 1 step of multistep DPM-Solver-2, then (K - 2) step of multistep DPM-Solver-3.
-            - 'singlestep_fixed':
-                Fixed order singlestep DPM-Solver (i.e. DPM-Solver-1 or singlestep DPM-Solver-2 or singlestep DPM-Solver-3).
-                We use singlestep DPM-Solver-`order` for `order`=1 or 2 or 3, with total [`steps` // `order`] * `order` NFE.
-            - 'adaptive':
-                Adaptive step size DPM-Solver (i.e. "DPM-Solver-12" and "DPM-Solver-23" in the paper).
-                We ignore `steps` and use adaptive step size DPM-Solver with a higher order of `order`.
-                You can adjust the absolute tolerance `atol` and the relative tolerance `rtol` to balance the computatation costs
-                (NFE) and the sample quality.
-                    - If `order` == 2, we use DPM-Solver-12 which combines DPM-Solver-1 and singlestep DPM-Solver-2.
-                    - If `order` == 3, we use DPM-Solver-23 which combines singlestep DPM-Solver-2 and singlestep DPM-Solver-3.
-
-        =====================================================
-
-        Some advices for choosing the algorithm:
-            - For **unconditional sampling** or **guided sampling with small guidance scale** by DPMs:
-                Use singlestep DPM-Solver or DPM-Solver++ ("DPM-Solver-fast" in the paper) with `order = 3`.
-                e.g., DPM-Solver:
-                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver")
-                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
-                            skip_type='time_uniform', method='singlestep')
-                e.g., DPM-Solver++:
-                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
-                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=3,
-                            skip_type='time_uniform', method='singlestep')
-            - For **guided sampling with large guidance scale** by DPMs:
-                Use multistep DPM-Solver with `algorithm_type="dpmsolver++"` and `order = 2`.
-                e.g.
-                    >>> dpm_solver = DPM_Solver(model_fn, noise_schedule, algorithm_type="dpmsolver++")
-                    >>> x_sample = dpm_solver.sample(x, steps=steps, t_start=t_start, t_end=t_end, order=2,
-                            skip_type='time_uniform', method='multistep')
-
-        We support three types of `skip_type`:
-            - 'logSNR': uniform logSNR for the time steps. **Recommended for low-resolutional images**
-            - 'time_uniform': uniform time for the time steps. **Recommended for high-resolutional images**.
-            - 'time_quadratic': quadratic time for the time steps.
-
-        =====================================================
-        Args:
-            x: A pytorch tensor. The initial value at time `t_start`
-                e.g. if `t_start` == T, then `x` is a sample from the standard normal distribution.
-            steps: A `int`. The total number of function evaluations (NFE).
-            t_start: A `float`. The starting time of the sampling.
-                If `T` is None, we use self.noise_schedule.T (default is 1.0).
-            t_end: A `float`. The ending time of the sampling.
-                If `t_end` is None, we use 1. / self.noise_schedule.total_N.
-                e.g. if total_N == 1000, we have `t_end` == 1e-3.
-                For discrete-time DPMs:
-                    - We recommend `t_end` == 1. / self.noise_schedule.total_N.
-                For continuous-time DPMs:
-                    - We recommend `t_end` == 1e-3 when `steps` <= 15; and `t_end` == 1e-4 when `steps` > 15.
-            order: A `int`. The order of DPM-Solver.
-            skip_type: A `str`. The type for the spacing of the time steps. 'time_uniform' or 'logSNR' or 'time_quadratic'.
-            method: A `str`. The method for sampling. 'singlestep' or 'multistep' or 'singlestep_fixed' or 'adaptive'.
-            denoise_to_zero: A `bool`. Whether to denoise to time 0 at the final step.
-                Default is `False`. If `denoise_to_zero` is `True`, the total NFE is (`steps` + 1).
-
-                This trick is firstly proposed by DDPM (https://arxiv.org/abs/2006.11239) and
-                score_sde (https://arxiv.org/abs/2011.13456). Such trick can improve the FID
-                for diffusion models sampling by diffusion SDEs for low-resolutional images
-                (such as CIFAR-10). However, we observed that such trick does not matter for
-                high-resolutional images. As it needs an additional NFE, we do not recommend
-                it for high-resolutional images.
-            lower_order_final: A `bool`. Whether to use lower order solvers at the final steps.
-                Only valid for `method=multistep` and `steps < 15`. We empirically find that
-                this trick is a key to stabilizing the sampling by DPM-Solver with very few steps
-                (especially for steps <= 10). So we recommend to set it to be `True`.
-            solver_type: A `str`. The taylor expansion type for the solver. `dpmsolver` or `taylor`. We recommend `dpmsolver`.
-            atol: A `float`. The absolute tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
-            rtol: A `float`. The relative tolerance of the adaptive step size solver. Valid when `method` == 'adaptive'.
-            return_intermediate: A `bool`. Whether to save the xt at each step.
-                When set to `True`, method returns a tuple (x0, intermediates); when set to False, method returns only x0.
-        Returns:
-            x_end: A pytorch tensor. The approximated solution at time `t_end`.
-
-        """
-        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
-        t_T = self.noise_schedule.T if t_start is None else t_start
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-        if return_intermediate:
-            assert method in [
-                "multistep",
-                "singlestep",
-                "singlestep_fixed",
-            ], "Cannot use adaptive solver when saving intermediate values"
-        if self.correcting_xt_fn is not None:
-            assert method in [
-                "multistep",
-                "singlestep",
-                "singlestep_fixed",
-            ], "Cannot use adaptive solver when correcting_xt_fn is not None"
-        device = x.device
-        intermediates = []
-        with torch.no_grad():
-            if method == "adaptive":
-                x = self.dpm_solver_adaptive(
-                    x,
-                    order=order,
-                    t_T=t_T,
-                    t_0=t_0,
-                    atol=atol,
-                    rtol=rtol,
-                    solver_type=solver_type,
-                )
-            elif method == "multistep":
-                assert steps >= order
-                timesteps = self.get_time_steps(
-                    skip_type=skip_type,
-                    t_T=t_T,
-                    t_0=t_0,
-                    N=steps,
-                    device=device,
-                    shift=flow_shift,
-                )
-                assert timesteps.shape[0] - 1 == steps
-                # Init the initial values.
-                step = 0
-                t = timesteps[step]
-                t_prev_list = [t]
-                model_prev_list = [self.model_fn(x, t)]
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-                self.update_progress(step + 1, len(timesteps))
-                # Init the first `order` values by lower order multistep DPM-Solver.
-                for step in range(1, order):
-                    t = timesteps[step]
-                    x = self.multistep_dpm_solver_update(
-                        x,
-                        model_prev_list,
-                        t_prev_list,
-                        t,
-                        step,
-                        solver_type=solver_type,
-                    )
-                    if self.correcting_xt_fn is not None:
-                        x = self.correcting_xt_fn(x, t, step)
-                    if return_intermediate:
-                        intermediates.append(x)
-                    t_prev_list.append(t)
-                    model_prev_list.append(self.model_fn(x, t))
-                    # update progress bar
-                    self.update_progress(step + 1, len(timesteps))
-                # Compute the remaining values by `order`-th order multistep DPM-Solver.
-                for step in tqdm(
-                    range(order, steps + 1),
-                    disable=os.getenv("DPM_TQDM", "False") == "True",
-                ):
-                    t = timesteps[step]
-                    # We only use lower order for steps < 10
-                    # if lower_order_final and steps < 10:
-                    if lower_order_final:  # recommended by Shuchen Xue
-                        step_order = min(order, steps + 1 - step)
-                    else:
-                        step_order = order
-                    x = self.multistep_dpm_solver_update(
-                        x,
-                        model_prev_list,
-                        t_prev_list,
-                        t,
-                        step_order,
-                        solver_type=solver_type,
-                    )
-                    if self.correcting_xt_fn is not None:
-                        x = self.correcting_xt_fn(x, t, step)
-                    if return_intermediate:
-                        intermediates.append(x)
-                    for i in range(order - 1):
-                        t_prev_list[i] = t_prev_list[i + 1]
-                        model_prev_list[i] = model_prev_list[i + 1]
-                    t_prev_list[-1] = t
-                    # We do not need to evaluate the final model value.
-                    if step < steps:
-                        model_prev_list[-1] = self.model_fn(x, t)
-                    # update progress bar
-                    self.update_progress(step + 1, len(timesteps))
-            elif method in ["singlestep", "singlestep_fixed"]:
-                if method == "singlestep":
-                    timesteps_outer, orders = (
-                        self.get_orders_and_timesteps_for_singlestep_solver(
-                            steps=steps,
-                            order=order,
-                            skip_type=skip_type,
-                            t_T=t_T,
-                            t_0=t_0,
-                            device=device,
-                        )
-                    )
-                elif method == "singlestep_fixed":
-                    K = steps // order
-                    orders = [
-                        order,
-                    ] * K
-                    timesteps_outer = self.get_time_steps(
-                        skip_type=skip_type, t_T=t_T, t_0=t_0, N=K, device=device
-                    )
-                for step, order in enumerate(orders):
-                    s, t = timesteps_outer[step], timesteps_outer[step + 1]
-                    timesteps_inner = self.get_time_steps(
-                        skip_type=skip_type,
-                        t_T=s.item(),
-                        t_0=t.item(),
-                        N=order,
-                        device=device,
-                    )
-                    lambda_inner = self.noise_schedule.marginal_lambda(timesteps_inner)
-                    h = lambda_inner[-1] - lambda_inner[0]
-                    r1 = None if order <= 1 else (lambda_inner[1] - lambda_inner[0]) / h
-                    r2 = None if order <= 2 else (lambda_inner[2] - lambda_inner[0]) / h
-                    x = self.singlestep_dpm_solver_update(
-                        x, s, t, order, solver_type=solver_type, r1=r1, r2=r2
-                    )
-                    if self.correcting_xt_fn is not None:
-                        x = self.correcting_xt_fn(x, t, step)
-                    if return_intermediate:
-                        intermediates.append(x)
-                    self.update_progress(step + 1, len(timesteps_outer))
-            else:
-                raise ValueError(f"Got wrong method {method}")
-            if denoise_to_zero:
-                t = torch.ones((1,)).to(device) * t_0
-                x = self.denoise_to_zero_fn(x, t)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step + 1)
-                if return_intermediate:
-                    intermediates.append(x)
-        if return_intermediate:
-            return x, intermediates
-        else:
-            return x
-
-
-#############################################################
-# other utility functions
-#############################################################
-
-
-def interpolate_fn(x, xp, yp):
-    """
-    A piecewise linear function y = f(x), using xp and yp as keypoints.
-    We implement f(x) in a differentiable way (i.e. applicable for autograd).
-    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
-
-    Args:
-        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
-        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
-        yp: PyTorch tensor with shape [C, K].
-    Returns:
-        The function values f(x), with shape [N, C].
-    """
-    N, K = x.shape[0], xp.shape[1]
-    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
-    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
-    x_idx = torch.argmin(x_indices, dim=2)
-    cand_start_idx = x_idx - 1
-    start_idx = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(1, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    end_idx = torch.where(
-        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
-    )
-    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
-    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
-    start_idx2 = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(0, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
-    start_y = torch.gather(
-        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
-    ).squeeze(2)
-    end_y = torch.gather(
-        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
-    ).squeeze(2)
-    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
-    return cand
-
-
-def expand_dims(v, dims):
-    """
-    Expand the tensor `v` to the dim `dims`.
-
-    Args:
-        `v`: a PyTorch tensor with shape [N].
-        `dim`: a `int`.
-    Returns:
-        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
-    """
-    return v[(...,) + (None,) * (dims - 1)]
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/edm_sample.py b/sana/sana_600M/packages/Sana/diffusion/model/edm_sample.py
deleted file mode 100755
index 9e99301eb..000000000
--- a/sana/sana_600M/packages/Sana/diffusion/model/edm_sample.py
+++ /dev/null
@@ -1,281 +0,0 @@
-# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-
-# Modified from OpenAI's diffusion repos
-#     GLIDE: https://github.com/openai/glide-text2im/blob/main/glide_text2im/gaussian_diffusion.py
-#     ADM:   https://github.com/openai/guided-diffusion/blob/main/guided_diffusion
-#     IDDPM: https://github.com/openai/improved-diffusion/blob/main/improved_diffusion/gaussian_diffusion.py
-
-
-import numpy as np
-from diffusion.model.utils import *
-from tqdm import tqdm
-
-# ----------------------------------------------------------------------------
-# Proposed EDM sampler (Algorithm 2).
-
-
-def edm_sampler(
-    net,
-    latents,
-    class_labels=None,
-    cfg_scale=None,
-    randn_like=torch.randn_like,
-    num_steps=18,
-    sigma_min=0.002,
-    sigma_max=80,
-    rho=7,
-    S_churn=0,
-    S_min=0,
-    S_max=float("inf"),
-    S_noise=1,
-    **kwargs,
-):
-    # Adjust noise levels based on what's supported by the network.
-    sigma_min = max(sigma_min, net.sigma_min)
-    sigma_max = min(sigma_max, net.sigma_max)
-
-    # Time step discretization.
-    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
-    t_steps = (
-        sigma_max ** (1 / rho)
-        + step_indices
-        / (num_steps - 1)
-        * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
-    ) ** rho
-    t_steps = torch.cat(
-        [net.round_sigma(t_steps), torch.zeros_like(t_steps[:1])]
-    )  # t_N = 0
-
-    # Main sampling loop.
-    x_next = latents.to(torch.float64) * t_steps[0]
-    for i, (t_cur, t_next) in tqdm(
-        list(enumerate(zip(t_steps[:-1], t_steps[1:])))
-    ):  # 0, ..., N-1
-        x_cur = x_next
-
-        # Increase noise temporarily.
-        gamma = (
-            min(S_churn / num_steps, np.sqrt(2) - 1) if S_min <= t_cur <= S_max else 0
-        )
-        t_hat = net.round_sigma(t_cur + gamma * t_cur)
-        x_hat = x_cur + (t_hat**2 - t_cur**2).sqrt() * S_noise * randn_like(x_cur)
-
-        # Euler step.
-        denoised = net(x_hat.float(), t_hat, class_labels, cfg_scale, **kwargs)["x"].to(
-            torch.float64
-        )
-        d_cur = (x_hat - denoised) / t_hat
-        x_next = x_hat + (t_next - t_hat) * d_cur
-
-        # Apply 2nd order correction.
-        if i < num_steps - 1:
-            denoised = net(x_next.float(), t_next, class_labels, cfg_scale, **kwargs)[
-                "x"
-            ].to(torch.float64)
-            d_prime = (x_next - denoised) / t_next
-            x_next = x_hat + (t_next - t_hat) * (0.5 * d_cur + 0.5 * d_prime)
-
-    return x_next
-
-
-# ----------------------------------------------------------------------------
-# Generalized ablation sampler, representing the superset of all sampling
-# methods discussed in the paper.
-
-
-def ablation_sampler(
-    net,
-    latents,
-    class_labels=None,
-    cfg_scale=None,
-    feat=None,
-    randn_like=torch.randn_like,
-    num_steps=18,
-    sigma_min=None,
-    sigma_max=None,
-    rho=7,
-    solver="heun",
-    discretization="edm",
-    schedule="linear",
-    scaling="none",
-    epsilon_s=1e-3,
-    C_1=0.001,
-    C_2=0.008,
-    M=1000,
-    alpha=1,
-    S_churn=0,
-    S_min=0,
-    S_max=float("inf"),
-    S_noise=1,
-):
-    assert solver in ["euler", "heun"]
-    assert discretization in ["vp", "ve", "iddpm", "edm"]
-    assert schedule in ["vp", "ve", "linear"]
-    assert scaling in ["vp", "none"]
-
-    # Helper functions for VP & VE noise level schedules.
-    vp_sigma = (
-        lambda beta_d, beta_min: lambda t: (
-            np.e ** (0.5 * beta_d * (t**2) + beta_min * t) - 1
-        )
-        ** 0.5
-    )
-    vp_sigma_deriv = (
-        lambda beta_d, beta_min: lambda t: 0.5
-        * (beta_min + beta_d * t)
-        * (sigma(t) + 1 / sigma(t))
-    )
-    vp_sigma_inv = (
-        lambda beta_d, beta_min: lambda sigma: (
-            (beta_min**2 + 2 * beta_d * (sigma**2 + 1).log()).sqrt() - beta_min
-        )
-        / beta_d
-    )
-    ve_sigma = lambda t: t.sqrt()
-    ve_sigma_deriv = lambda t: 0.5 / t.sqrt()
-    ve_sigma_inv = lambda sigma: sigma**2
-
-    # Select default noise level range based on the specified time step discretization.
-    if sigma_min is None:
-        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=epsilon_s)
-        sigma_min = {"vp": vp_def, "ve": 0.02, "iddpm": 0.002, "edm": 0.002}[
-            discretization
-        ]
-    if sigma_max is None:
-        vp_def = vp_sigma(beta_d=19.1, beta_min=0.1)(t=1)
-        sigma_max = {"vp": vp_def, "ve": 100, "iddpm": 81, "edm": 80}[discretization]
-
-    # Adjust noise levels based on what's supported by the network.
-    sigma_min = max(sigma_min, net.sigma_min)
-    sigma_max = min(sigma_max, net.sigma_max)
-
-    # Compute corresponding betas for VP.
-    vp_beta_d = (
-        2
-        * (np.log(sigma_min**2 + 1) / epsilon_s - np.log(sigma_max**2 + 1))
-        / (epsilon_s - 1)
-    )
-    vp_beta_min = np.log(sigma_max**2 + 1) - 0.5 * vp_beta_d
-
-    # Define time steps in terms of noise level.
-    step_indices = torch.arange(num_steps, dtype=torch.float64, device=latents.device)
-    if discretization == "vp":
-        orig_t_steps = 1 + step_indices / (num_steps - 1) * (epsilon_s - 1)
-        sigma_steps = vp_sigma(vp_beta_d, vp_beta_min)(orig_t_steps)
-    elif discretization == "ve":
-        orig_t_steps = (sigma_max**2) * (
-            (sigma_min**2 / sigma_max**2) ** (step_indices / (num_steps - 1))
-        )
-        sigma_steps = ve_sigma(orig_t_steps)
-    elif discretization == "iddpm":
-        u = torch.zeros(M + 1, dtype=torch.float64, device=latents.device)
-        alpha_bar = lambda j: (0.5 * np.pi * j / M / (C_2 + 1)).sin() ** 2
-        for j in torch.arange(M, 0, -1, device=latents.device):  # M, ..., 1
-            u[j - 1] = (
-                (u[j] ** 2 + 1) / (alpha_bar(j - 1) / alpha_bar(j)).clip(min=C_1) - 1
-            ).sqrt()
-        u_filtered = u[torch.logical_and(u >= sigma_min, u <= sigma_max)]
-        sigma_steps = u_filtered[
-            ((len(u_filtered) - 1) / (num_steps - 1) * step_indices)
-            .round()
-            .to(torch.int64)
-        ]
-    else:
-        assert discretization == "edm"
-        sigma_steps = (
-            sigma_max ** (1 / rho)
-            + step_indices
-            / (num_steps - 1)
-            * (sigma_min ** (1 / rho) - sigma_max ** (1 / rho))
-        ) ** rho
-
-    # Define noise level schedule.
-    if schedule == "vp":
-        sigma = vp_sigma(vp_beta_d, vp_beta_min)
-        sigma_deriv = vp_sigma_deriv(vp_beta_d, vp_beta_min)
-        sigma_inv = vp_sigma_inv(vp_beta_d, vp_beta_min)
-    elif schedule == "ve":
-        sigma = ve_sigma
-        sigma_deriv = ve_sigma_deriv
-        sigma_inv = ve_sigma_inv
-    else:
-        assert schedule == "linear"
-        sigma = lambda t: t
-        sigma_deriv = lambda t: 1
-        sigma_inv = lambda sigma: sigma
-
-    # Define scaling schedule.
-    if scaling == "vp":
-        s = lambda t: 1 / (1 + sigma(t) ** 2).sqrt()
-        s_deriv = lambda t: -sigma(t) * sigma_deriv(t) * (s(t) ** 3)
-    else:
-        assert scaling == "none"
-        s = lambda t: 1
-        s_deriv = lambda t: 0
-
-    # Compute final time steps based on the corresponding noise levels.
-    t_steps = sigma_inv(net.round_sigma(sigma_steps))
-    t_steps = torch.cat([t_steps, torch.zeros_like(t_steps[:1])])  # t_N = 0
-
-    # Main sampling loop.
-    t_next = t_steps[0]
-    x_next = latents.to(torch.float64) * (sigma(t_next) * s(t_next))
-    for i, (t_cur, t_next) in enumerate(zip(t_steps[:-1], t_steps[1:])):  # 0, ..., N-1
-        x_cur = x_next
-
-        # Increase noise temporarily.
-        gamma = (
-            min(S_churn / num_steps, np.sqrt(2) - 1)
-            if S_min <= sigma(t_cur) <= S_max
-            else 0
-        )
-        t_hat = sigma_inv(net.round_sigma(sigma(t_cur) + gamma * sigma(t_cur)))
-        x_hat = s(t_hat) / s(t_cur) * x_cur + (
-            sigma(t_hat) ** 2 - sigma(t_cur) ** 2
-        ).clip(min=0).sqrt() * s(t_hat) * S_noise * randn_like(x_cur)
-
-        # Euler step.
-        h = t_next - t_hat
-        denoised = net(
-            x_hat.float() / s(t_hat), sigma(t_hat), class_labels, cfg_scale, feat=feat
-        )["x"].to(torch.float64)
-        d_cur = (
-            sigma_deriv(t_hat) / sigma(t_hat) + s_deriv(t_hat) / s(t_hat)
-        ) * x_hat - sigma_deriv(t_hat) * s(t_hat) / sigma(t_hat) * denoised
-        x_prime = x_hat + alpha * h * d_cur
-        t_prime = t_hat + alpha * h
-
-        # Apply 2nd order correction.
-        if solver == "euler" or i == num_steps - 1:
-            x_next = x_hat + h * d_cur
-        else:
-            assert solver == "heun"
-            denoised = net(
-                x_prime.float() / s(t_prime),
-                sigma(t_prime),
-                class_labels,
-                cfg_scale,
-                feat=feat,
-            )["x"].to(torch.float64)
-            d_prime = (
-                sigma_deriv(t_prime) / sigma(t_prime) + s_deriv(t_prime) / s(t_prime)
-            ) * x_prime - sigma_deriv(t_prime) * s(t_prime) / sigma(t_prime) * denoised
-            x_next = x_hat + h * (
-                (1 - 1 / (2 * alpha)) * d_cur + 1 / (2 * alpha) * d_prime
-            )
-
-    return x_next
diff --git a/sana/sana_600M/packages/Sana/diffusion/model/sa_solver.py b/sana/sana_600M/packages/Sana/diffusion/model/sa_solver.py
deleted file mode 100755
index b7566cef5..000000000
--- a/sana/sana_600M/packages/Sana/diffusion/model/sa_solver.py
+++ /dev/null
@@ -1,1618 +0,0 @@
-# Copyright 2024 NVIDIA CORPORATION & AFFILIATES
-#
-# Licensed under the Apache License, Version 2.0 (the "License");
-# you may not use this file except in compliance with the License.
-# You may obtain a copy of the License at
-#
-#     http://www.apache.org/licenses/LICENSE-2.0
-#
-# Unless required by applicable law or agreed to in writing, software
-# distributed under the License is distributed on an "AS IS" BASIS,
-# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
-# See the License for the specific language governing permissions and
-# limitations under the License.
-#
-# SPDX-License-Identifier: Apache-2.0
-
-import math
-
-import torch
-from tqdm import tqdm
-
-
-class NoiseScheduleVP:
-    def __init__(
-        self,
-        schedule="discrete",
-        betas=None,
-        alphas_cumprod=None,
-        continuous_beta_0=0.1,
-        continuous_beta_1=20.0,
-        dtype=torch.float32,
-    ):
-        """Thanks to DPM-Solver for their code base"""
-        r"""Create a wrapper class for the forward SDE (VP type).
-        ***
-        Update: We support discrete-time diffusion models by implementing a picewise linear interpolation for log_alpha_t.
-                We recommend to use schedule='discrete' for the discrete-time diffusion models, especially for high-resolution images.
-        ***
-        The forward SDE ensures that the condition distribution q_{t|0}(x_t | x_0) = N ( alpha_t * x_0, sigma_t^2 * I ).
-        We further define lambda_t = log(alpha_t) - log(sigma_t), which is the half-logSNR (described in the DPM-Solver paper).
-        Therefore, we implement the functions for computing alpha_t, sigma_t and lambda_t. For t in [0, T], we have:
-            log_alpha_t = self.marginal_log_mean_coeff(t)
-            sigma_t = self.marginal_std(t)
-            lambda_t = self.marginal_lambda(t)
-        Moreover, as lambda(t) is an invertible function, we also support its inverse function:
-            t = self.inverse_lambda(lambda_t)
-        ===============================================================
-        We support both discrete-time DPMs (trained on n = 0, 1, ..., N-1) and continuous-time DPMs (trained on t in [t_0, T]).
-        1. For discrete-time DPMs:
-            For discrete-time DPMs trained on n = 0, 1, ..., N-1, we convert the discrete steps to continuous time steps by:
-                t_i = (i + 1) / N
-            e.g. for N = 1000, we have t_0 = 1e-3 and T = t_{N-1} = 1.
-            We solve the corresponding diffusion ODE from time T = 1 to time t_0 = 1e-3.
-            Args:
-                betas: A `torch.Tensor`. The beta array for the discrete-time DPM. (See the original DDPM paper for details)
-                alphas_cumprod: A `torch.Tensor`. The cumprod alphas for the discrete-time DPM. (See the original DDPM paper for details)
-            Note that we always have alphas_cumprod = cumprod(1 - betas). Therefore, we only need to set one of `betas` and `alphas_cumprod`.
-            **Important**:  Please pay special attention for the args for `alphas_cumprod`:
-                The `alphas_cumprod` is the \hat{alpha_n} arrays in the notations of DDPM. Specifically, DDPMs assume that
-                    q_{t_n | 0}(x_{t_n} | x_0) = N ( \sqrt{\hat{alpha_n}} * x_0, (1 - \hat{alpha_n}) * I ).
-                Therefore, the notation \hat{alpha_n} is different from the notation alpha_t in DPM-Solver. In fact, we have
-                    alpha_{t_n} = \sqrt{\hat{alpha_n}},
-                and
-                    log(alpha_{t_n}) = 0.5 * log(\hat{alpha_n}).
-        2. For continuous-time DPMs:
-            We support two types of VPSDEs: linear (DDPM) and cosine (improved-DDPM). The hyperparameters for the noise
-            schedule are the default settings in DDPM and improved-DDPM:
-            Args:
-                beta_min: A `float` number. The smallest beta for the linear schedule.
-                beta_max: A `float` number. The largest beta for the linear schedule.
-                cosine_s: A `float` number. The hyperparameter in the cosine schedule.
-                cosine_beta_max: A `float` number. The hyperparameter in the cosine schedule.
-                T: A `float` number. The ending time of the forward process.
-        ===============================================================
-        Args:
-            schedule: A `str`. The noise schedule of the forward SDE. 'discrete' for discrete-time DPMs,
-                    'linear' or 'cosine' for continuous-time DPMs.
-        Returns:
-            A wrapper object of the forward SDE (VP type).
-
-        ===============================================================
-        Example:
-        # For discrete-time DPMs, given betas (the beta array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', betas=betas)
-        # For discrete-time DPMs, given alphas_cumprod (the \hat{alpha_n} array for n = 0, 1, ..., N - 1):
-        >>> ns = NoiseScheduleVP('discrete', alphas_cumprod=alphas_cumprod)
-        # For continuous-time DPMs (VPSDE), linear schedule:
-        >>> ns = NoiseScheduleVP('linear', continuous_beta_0=0.1, continuous_beta_1=20.)
-        """
-
-        if schedule not in ["discrete", "linear", "cosine"]:
-            raise ValueError(
-                "Unsupported noise schedule {}. The schedule needs to be 'discrete' or 'linear' or 'cosine'".format(
-                    schedule
-                )
-            )
-
-        self.schedule = schedule
-        if schedule == "discrete":
-            if betas is not None:
-                log_alphas = 0.5 * torch.log(1 - betas).cumsum(dim=0)
-            else:
-                assert alphas_cumprod is not None
-                log_alphas = 0.5 * torch.log(alphas_cumprod)
-            self.total_N = len(log_alphas)
-            self.T = 1.0
-            self.t_array = (
-                torch.linspace(0.0, 1.0, self.total_N + 1)[1:]
-                .reshape((1, -1))
-                .to(dtype=dtype)
-            )
-            self.log_alpha_array = log_alphas.reshape(
-                (
-                    1,
-                    -1,
-                )
-            ).to(dtype=dtype)
-        else:
-            self.total_N = 1000
-            self.beta_0 = continuous_beta_0
-            self.beta_1 = continuous_beta_1
-            self.cosine_s = 0.008
-            self.cosine_beta_max = 999.0
-            self.cosine_t_max = (
-                math.atan(self.cosine_beta_max * (1.0 + self.cosine_s) / math.pi)
-                * 2.0
-                * (1.0 + self.cosine_s)
-                / math.pi
-                - self.cosine_s
-            )
-            self.cosine_log_alpha_0 = math.log(
-                math.cos(self.cosine_s / (1.0 + self.cosine_s) * math.pi / 2.0)
-            )
-            self.schedule = schedule
-            if schedule == "cosine":
-                # For the cosine schedule, T = 1 will have numerical issues. So we manually set the ending time T.
-                # Note that T = 0.9946 may be not the optimal setting. However, we find it works well.
-                self.T = 0.9946
-            else:
-                self.T = 1.0
-
-    def marginal_log_mean_coeff(self, t):
-        """
-        Compute log(alpha_t) of a given continuous-time label t in [0, T].
-        """
-        if self.schedule == "discrete":
-            return interpolate_fn(
-                t.reshape((-1, 1)),
-                self.t_array.to(t.device),
-                self.log_alpha_array.to(t.device),
-            ).reshape(-1)
-        elif self.schedule == "linear":
-            return -0.25 * t**2 * (self.beta_1 - self.beta_0) - 0.5 * t * self.beta_0
-        elif self.schedule == "cosine":
-            log_alpha_fn = lambda s: torch.log(
-                torch.cos((s + self.cosine_s) / (1.0 + self.cosine_s) * math.pi / 2.0)
-            )
-            log_alpha_t = log_alpha_fn(t) - self.cosine_log_alpha_0
-            return log_alpha_t
-
-    def marginal_alpha(self, t):
-        """
-        Compute alpha_t of a given continuous-time label t in [0, T].
-        """
-        return torch.exp(self.marginal_log_mean_coeff(t))
-
-    def marginal_std(self, t):
-        """
-        Compute sigma_t of a given continuous-time label t in [0, T].
-        """
-        return torch.sqrt(1.0 - torch.exp(2.0 * self.marginal_log_mean_coeff(t)))
-
-    def marginal_lambda(self, t):
-        """
-        Compute lambda_t = log(alpha_t) - log(sigma_t) of a given continuous-time label t in [0, T].
-        """
-        log_mean_coeff = self.marginal_log_mean_coeff(t)
-        log_std = 0.5 * torch.log(1.0 - torch.exp(2.0 * log_mean_coeff))
-        return log_mean_coeff - log_std
-
-    def inverse_lambda(self, lamb):
-        """
-        Compute the continuous-time label t in [0, T] of a given half-logSNR lambda_t.
-        """
-        if self.schedule == "linear":
-            tmp = (
-                2.0
-                * (self.beta_1 - self.beta_0)
-                * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
-            )
-            Delta = self.beta_0**2 + tmp
-            return tmp / (torch.sqrt(Delta) + self.beta_0) / (self.beta_1 - self.beta_0)
-        elif self.schedule == "discrete":
-            log_alpha = -0.5 * torch.logaddexp(
-                torch.zeros((1,)).to(lamb.device), -2.0 * lamb
-            )
-            t = interpolate_fn(
-                log_alpha.reshape((-1, 1)),
-                torch.flip(self.log_alpha_array.to(lamb.device), [1]),
-                torch.flip(self.t_array.to(lamb.device), [1]),
-            )
-            return t.reshape((-1,))
-        else:
-            log_alpha = -0.5 * torch.logaddexp(-2.0 * lamb, torch.zeros((1,)).to(lamb))
-            t_fn = (
-                lambda log_alpha_t: torch.arccos(
-                    torch.exp(log_alpha_t + self.cosine_log_alpha_0)
-                )
-                * 2.0
-                * (1.0 + self.cosine_s)
-                / math.pi
-                - self.cosine_s
-            )
-            t = t_fn(log_alpha)
-            return t
-
-    def edm_sigma(self, t):
-        return self.marginal_std(t) / self.marginal_alpha(t)
-
-    def edm_inverse_sigma(self, edmsigma):
-        alpha = 1 / (edmsigma**2 + 1).sqrt()
-        sigma = alpha * edmsigma
-        lambda_t = torch.log(alpha / sigma)
-        t = self.inverse_lambda(lambda_t)
-        return t
-
-
-def model_wrapper(
-    model,
-    noise_schedule,
-    model_type="noise",
-    model_kwargs={},
-    guidance_type="uncond",
-    condition=None,
-    unconditional_condition=None,
-    guidance_scale=1.0,
-    classifier_fn=None,
-    classifier_kwargs={},
-):
-    """Thanks to DPM-Solver for their code base"""
-    """Create a wrapper function for the noise prediction model.
-    SA-Solver needs to solve the continuous-time diffusion SDEs. For DPMs trained on discrete-time labels, we need to
-    firstly wrap the model function to a noise prediction model that accepts the continuous time as the input.
-    We support four types of the diffusion model by setting `model_type`:
-        1. "noise": noise prediction model. (Trained by predicting noise).
-        2. "x_start": data prediction model. (Trained by predicting the data x_0 at time 0).
-        3. "v": velocity prediction model. (Trained by predicting the velocity).
-            The "v" prediction is derivation detailed in Appendix D of [1], and is used in Imagen-Video [2].
-            [1] Salimans, Tim, and Jonathan Ho. "Progressive distillation for fast sampling of diffusion models."
-                arXiv preprint arXiv:2202.00512 (2022).
-            [2] Ho, Jonathan, et al. "Imagen Video: High Definition Video Generation with Diffusion Models."
-                arXiv preprint arXiv:2210.02303 (2022).
-
-        4. "score": marginal score function. (Trained by denoising score matching).
-            Note that the score function and the noise prediction model follows a simple relationship:
-            ```
-                noise(x_t, t) = -sigma_t * score(x_t, t)
-            ```
-    We support three types of guided sampling by DPMs by setting `guidance_type`:
-        1. "uncond": unconditional sampling by DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-        2. "classifier": classifier guidance sampling [3] by DPMs and another classifier.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, **model_kwargs) -> noise | x_start | v | score
-            ``
-            The input `classifier_fn` has the following format:
-            ``
-                classifier_fn(x, t_input, cond, **classifier_kwargs) -> logits(x, t_input, cond)
-            ``
-            [3] P. Dhariwal and A. Q. Nichol, "Diffusion models beat GANs on image synthesis,"
-                in Advances in Neural Information Processing Systems, vol. 34, 2021, pp. 8780-8794.
-        3. "classifier-free": classifier-free guidance sampling by conditional DPMs.
-            The input `model` has the following format:
-            ``
-                model(x, t_input, cond, **model_kwargs) -> noise | x_start | v | score
-            ``
-            And if cond == `unconditional_condition`, the model output is the unconditional DPM output.
-            [4] Ho, Jonathan, and Tim Salimans. "Classifier-free diffusion guidance."
-                arXiv preprint arXiv:2207.12598 (2022).
-
-    The `t_input` is the time label of the model, which may be discrete-time labels (i.e. 0 to 999)
-    or continuous-time labels (i.e. epsilon to T).
-    We wrap the model function to accept only `x` and `t_continuous` as inputs, and outputs the predicted noise:
-    ``
-        def model_fn(x, t_continuous) -> noise:
-            t_input = get_model_input_time(t_continuous)
-            return noise_pred(model, x, t_input, **model_kwargs)
-    ``
-    where `t_continuous` is the continuous time labels (i.e. epsilon to T). And we use `model_fn` for SA-Solver.
-    ===============================================================
-    Args:
-        model: A diffusion model with the corresponding format described above.
-        noise_schedule: A noise schedule object, such as NoiseScheduleVP.
-        model_type: A `str`. The parameterization type of the diffusion model.
-                    "noise" or "x_start" or "v" or "score".
-        model_kwargs: A `dict`. A dict for the other inputs of the model function.
-        guidance_type: A `str`. The type of the guidance for sampling.
-                    "uncond" or "classifier" or "classifier-free".
-        condition: A pytorch tensor. The condition for the guided sampling.
-                    Only used for "classifier" or "classifier-free" guidance type.
-        unconditional_condition: A pytorch tensor. The condition for the unconditional sampling.
-                    Only used for "classifier-free" guidance type.
-        guidance_scale: A `float`. The scale for the guided sampling.
-        classifier_fn: A classifier function. Only used for the classifier guidance.
-        classifier_kwargs: A `dict`. A dict for the other inputs of the classifier function.
-    Returns:
-        A noise prediction model that accepts the noised data and the continuous time as the inputs.
-    """
-
-    def get_model_input_time(t_continuous):
-        """
-        Convert the continuous-time `t_continuous` (in [epsilon, T]) to the model input time.
-        For discrete-time DPMs, we convert `t_continuous` in [1 / N, 1] to `t_input` in [0, 1000 * (N - 1) / N].
-        For continuous-time DPMs, we just use `t_continuous`.
-        """
-        if noise_schedule.schedule == "discrete":
-            return (t_continuous - 1.0 / noise_schedule.total_N) * 1000.0
-        else:
-            return t_continuous
-
-    def noise_pred_fn(x, t_continuous, cond=None):
-        t_input = get_model_input_time(t_continuous)
-        if cond is None:
-            output = model(x, t_input, **model_kwargs)
-        else:
-            output = model(x, t_input, cond, **model_kwargs)
-        if model_type == "noise":
-            return output
-        elif model_type == "x_start":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return (x - alpha_t[0] * output) / sigma_t[0]
-        elif model_type == "v":
-            alpha_t, sigma_t = (
-                noise_schedule.marginal_alpha(t_continuous),
-                noise_schedule.marginal_std(t_continuous),
-            )
-            return alpha_t[0] * output + sigma_t[0] * x
-        elif model_type == "score":
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            return -sigma_t[0] * output
-
-    def cond_grad_fn(x, t_input):
-        """
-        Compute the gradient of the classifier, i.e. nabla_{x} log p_t(cond | x_t).
-        """
-        with torch.enable_grad():
-            x_in = x.detach().requires_grad_(True)
-            log_prob = classifier_fn(x_in, t_input, condition, **classifier_kwargs)
-            return torch.autograd.grad(log_prob.sum(), x_in)[0]
-
-    def model_fn(x, t_continuous):
-        """
-        The noise predicition model function that is used for DPM-Solver.
-        """
-        if guidance_type == "uncond":
-            return noise_pred_fn(x, t_continuous)
-        elif guidance_type == "classifier":
-            assert classifier_fn is not None
-            t_input = get_model_input_time(t_continuous)
-            cond_grad = cond_grad_fn(x, t_input)
-            sigma_t = noise_schedule.marginal_std(t_continuous)
-            noise = noise_pred_fn(x, t_continuous)
-            return noise - guidance_scale * sigma_t * cond_grad
-        elif guidance_type == "classifier-free":
-            if guidance_scale == 1.0 or unconditional_condition is None:
-                return noise_pred_fn(x, t_continuous, cond=condition)
-            else:
-                x_in = torch.cat([x] * 2)
-                t_in = torch.cat([t_continuous] * 2)
-                c_in = torch.cat([unconditional_condition, condition])
-                noise_uncond, noise = noise_pred_fn(x_in, t_in, cond=c_in).chunk(2)
-                return noise_uncond + guidance_scale * (noise - noise_uncond)
-
-    assert model_type in ["noise", "x_start", "v", "score"]
-    assert guidance_type in ["uncond", "classifier", "classifier-free"]
-    return model_fn
-
-
-class SASolver:
-    def __init__(
-        self,
-        model_fn,
-        noise_schedule,
-        algorithm_type="data_prediction",
-        correcting_x0_fn=None,
-        correcting_xt_fn=None,
-        thresholding_max_val=1.0,
-        dynamic_thresholding_ratio=0.995,
-    ):
-        """
-        Construct a SA-Solver
-        The default value for algorithm_type is "data_prediction" and we recommend not to change it to
-        "noise_prediction". For details, please see Appendix A.2.4 in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        self.model = lambda x, t: model_fn(x, t.expand(x.shape[0]))
-        self.noise_schedule = noise_schedule
-        assert algorithm_type in ["data_prediction", "noise_prediction"]
-
-        if correcting_x0_fn == "dynamic_thresholding":
-            self.correcting_x0_fn = self.dynamic_thresholding_fn
-        else:
-            self.correcting_x0_fn = correcting_x0_fn
-
-        self.correcting_xt_fn = correcting_xt_fn
-        self.dynamic_thresholding_ratio = dynamic_thresholding_ratio
-        self.thresholding_max_val = thresholding_max_val
-
-        self.predict_x0 = algorithm_type == "data_prediction"
-
-        self.sigma_min = float(self.noise_schedule.edm_sigma(torch.tensor([1e-3])))
-        self.sigma_max = float(self.noise_schedule.edm_sigma(torch.tensor([1])))
-
-    def dynamic_thresholding_fn(self, x0, t=None):
-        """
-        The dynamic thresholding method.
-        """
-        dims = x0.dim()
-        p = self.dynamic_thresholding_ratio
-        s = torch.quantile(torch.abs(x0).reshape((x0.shape[0], -1)), p, dim=1)
-        s = expand_dims(
-            torch.maximum(
-                s, self.thresholding_max_val * torch.ones_like(s).to(s.device)
-            ),
-            dims,
-        )
-        x0 = torch.clamp(x0, -s, s) / s
-        return x0
-
-    def noise_prediction_fn(self, x, t):
-        """
-        Return the noise prediction model.
-        """
-        return self.model(x, t)
-
-    def data_prediction_fn(self, x, t):
-        """
-        Return the data prediction model (with corrector).
-        """
-        noise = self.noise_prediction_fn(x, t)
-        alpha_t, sigma_t = (
-            self.noise_schedule.marginal_alpha(t),
-            self.noise_schedule.marginal_std(t),
-        )
-        x0 = (x - sigma_t * noise) / alpha_t
-        if self.correcting_x0_fn is not None:
-            x0 = self.correcting_x0_fn(x0)
-        return x0
-
-    def model_fn(self, x, t):
-        """
-        Convert the model to the noise prediction model or the data prediction model.
-        """
-
-        if self.predict_x0:
-            return self.data_prediction_fn(x, t)
-        else:
-            return self.noise_prediction_fn(x, t)
-
-    def get_time_steps(self, skip_type, t_T, t_0, N, order, device):
-        """Compute the intermediate time steps for sampling."""
-        if skip_type == "logSNR":
-            lambda_T = self.noise_schedule.marginal_lambda(torch.tensor(t_T).to(device))
-            lambda_0 = self.noise_schedule.marginal_lambda(torch.tensor(t_0).to(device))
-            logSNR_steps = lambda_T + torch.linspace(
-                torch.tensor(0.0).cpu().item(),
-                (lambda_0 - lambda_T).cpu().item() ** (1.0 / order),
-                N + 1,
-            ).pow(order).to(device)
-            return self.noise_schedule.inverse_lambda(logSNR_steps)
-        elif skip_type == "time":
-            t = (
-                torch.linspace(t_T ** (1.0 / order), t_0 ** (1.0 / order), N + 1)
-                .pow(order)
-                .to(device)
-            )
-            return t
-        elif skip_type == "karras":
-            sigma_min = max(0.002, self.sigma_min)
-            sigma_max = min(80, self.sigma_max)
-            sigma_steps = (
-                torch.linspace(sigma_max ** (1.0 / 7), sigma_min ** (1.0 / 7), N + 1)
-                .pow(7)
-                .to(device)
-            )
-            t = self.noise_schedule.edm_inverse_sigma(sigma_steps)
-            return t
-        else:
-            raise ValueError(
-                f"Unsupported skip_type {skip_type}, need to be 'logSNR' or 'time' or 'karras'"
-            )
-
-    def denoise_to_zero_fn(self, x, s):
-        """
-        Denoise at the final step, which is equivalent to solve the ODE from lambda_s to infty by first-order discretization.
-        """
-        return self.data_prediction_fn(x, s)
-
-    def get_coefficients_exponential_negative(
-        self, order, interval_start, interval_end
-    ):
-        """
-        Calculate the integral of exp(-x) * x^order dx from interval_start to interval_end
-        For calculating the coefficient of gradient terms after the lagrange interpolation,
-        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        For noise_prediction formula.
-        """
-        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
-
-        if order == 0:
-            return torch.exp(-interval_end) * (
-                torch.exp(interval_end - interval_start) - 1
-            )
-        elif order == 1:
-            return torch.exp(-interval_end) * (
-                (interval_start + 1) * torch.exp(interval_end - interval_start)
-                - (interval_end + 1)
-            )
-        elif order == 2:
-            return torch.exp(-interval_end) * (
-                (interval_start**2 + 2 * interval_start + 2)
-                * torch.exp(interval_end - interval_start)
-                - (interval_end**2 + 2 * interval_end + 2)
-            )
-        elif order == 3:
-            return torch.exp(-interval_end) * (
-                (interval_start**3 + 3 * interval_start**2 + 6 * interval_start + 6)
-                * torch.exp(interval_end - interval_start)
-                - (interval_end**3 + 3 * interval_end**2 + 6 * interval_end + 6)
-            )
-
-    def get_coefficients_exponential_positive(
-        self, order, interval_start, interval_end, tau
-    ):
-        """
-        Calculate the integral of exp(x(1+tau^2)) * x^order dx from interval_start to interval_end
-        For calculating the coefficient of gradient terms after the lagrange interpolation,
-        see Eq.(15) and Eq.(18) in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        For data_prediction formula.
-        """
-        assert order in [0, 1, 2, 3], "order is only supported for 0, 1, 2 and 3"
-
-        # after change of variable(cov)
-        interval_end_cov = (1 + tau**2) * interval_end
-        interval_start_cov = (1 + tau**2) * interval_start
-
-        if order == 0:
-            return (
-                torch.exp(interval_end_cov)
-                * (1 - torch.exp(-(interval_end_cov - interval_start_cov)))
-                / (1 + tau**2)
-            )
-        elif order == 1:
-            return (
-                torch.exp(interval_end_cov)
-                * (
-                    (interval_end_cov - 1)
-                    - (interval_start_cov - 1)
-                    * torch.exp(-(interval_end_cov - interval_start_cov))
-                )
-                / ((1 + tau**2) ** 2)
-            )
-        elif order == 2:
-            return (
-                torch.exp(interval_end_cov)
-                * (
-                    (interval_end_cov**2 - 2 * interval_end_cov + 2)
-                    - (interval_start_cov**2 - 2 * interval_start_cov + 2)
-                    * torch.exp(-(interval_end_cov - interval_start_cov))
-                )
-                / ((1 + tau**2) ** 3)
-            )
-        elif order == 3:
-            return (
-                torch.exp(interval_end_cov)
-                * (
-                    (
-                        interval_end_cov**3
-                        - 3 * interval_end_cov**2
-                        + 6 * interval_end_cov
-                        - 6
-                    )
-                    - (
-                        interval_start_cov**3
-                        - 3 * interval_start_cov**2
-                        + 6 * interval_start_cov
-                        - 6
-                    )
-                    * torch.exp(-(interval_end_cov - interval_start_cov))
-                )
-                / ((1 + tau**2) ** 4)
-            )
-
-    def lagrange_polynomial_coefficient(self, order, lambda_list):
-        """
-        Calculate the coefficient of lagrange polynomial
-        For lagrange interpolation
-        """
-        assert order in [0, 1, 2, 3]
-        assert order == len(lambda_list) - 1
-        if order == 0:
-            return [[1]]
-        elif order == 1:
-            return [
-                [
-                    1 / (lambda_list[0] - lambda_list[1]),
-                    -lambda_list[1] / (lambda_list[0] - lambda_list[1]),
-                ],
-                [
-                    1 / (lambda_list[1] - lambda_list[0]),
-                    -lambda_list[0] / (lambda_list[1] - lambda_list[0]),
-                ],
-            ]
-        elif order == 2:
-            denominator1 = (lambda_list[0] - lambda_list[1]) * (
-                lambda_list[0] - lambda_list[2]
-            )
-            denominator2 = (lambda_list[1] - lambda_list[0]) * (
-                lambda_list[1] - lambda_list[2]
-            )
-            denominator3 = (lambda_list[2] - lambda_list[0]) * (
-                lambda_list[2] - lambda_list[1]
-            )
-            return [
-                [
-                    1 / denominator1,
-                    (-lambda_list[1] - lambda_list[2]) / denominator1,
-                    lambda_list[1] * lambda_list[2] / denominator1,
-                ],
-                [
-                    1 / denominator2,
-                    (-lambda_list[0] - lambda_list[2]) / denominator2,
-                    lambda_list[0] * lambda_list[2] / denominator2,
-                ],
-                [
-                    1 / denominator3,
-                    (-lambda_list[0] - lambda_list[1]) / denominator3,
-                    lambda_list[0] * lambda_list[1] / denominator3,
-                ],
-            ]
-        elif order == 3:
-            denominator1 = (
-                (lambda_list[0] - lambda_list[1])
-                * (lambda_list[0] - lambda_list[2])
-                * (lambda_list[0] - lambda_list[3])
-            )
-            denominator2 = (
-                (lambda_list[1] - lambda_list[0])
-                * (lambda_list[1] - lambda_list[2])
-                * (lambda_list[1] - lambda_list[3])
-            )
-            denominator3 = (
-                (lambda_list[2] - lambda_list[0])
-                * (lambda_list[2] - lambda_list[1])
-                * (lambda_list[2] - lambda_list[3])
-            )
-            denominator4 = (
-                (lambda_list[3] - lambda_list[0])
-                * (lambda_list[3] - lambda_list[1])
-                * (lambda_list[3] - lambda_list[2])
-            )
-            return [
-                [
-                    1 / denominator1,
-                    (-lambda_list[1] - lambda_list[2] - lambda_list[3]) / denominator1,
-                    (
-                        lambda_list[1] * lambda_list[2]
-                        + lambda_list[1] * lambda_list[3]
-                        + lambda_list[2] * lambda_list[3]
-                    )
-                    / denominator1,
-                    (-lambda_list[1] * lambda_list[2] * lambda_list[3]) / denominator1,
-                ],
-                [
-                    1 / denominator2,
-                    (-lambda_list[0] - lambda_list[2] - lambda_list[3]) / denominator2,
-                    (
-                        lambda_list[0] * lambda_list[2]
-                        + lambda_list[0] * lambda_list[3]
-                        + lambda_list[2] * lambda_list[3]
-                    )
-                    / denominator2,
-                    (-lambda_list[0] * lambda_list[2] * lambda_list[3]) / denominator2,
-                ],
-                [
-                    1 / denominator3,
-                    (-lambda_list[0] - lambda_list[1] - lambda_list[3]) / denominator3,
-                    (
-                        lambda_list[0] * lambda_list[1]
-                        + lambda_list[0] * lambda_list[3]
-                        + lambda_list[1] * lambda_list[3]
-                    )
-                    / denominator3,
-                    (-lambda_list[0] * lambda_list[1] * lambda_list[3]) / denominator3,
-                ],
-                [
-                    1 / denominator4,
-                    (-lambda_list[0] - lambda_list[1] - lambda_list[2]) / denominator4,
-                    (
-                        lambda_list[0] * lambda_list[1]
-                        + lambda_list[0] * lambda_list[2]
-                        + lambda_list[1] * lambda_list[2]
-                    )
-                    / denominator4,
-                    (-lambda_list[0] * lambda_list[1] * lambda_list[2]) / denominator4,
-                ],
-            ]
-
-    def get_coefficients_fn(
-        self, order, interval_start, interval_end, lambda_list, tau
-    ):
-        """
-        Calculate the coefficient of gradients.
-        """
-        assert order in [1, 2, 3, 4]
-        assert order == len(lambda_list), (
-            "the length of lambda list must be equal to the order"
-        )
-        coefficients = []
-        lagrange_coefficient = self.lagrange_polynomial_coefficient(
-            order - 1, lambda_list
-        )
-        for i in range(order):
-            coefficient = 0
-            for j in range(order):
-                if self.predict_x0:
-                    coefficient += lagrange_coefficient[i][
-                        j
-                    ] * self.get_coefficients_exponential_positive(
-                        order - 1 - j, interval_start, interval_end, tau
-                    )
-                else:
-                    coefficient += lagrange_coefficient[i][
-                        j
-                    ] * self.get_coefficients_exponential_negative(
-                        order - 1 - j, interval_start, interval_end
-                    )
-            coefficients.append(coefficient)
-        assert len(coefficients) == order, (
-            "the length of coefficients does not match the order"
-        )
-        return coefficients
-
-    def adams_bashforth_update(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Predictor, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def adams_moulton_update(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        t_list = t_prev_list + [t]
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def adams_bashforth_update_few_steps(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Predictor, with the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_prev_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        if self.predict_x0:
-            if (
-                order == 2
-            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to unipc. Note: This is used only for few steps sampling.
-                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
-                # ODE case
-                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
-                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h ** 2 / 2 - (h - 1 + torch.exp(-h))) / (ns.marginal_lambda(t_prev_list[-1]) - ns.marginal_lambda(t_prev_list[-2]))
-                gradient_coefficients[0] += (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h**2 / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2)
-                    )
-                    / (
-                        ns.marginal_lambda(t_prev_list[-1])
-                        - ns.marginal_lambda(t_prev_list[-2])
-                    )
-                )
-                gradient_coefficients[1] -= (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h**2 / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2)
-                    )
-                    / (
-                        ns.marginal_lambda(t_prev_list[-1])
-                        - ns.marginal_lambda(t_prev_list[-2])
-                    )
-                )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def adams_moulton_update_few_steps(
-        self, order, x, tau, model_prev_list, t_prev_list, noise, t
-    ):
-        """
-        SA-Corrector, without the "rescaling" trick in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        """
-
-        assert order in [
-            1,
-            2,
-            3,
-            4,
-        ], (
-            "order of stochastic adams bashforth method is only supported for 1, 2, 3 and 4"
-        )
-
-        # get noise schedule
-        ns = self.noise_schedule
-        alpha_t = ns.marginal_alpha(t)
-        sigma_t = ns.marginal_std(t)
-        lambda_t = ns.marginal_lambda(t)
-        alpha_prev = ns.marginal_alpha(t_prev_list[-1])
-        sigma_prev = ns.marginal_std(t_prev_list[-1])
-        gradient_part = torch.zeros_like(x)
-        h = lambda_t - ns.marginal_lambda(t_prev_list[-1])
-        lambda_list = []
-        t_list = t_prev_list + [t]
-        for i in range(order):
-            lambda_list.append(ns.marginal_lambda(t_list[-(i + 1)]))
-        gradient_coefficients = self.get_coefficients_fn(
-            order, ns.marginal_lambda(t_prev_list[-1]), lambda_t, lambda_list, tau
-        )
-
-        if self.predict_x0:
-            if (
-                order == 2
-            ):  ## if order = 2 we do a modification that does not influence the convergence order similar to UniPC. Note: This is used only for few steps sampling.
-                # The added term is O(h^3). Empirically we find it will slightly improve the image quality.
-                # ODE case
-                # gradient_coefficients[0] += 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
-                # gradient_coefficients[1] -= 1.0 * torch.exp(lambda_t) * (h / 2 - (h - 1 + torch.exp(-h)) / h)
-                gradient_coefficients[0] += (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2 * h)
-                    )
-                )
-                gradient_coefficients[1] -= (
-                    1.0
-                    * torch.exp((1 + tau**2) * lambda_t)
-                    * (
-                        h / 2
-                        - (h * (1 + tau**2) - 1 + torch.exp((1 + tau**2) * (-h)))
-                        / ((1 + tau**2) ** 2 * h)
-                    )
-                )
-
-        for i in range(order):
-            if self.predict_x0:
-                gradient_part += (
-                    (1 + tau**2)
-                    * sigma_t
-                    * torch.exp(-(tau**2) * lambda_t)
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-            else:
-                gradient_part += (
-                    -(1 + tau**2)
-                    * alpha_t
-                    * gradient_coefficients[i]
-                    * model_prev_list[-(i + 1)]
-                )
-
-        if self.predict_x0:
-            noise_part = sigma_t * torch.sqrt(1 - torch.exp(-2 * tau**2 * h)) * noise
-        else:
-            noise_part = tau * sigma_t * torch.sqrt(torch.exp(2 * h) - 1) * noise
-
-        if self.predict_x0:
-            x_t = (
-                torch.exp(-(tau**2) * h) * (sigma_t / sigma_prev) * x
-                + gradient_part
-                + noise_part
-            )
-        else:
-            x_t = (alpha_t / alpha_prev) * x + gradient_part + noise_part
-
-        return x_t
-
-    def sample_few_steps(
-        self,
-        x,
-        tau,
-        steps=5,
-        t_start=None,
-        t_end=None,
-        skip_type="time",
-        skip_order=1,
-        predictor_order=3,
-        corrector_order=4,
-        pc_mode="PEC",
-        return_intermediate=False,
-    ):
-        """
-        For the PC-mode, please refer to the wiki page
-        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
-        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
-        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
-        """
-
-        skip_first_step = False
-        skip_final_step = True
-        lower_order_final = True
-        denoise_to_zero = False
-
-        assert pc_mode in [
-            "PEC",
-            "PECE",
-        ], "Predictor-corrector mode only supports PEC and PECE"
-        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
-        t_T = self.noise_schedule.T if t_start is None else t_start
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-
-        device = x.device
-        intermediates = []
-        with torch.no_grad():
-            assert steps >= max(predictor_order, corrector_order - 1)
-            timesteps = self.get_time_steps(
-                skip_type=skip_type,
-                t_T=t_T,
-                t_0=t_0,
-                N=steps,
-                order=skip_order,
-                device=device,
-            )
-            assert timesteps.shape[0] - 1 == steps
-            # Init the initial values.
-            step = 0
-            t = timesteps[step]
-            noise = torch.randn_like(x)
-            t_prev_list = [t]
-            # do not evaluate if skip_first_step
-            if skip_first_step:
-                if self.predict_x0:
-                    alpha_t = self.noise_schedule.marginal_alpha(t)
-                    sigma_t = self.noise_schedule.marginal_std(t)
-                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
-                else:
-                    model_prev_list = [x]
-            else:
-                model_prev_list = [self.model_fn(x, t)]
-
-            if self.correcting_xt_fn is not None:
-                x = self.correcting_xt_fn(x, t, step)
-            if return_intermediate:
-                intermediates.append(x)
-
-            # determine the first several values
-            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
-                t = timesteps[step]
-                predictor_order_used = min(predictor_order, step)
-                corrector_order_used = min(corrector_order, step + 1)
-                noise = torch.randn_like(x)
-                # predictor step
-                x_p = self.adams_bashforth_update_few_steps(
-                    order=predictor_order_used,
-                    x=x,
-                    tau=tau(t),
-                    model_prev_list=model_prev_list,
-                    t_prev_list=t_prev_list,
-                    noise=noise,
-                    t=t,
-                )
-                # evaluation step
-                model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                model_prev_list.append(model_x)
-                # corrector step
-                if corrector_order > 0:
-                    x = self.adams_moulton_update_few_steps(
-                        order=corrector_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x = x_p
-
-                # evaluation step if correction and mode = pece
-                if corrector_order > 0:
-                    if pc_mode == "PECE":
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-
-            for step in tqdm(
-                range(max(predictor_order, corrector_order - 1), steps + 1)
-            ):
-                if lower_order_final:
-                    predictor_order_used = min(predictor_order, steps - step + 1)
-                    corrector_order_used = min(corrector_order, steps - step + 2)
-
-                else:
-                    predictor_order_used = predictor_order
-                    corrector_order_used = corrector_order
-                t = timesteps[step]
-                noise = torch.randn_like(x)
-
-                # predictor step
-                if skip_final_step and step == steps and not denoise_to_zero:
-                    x_p = self.adams_bashforth_update_few_steps(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=0,
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x_p = self.adams_bashforth_update_few_steps(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-
-                # evaluation step
-                # do not evaluate if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                # do not update if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_prev_list.append(model_x)
-
-                # corrector step
-                # do not correct if skip_final_step and step = steps
-                if corrector_order > 0:
-                    if not skip_final_step or step < steps:
-                        x = self.adams_moulton_update_few_steps(
-                            order=corrector_order_used,
-                            x=x,
-                            tau=tau(t),
-                            model_prev_list=model_prev_list,
-                            t_prev_list=t_prev_list,
-                            noise=noise,
-                            t=t,
-                        )
-                    else:
-                        x = x_p
-                else:
-                    x = x_p
-
-                # evaluation step if mode = pece and step != steps
-                if corrector_order > 0:
-                    if pc_mode == "PECE" and step < steps:
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-                del model_prev_list[0]
-
-            if denoise_to_zero:
-                t = torch.ones((1,)).to(device) * t_0
-                x = self.denoise_to_zero_fn(x, t)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step + 1)
-                if return_intermediate:
-                    intermediates.append(x)
-        if return_intermediate:
-            return x, intermediates
-        else:
-            return x
-
-    def sample_more_steps(
-        self,
-        x,
-        tau,
-        steps=20,
-        t_start=None,
-        t_end=None,
-        skip_type="time",
-        skip_order=1,
-        predictor_order=3,
-        corrector_order=4,
-        pc_mode="PEC",
-        return_intermediate=False,
-    ):
-        """
-        For the PC-mode, please refer to the wiki page
-        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
-        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
-        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
-        """
-
-        skip_first_step = False
-        skip_final_step = False
-        lower_order_final = True
-        denoise_to_zero = True
-
-        assert pc_mode in [
-            "PEC",
-            "PECE",
-        ], "Predictor-corrector mode only supports PEC and PECE"
-        t_0 = 1.0 / self.noise_schedule.total_N if t_end is None else t_end
-        t_T = self.noise_schedule.T if t_start is None else t_start
-        assert t_0 > 0 and t_T > 0, (
-            "Time range needs to be greater than 0. For discrete-time DPMs, it needs to be in [1 / N, 1], where N is the length of betas array"
-        )
-
-        device = x.device
-        intermediates = []
-        with torch.no_grad():
-            assert steps >= max(predictor_order, corrector_order - 1)
-            timesteps = self.get_time_steps(
-                skip_type=skip_type,
-                t_T=t_T,
-                t_0=t_0,
-                N=steps,
-                order=skip_order,
-                device=device,
-            )
-            assert timesteps.shape[0] - 1 == steps
-            # Init the initial values.
-            step = 0
-            t = timesteps[step]
-            noise = torch.randn_like(x)
-            t_prev_list = [t]
-            # do not evaluate if skip_first_step
-            if skip_first_step:
-                if self.predict_x0:
-                    alpha_t = self.noise_schedule.marginal_alpha(t)
-                    sigma_t = self.noise_schedule.marginal_std(t)
-                    model_prev_list = [(1 - sigma_t) / alpha_t * x]
-                else:
-                    model_prev_list = [x]
-            else:
-                model_prev_list = [self.model_fn(x, t)]
-
-            if self.correcting_xt_fn is not None:
-                x = self.correcting_xt_fn(x, t, step)
-            if return_intermediate:
-                intermediates.append(x)
-
-            # determine the first several values
-            for step in tqdm(range(1, max(predictor_order, corrector_order - 1))):
-                t = timesteps[step]
-                predictor_order_used = min(predictor_order, step)
-                corrector_order_used = min(corrector_order, step + 1)
-                noise = torch.randn_like(x)
-                # predictor step
-                x_p = self.adams_bashforth_update(
-                    order=predictor_order_used,
-                    x=x,
-                    tau=tau(t),
-                    model_prev_list=model_prev_list,
-                    t_prev_list=t_prev_list,
-                    noise=noise,
-                    t=t,
-                )
-                # evaluation step
-                model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                model_prev_list.append(model_x)
-                # corrector step
-                if corrector_order > 0:
-                    x = self.adams_moulton_update(
-                        order=corrector_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x = x_p
-
-                # evaluation step if mode = pece
-                if corrector_order > 0:
-                    if pc_mode == "PECE":
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-
-            for step in tqdm(
-                range(max(predictor_order, corrector_order - 1), steps + 1)
-            ):
-                if lower_order_final:
-                    predictor_order_used = min(predictor_order, steps - step + 1)
-                    corrector_order_used = min(corrector_order, steps - step + 2)
-
-                else:
-                    predictor_order_used = predictor_order
-                    corrector_order_used = corrector_order
-                t = timesteps[step]
-                noise = torch.randn_like(x)
-
-                # predictor step
-                if skip_final_step and step == steps and not denoise_to_zero:
-                    x_p = self.adams_bashforth_update(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=0,
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-                else:
-                    x_p = self.adams_bashforth_update(
-                        order=predictor_order_used,
-                        x=x,
-                        tau=tau(t),
-                        model_prev_list=model_prev_list,
-                        t_prev_list=t_prev_list,
-                        noise=noise,
-                        t=t,
-                    )
-
-                # evaluation step
-                # do not evaluate if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_x = self.model_fn(x_p, t)
-
-                # update model_list
-                # do not update if skip_final_step and step = steps
-                if not skip_final_step or step < steps:
-                    model_prev_list.append(model_x)
-
-                # corrector step
-                # do not correct if skip_final_step and step = steps
-                if corrector_order > 0:
-                    if not skip_final_step or step < steps:
-                        x = self.adams_moulton_update(
-                            order=corrector_order_used,
-                            x=x,
-                            tau=tau(t),
-                            model_prev_list=model_prev_list,
-                            t_prev_list=t_prev_list,
-                            noise=noise,
-                            t=t,
-                        )
-                    else:
-                        x = x_p
-                else:
-                    x = x_p
-
-                # evaluation step if mode = pece and step != steps
-                if corrector_order > 0:
-                    if pc_mode == "PECE" and step < steps:
-                        model_x = self.model_fn(x, t)
-                        del model_prev_list[-1]
-                        model_prev_list.append(model_x)
-
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step)
-                if return_intermediate:
-                    intermediates.append(x)
-
-                t_prev_list.append(t)
-                del model_prev_list[0]
-
-            if denoise_to_zero:
-                t = torch.ones((1,)).to(device) * t_0
-                x = self.denoise_to_zero_fn(x, t)
-                if self.correcting_xt_fn is not None:
-                    x = self.correcting_xt_fn(x, t, step + 1)
-                if return_intermediate:
-                    intermediates.append(x)
-        if return_intermediate:
-            return x, intermediates
-        else:
-            return x
-
-    def sample(
-        self,
-        mode,
-        x,
-        tau,
-        steps,
-        t_start=None,
-        t_end=None,
-        skip_type="time",
-        skip_order=1,
-        predictor_order=3,
-        corrector_order=4,
-        pc_mode="PEC",
-        return_intermediate=False,
-    ):
-        """
-        For the PC-mode, please refer to the wiki page
-        https://en.wikipedia.org/wiki/Predictor%E2%80%93corrector_method#PEC_mode_and_PECE_mode
-        'PEC' needs one model evaluation per step while 'PECE' needs two model evaluations
-        We recommend use pc_mode='PEC' for NFEs is limited. 'PECE' mode is only for test with sufficient NFEs.
-
-        'few_steps' mode is recommended. The differences between 'few_steps' and 'more_steps' are as below:
-        1) 'few_steps' do not correct at final step and do not denoise to zero, while 'more_steps' do these two.
-        Thus the NFEs for 'few_steps' = steps, NFEs for 'more_steps' = steps + 2
-        For most of the experiments and tasks, we find these two operations do not have much help to sample quality.
-        2) 'few_steps' use a rescaling trick as in Appendix D in SA-Solver paper https://arxiv.org/pdf/2309.05019.pdf
-        We find it will slightly improve the sample quality especially in few steps.
-        """
-        assert mode in [
-            "few_steps",
-            "more_steps",
-        ], "mode must be either 'few_steps' or 'more_steps'"
-        if mode == "few_steps":
-            return self.sample_few_steps(
-                x=x,
-                tau=tau,
-                steps=steps,
-                t_start=t_start,
-                t_end=t_end,
-                skip_type=skip_type,
-                skip_order=skip_order,
-                predictor_order=predictor_order,
-                corrector_order=corrector_order,
-                pc_mode=pc_mode,
-                return_intermediate=return_intermediate,
-            )
-        else:
-            return self.sample_more_steps(
-                x=x,
-                tau=tau,
-                steps=steps,
-                t_start=t_start,
-                t_end=t_end,
-                skip_type=skip_type,
-                skip_order=skip_order,
-                predictor_order=predictor_order,
-                corrector_order=corrector_order,
-                pc_mode=pc_mode,
-                return_intermediate=return_intermediate,
-            )
-
-
-#############################################################
-# other utility functions
-#############################################################
-
-
-def interpolate_fn(x, xp, yp):
-    """
-    A piecewise linear function y = f(x), using xp and yp as keypoints.
-    We implement f(x) in a differentiable way (i.e. applicable for autograd).
-    The function f(x) is well-defined for all x-axis. (For x beyond the bounds of xp, we use the outmost points of xp to define the linear function.)
-    Args:
-        x: PyTorch tensor with shape [N, C], where N is the batch size, C is the number of channels (we use C = 1 for DPM-Solver).
-        xp: PyTorch tensor with shape [C, K], where K is the number of keypoints.
-        yp: PyTorch tensor with shape [C, K].
-    Returns:
-        The function values f(x), with shape [N, C].
-    """
-    N, K = x.shape[0], xp.shape[1]
-    all_x = torch.cat([x.unsqueeze(2), xp.unsqueeze(0).repeat((N, 1, 1))], dim=2)
-    sorted_all_x, x_indices = torch.sort(all_x, dim=2)
-    x_idx = torch.argmin(x_indices, dim=2)
-    cand_start_idx = x_idx - 1
-    start_idx = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(1, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    end_idx = torch.where(
-        torch.eq(start_idx, cand_start_idx), start_idx + 2, start_idx + 1
-    )
-    start_x = torch.gather(sorted_all_x, dim=2, index=start_idx.unsqueeze(2)).squeeze(2)
-    end_x = torch.gather(sorted_all_x, dim=2, index=end_idx.unsqueeze(2)).squeeze(2)
-    start_idx2 = torch.where(
-        torch.eq(x_idx, 0),
-        torch.tensor(0, device=x.device),
-        torch.where(
-            torch.eq(x_idx, K),
-            torch.tensor(K - 2, device=x.device),
-            cand_start_idx,
-        ),
-    )
-    y_positions_expanded = yp.unsqueeze(0).expand(N, -1, -1)
-    start_y = torch.gather(
-        y_positions_expanded, dim=2, index=start_idx2.unsqueeze(2)
-    ).squeeze(2)
-    end_y = torch.gather(
-        y_positions_expanded, dim=2, index=(start_idx2 + 1).unsqueeze(2)
-    ).squeeze(2)
-    cand = start_y + (x - start_x) * (end_y - start_y) / (end_x - start_x)
-    return cand
-
-
-def expand_dims(v, dims):
-    """
-    Expand the tensor `v` to the dim `dims`.
-    Args:
-        `v`: a PyTorch tensor with shape [N].
-        `dim`: a `int`.
-    Returns:
-        a PyTorch tensor with shape [N, 1, 1, ..., 1] and the total dimension is `dims`.
-    """
-    return v[(...,) + (None,) * (dims - 1)]
diff --git a/seed/seed_llm/config.yaml b/seed/seed_llm/config.yaml
deleted file mode 100644
index ef7027cb4..000000000
--- a/seed/seed_llm/config.yaml
+++ /dev/null
@@ -1,37 +0,0 @@
-base_image:
-  image: public.ecr.aws/q9t5s3a7/vllm-ci-postmerge-repo:78336a0c3ee4eb9dba6e37959d926160e91623fd
-build_commands:
-  - pip install --pre --upgrade transformers
-  - pip uninstall -y vllm
-  - VLLM_USE_PRECOMPILED=1 VLLM_TEST_USE_PRECOMPILED_NIGHTLY_WHEEL=1 pip install git+https://github.com/vllm-project/vllm.git@78336a0c3ee4eb9dba6e37959d926160e91623fd
-model_metadata:
-  repo_id: ByteDance-Seed/Seed-OSS-36B-Instruct
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write FizzBuzz in Python"
-    stream: true
-    model: "ByteDance-Seed/Seed-OSS-36B-Instruct"
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-docker_server:
-  start_command: python3 -m vllm.entrypoints.openai.api_server --model  ByteDance-Seed/Seed-OSS-36B-Instruct -O3 --tensor-parallel-size 2 --tool-call-parser seed_oss --served-model-name ByteDance-Seed/Seed-OSS-36B-Instruct --enable-auto-tool-choice --max-model-len 65536 --gpu-memory-utilization=0.95 --host 0.0.0.0 --port 8000
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-model_cache:
-  - repo_id: ByteDance-Seed/Seed-OSS-36B-Instruct
-    revision: 497f1dca95ebdec98e41d517b9f060ee753c902f
-    use_volume: true
-    volume_folder: glm
-model_name: Seed-OSS-36B-Instruct
diff --git a/segment-anything/README.md b/segment-anything/README.md
deleted file mode 100644
index 245d448db..000000000
--- a/segment-anything/README.md
+++ /dev/null
@@ -1,17 +0,0 @@
-# Segment Anything Model
-
-This is an example deploying Segment Anything Model (SAM) with truss weights preloaded
-
-## Deploy to Baseten
-To deploy the model, run the following from the root of the directory
-
-```
-truss push --publish
-```
-
-## Predict
-Example prediction:
-
-```
-truss predict --published -d '{"image_url": "https://as2.ftcdn.net/v2/jpg/00/66/26/87/1000_F_66268784_jccdcfdpf2vmq5X8raYA8JQT0sziZ1H9.jpg"}'
-```
diff --git a/segment-anything/config.yaml b/segment-anything/config.yaml
deleted file mode 100644
index 05f1991f1..000000000
--- a/segment-anything/config.yaml
+++ /dev/null
@@ -1,24 +0,0 @@
-environment_variables: {}
-external_data:
-- local_data_path: sam_vit_h_4b8939.pth
-  url: https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    image_url: https://as2.ftcdn.net/v2/jpg/00/66/26/87/1000_F_66268784_jccdcfdpf2vmq5X8raYA8JQT0sziZ1H9.jpg
-model_name: Segment Anything
-python_version: py310
-requirements:
-- git+https://github.com/facebookresearch/segment-anything.git@6fdee8f2727f4506cfbbe553e23b895e27956588
-- opencv-python==4.8.1.78
-- torch==2.1.0
-- torchvision==0.16.0
-- pycocotools==2.0.7
-resources:
-  accelerator: A10G
-  cpu: 1000m
-  memory: 10Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- python3-opencv
diff --git a/sesame-csm-1b/config.yaml b/sesame-csm-1b/config.yaml
deleted file mode 100644
index be19e9d10..000000000
--- a/sesame-csm-1b/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-model_name: sesame-csm-1b
-python_version: py310
-model_metadata:
-  example_model_input:
-    text: "Hello from Sesame."
-    speaker: 0
-requirements:
-  - torch==2.4.0
-  - torchaudio==2.4.0
-  - tokenizers==0.21.0
-  - transformers==4.49.0
-  - huggingface_hub==0.28.1
-  - moshi==0.2.2
-  - torchtune==0.4.0
-  - torchao==0.9.0
-  - silentcipher @ git+https://github.com/SesameAILabs/silentcipher@master
-  - ffmpeg
-  - git+https://github.com/veerbia/csm.git
-resources:
-  accelerator: T4
-  cpu: '1'
-  memory: 10Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages: []
-environment_variables: {}
-external_package_dirs: []
diff --git a/stable-diffusion/dreamshaper-lcm/README.md b/stable-diffusion/dreamshaper-lcm/README.md
deleted file mode 100644
index d43c92122..000000000
--- a/stable-diffusion/dreamshaper-lcm/README.md
+++ /dev/null
@@ -1,3 +0,0 @@
-# Dreamshaper Latent Consistency Model
-
-A Truss for [Dreamshaper LCM](https://huggingface.co/spaces/SimianLuo/Latent_Consistency_Model), a distillation of Dreamshaper (a Stable Diffusion 1.5 fine-tune), that can achieve similar quality in ~1-8 steps. Generate high quality 768 x 768 images in under a second.
diff --git a/stable-diffusion/dreamshaper-lcm/config.yaml b/stable-diffusion/dreamshaper-lcm/config.yaml
deleted file mode 100644
index fd170844b..000000000
--- a/stable-diffusion/dreamshaper-lcm/config.yaml
+++ /dev/null
@@ -1,16 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: Dreamshaper Latent Consistency Model
-python_version: py311
-requirements:
-- diffusers=0.21.4
-- transformers=4.34.1
-- accelerate=0.23.0
-- torch=2.1.0
-resources:
-  accelerator: A10G
-  cpu: '1'
-  memory: 2Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/stable-diffusion/playground-v2-trt/README.md b/stable-diffusion/playground-v2-trt/README.md
deleted file mode 100644
index b447d5d6f..000000000
--- a/stable-diffusion/playground-v2-trt/README.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# Playground v2 TensorRT Truss
-
-Playground is a diffusion-based text-to-image generative model. This README covers deploying and invoking this model.
-
-This model is packaged using [Truss](https://trussml.com), the simplest way to serve AI/ML models in production.
-
-## Deploy Playground v2 TensorRT
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion/playground-v2-trt
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `playground-v2-trt` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using SDXL through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-### Hardware notes
-
-Running inference on an A100 cuts invocation time to ~3.5 seconds.
-
-## Invoking Playground v2 TensorRT
-
-Playground v2 TensorRT returns an image in Base 64, which is not super useful as a string in your terminal. So we included a helpful utility script to show and save the image. Pipe the model results into the script.
-
-```sh
-truss predict -d '{"prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"}' | python show.py
-```
-
-The output will be a dictionary with a key `data` mapping to a base64 encoded image. It's processed with this script:
-
-```python
-import json
-import base64
-import os, sys
-
-resp = sys.stdin.read()
-image = json.loads(resp)["data"]
-img=base64.b64decode(image)
-
-file_name = f'{image[-10:].replace("/", "")}.jpeg'
-img_file = open(file_name, 'wb')
-img_file.write(img)
-img_file.close()
-os.system(f'open {file_name}')
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://app.baseten.co/models/{MODEL_ID}/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
-         }'
-```
-
-Again, the model will return a dictionary containing the base64-encoded image, which will need to be decoded and saved.
diff --git a/stable-diffusion/playground-v2-trt/config.yaml b/stable-diffusion/playground-v2-trt/config.yaml
deleted file mode 100644
index 380979254..000000000
--- a/stable-diffusion/playground-v2-trt/config.yaml
+++ /dev/null
@@ -1,55 +0,0 @@
-base_image:
-  image: nvcr.io/nvidia/pytorch:23.11-py3
-  python_executable_path: /usr/bin/python
-description: Generate original images from text prompts.
-environment_variables:
-  HF_HUB_ENABLE_HF_TRANSFER: 1
-external_package_dirs: []
-model_cache:
-- repo_id: baseten/playground-v2-trt-8.6.1.post1-engine-A100
-  use_volume: false
-- allow_patterns:
-  - config.json
-  - diffusion_pytorch_model.safetensors
-  repo_id: madebyollin/sdxl-vae-fp16-fix
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - playground-v2.safetensors
-  repo_id: playgroundai/playground-v2-1024px-aesthetic
-model_metadata:
-  example_model_input:
-    prompt: Astronaut in a jungle, cold color palette, muted colors, detailed, 8k
-  pretty_name: Playground v2 - TensorRT
-  tags:
-  - image-generation
-model_name: Playground v2 - TensorRT
-python_version: py39
-requirements:
-- accelerate
-- colored
-- cuda-python
-- ftfy
-- nvtx
-- opencv-python==4.8.0.74
-- scipy
-- transformers==4.31.0
-- safetensors
-- hf_transfer
-- diffusers==0.23.1
-- invisible-watermark>=0.2.0
-- --extra-index-url https://pypi.ngc.nvidia.com
-- polygraphy
-- --extra-index-url https://pypi.nvidia.com
-- tensorrt==8.6.1.post1
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  predict_concurrency: 1
-secrets: {}
-system_packages:
-- python3.10-venv
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/sd-textual-inversion/README.md b/stable-diffusion/sd-textual-inversion/README.md
deleted file mode 100644
index b2ce912c7..000000000
--- a/stable-diffusion/sd-textual-inversion/README.md
+++ /dev/null
@@ -1,8 +0,0 @@
-# Textual Inversion with Stable Diffusion
-
-The following example demonstrates how to use Stable Diffusion with
-textual inversion embeddings.
-
-This truss combines concepts from:
-1. [This colab](https://colab.research.google.com/github/huggingface/notebooks/blob/main/diffusers/stable_conceptualizer_inference.ipynb#scrollTo=JkIeuLEfqi-g) which demonstrates how to load textual inversion embeddings from hugginface repos
-2. [This diffusers issue](https://github.com/huggingface/diffusers/issues/3097#issuecomment-1516138396) which demonstrates how to load an embedding directly.
diff --git a/stable-diffusion/sd-textual-inversion/config.yaml b/stable-diffusion/sd-textual-inversion/config.yaml
deleted file mode 100644
index 2008f9d2f..000000000
--- a/stable-diffusion/sd-textual-inversion/config.yaml
+++ /dev/null
@@ -1,22 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  pretty_name: Stable Diffusion - Textual Inversion
-  tags:
-  - image-generation
-model_name: SD_Textual_Inversion
-python_version: py311
-requirements:
-- diffusers==0.16.1
-- transformers
-- ftfy
-- accelerate
-- torch
-- pillow
-resources:
-  accelerator: T4
-  cpu: 500m
-  memory: 512Mi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/stable-diffusion/sd-turbo/README.md b/stable-diffusion/sd-turbo/README.md
deleted file mode 100644
index ebbb912a6..000000000
--- a/stable-diffusion/sd-turbo/README.md
+++ /dev/null
@@ -1,79 +0,0 @@
-# SD Turbo
-
-SD Turbo is a fast text-to-image model built for real time image synthesis. It is a distilled version of Stable Diffusion 2.1.
-
-SD Turbo is a smaller model than [SDXL Turbo](https://github.com/basetenlabs/truss-examples/tree/main/stable-diffusion/sdxl-turbo). Comparatively, it achieves faster latency and tends to produce outputs of lower quality and prompt alignment.
-
-## Deploying SD Turbo
-
-First, clone this repository and navigate to the `sd-turbo` directory:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd sd-turbo
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `sd-turbo` as your working directory, deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-On a T4 with 16 GiB VRAM, we have observed that the SD Turbo mean inference time is < 0.5 seconds.
-
-## Invoking SD Turbo
-
-The SD Turbo truss takes in a JSON payload with two fields:
-- `prompt` (required): Text describing the desired image.
-- `num_steps` (optional, default = 1): Number of steps the model should iterate. Must be between 1-4.
-
-It returns a JSON object with a `result` field, which contains a generated image of size 512 x 512 encoded as a base 64 string.
-
-### Example usage
-
-The example code below invokes the SD Turbo truss, parses the image from the response, and saves the image.
-
-```python
-from PIL import Image
-from io import BytesIO
-import base64
-import requests
-
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-# Paste your model ID here. This can be grabbed from the Baseten UI or from the
-# output of `truss push`.
-MODEL_ID = ""
-# Development model endpoint URL. To call the production deployment or another
-# deployment, replace this with the desired endpoint URL from the Baseten UI.
-MODEL_ENDPOINT = f"https://model-{MODEL_ID}.api.baseten.co/development/predict"
-
-# Paste your Baseten API key here.
-API_KEY = ""
-HEADERS = {"Authorization": f"Api-Key {API_KEY}"}
-
-resp = requests.post(
-    MODEL_ENDPOINT,
-    headers=HEADERS,
-    json={"prompt": "A tree in a field under the night sky"},
-)
-resp = resp.json()
-img = b64_to_pil(resp.get("result"))
-
-# Save the image.
-img.save("sd_turbo_output.png")
-```
diff --git a/stable-diffusion/sd-turbo/config.yaml b/stable-diffusion/sd-turbo/config.yaml
deleted file mode 100644
index 63469a4a4..000000000
--- a/stable-diffusion/sd-turbo/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_cache:
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - '*.txt'
-  repo_id: stabilityai/sdxl-turbo
-  use_volume: false
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: A tree in a field under the night sky
-  pretty_name: SD Turbo
-  tags:
-  - image-generation
-model_name: SD Turbo
-python_version: py311
-requirements:
-- torch==2.0.1
-- transformers==4.35.2
-- diffusers==0.23.1
-- accelerate==0.24.1
-resources:
-  accelerator: T4
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/stable-diffusion/sdxl-controlnet-canny/README.md b/stable-diffusion/sdxl-controlnet-canny/README.md
deleted file mode 100644
index 83a720d54..000000000
--- a/stable-diffusion/sdxl-controlnet-canny/README.md
+++ /dev/null
@@ -1,91 +0,0 @@
-# Stable Diffusion XL + ControlNet Canny Truss
-
-This Truss uses Stable Diffusion XL and ControlNet with the Canny preprocessor to generate images guided by input image edges. The inputs are a prompt and an image. A Canny filter is applied to the image to generate a outline, which is then passed to SDXL with the prompt.
-
-![baseten_controlnet](baseten-logo.gif)
-
-## Deploying the Truss
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd sdxl-controlnet-canny-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `sdxl-controlnet-canny-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Using the model
-
-The model takes a JSON payload with two fields:
-
-- `prompt`: Text describing the desired image.
-- `image`: Base64 encoded input image.
-
-It returns a JSON object with the `result` field containing the generated image.
-
-## Example Usage
-
-You can also invoke the SDXL + ControlNet model from Python using the `baseten` SDK:
-
-```python
-import baseten
-from PIL import Image
-from io import BytesIO
-import base64
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-model = baseten.deployed_model_version_id("MODEL_VERSION_ID") # you can get this from the Baseten web UI
-
-image = Image.open("cat.png")
-image_b64 = pil_to_b64()
-
-request = {
-  "prompt": "A painting of a cat",
-  "image": image_b64
-}
-
-response = model.predict(request)
-```
-
-The response will contain a base64 encoded image that you can save:
-
-```python
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-
-img = b64_to_pil(response["result"])
-
-img.save('generated.png')
-```
-
-You can also invoke the model via REST API:
-
-```bash
-curl -X POST "https://app.baseten.co/model_versions/VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H "Authorization: Api-Key {API_KEY}" \
-     -d '{"prompt": "A painting of a cat",
-           "image": "data:image/png;base64,..."}'
-```
-
-The API will return a JSON response containing the generated image encoded in base64.
diff --git a/stable-diffusion/sdxl-controlnet-canny/config.yaml b/stable-diffusion/sdxl-controlnet-canny/config.yaml
deleted file mode 100644
index a6acda3f9..000000000
--- a/stable-diffusion/sdxl-controlnet-canny/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: aerial view, a futuristic research complex in a bright foggy jungle, hard
-      lighting
-  model_metadata: null
-  pretty_name: Stable Diffusion ControlNet
-  tags:
-  - image-generation
-model_name: SDXL ControlNet Canny
-python_version: py39
-requirements:
-- accelerate==0.23.0
-- transformers==4.33.2
-- safetensors==0.3.3
-- opencv-python==4.8.0.76
-- diffusers==0.21.2
-resources:
-  accelerator: A10G:2
-  cpu: 3500m
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/sdxl-controlnet-depth/README.md b/stable-diffusion/sdxl-controlnet-depth/README.md
deleted file mode 100644
index 3d66358b0..000000000
--- a/stable-diffusion/sdxl-controlnet-depth/README.md
+++ /dev/null
@@ -1,89 +0,0 @@
-# Stable Diffusion XL + ControlNet Depth Truss
-
-This Truss uses Stable Diffusion XL and ControlNet with the Depth preprocessor to generate images guided by the depth map of an input image. The inputs are a prompt and an image.
-
-## Deploying the Truss
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd sdxl-controlnet-depth-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `sdxl-controlnet-depth-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Using the model
-
-The model takes a JSON payload with two fields:
-
-- `prompt`: Text describing the desired image.
-- `image`: Base64 encoded input image.
-
-It returns a JSON object with the `result` field containing the generated image.
-
-## Example Usage
-
-You can also invoke the SDXL + ControlNet model from Python using the `baseten` SDK:
-
-```python
-import baseten
-from PIL import Image
-from io import BytesIO
-import base64
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-model = baseten.deployed_model_version_id("MODEL_VERSION_ID") # you can get this from the Baseten web UI
-
-image = Image.open("cat.png")
-image_b64 = pil_to_b64()
-
-request = {
-  "prompt": "A painting of a cat",
-  "image": image_b64
-}
-
-response = model.predict(request)
-```
-
-The response will contain a base64 encoded image that you can save:
-
-```python
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-
-img = b64_to_pil(response["result"])
-
-img.save('generated.png')
-```
-
-You can also invoke the model via REST API:
-
-```bash
-curl -X POST "https://app.baseten.co/model_versions/VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H "Authorization: Api-Key {API_KEY}" \
-     -d '{"prompt": "A painting of a cat",
-           "image": "data:image/png;base64,..."}'
-```
-
-The API will return a JSON response containing the generated image encoded in base64.
diff --git a/stable-diffusion/sdxl-controlnet-depth/config.yaml b/stable-diffusion/sdxl-controlnet-depth/config.yaml
deleted file mode 100644
index b0ca0d71b..000000000
--- a/stable-diffusion/sdxl-controlnet-depth/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: large bed, abstract painting on the wall, fluffy rug on the floor, ambient
-      lighting, extremely detailed
-  pretty_name: Stable Diffusion ControlNet Depth
-  tags:
-  - image-generation
-model_name: SDXL ControlNet Depth
-python_version: py39
-requirements:
-- accelerate==0.23.0
-- transformers==4.33.2
-- safetensors==0.3.3
-- opencv-python==4.8.0.76
-- diffusers==0.21.2
-resources:
-  accelerator: A10G:2
-  cpu: 3500m
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/sdxl-controlnet/README.md b/stable-diffusion/sdxl-controlnet/README.md
deleted file mode 100644
index e2b621fda..000000000
--- a/stable-diffusion/sdxl-controlnet/README.md
+++ /dev/null
@@ -1,81 +0,0 @@
-# Stable Diffusion XL + ControlNet Truss
-
-This Truss uses Stable Diffusion XL and ControlNet to generate images guided by input image edges. The inputs are a prompt and an image. A Canny filter is applied to the image to generate a outline, which is then passed to SDXL with the prompt.
-
-![baseten_controlnet](baseten-logo.gif)
-
-## Deploying the Truss
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd sdxl-controlnet-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `sdxl-controlnet-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Using the model
-
-The model takes a JSON payload with two fields:
-
-- `prompt`: Text describing the desired image.
-- `image`: Base64 encoded input image.
-
-It returns a JSON object with the `result` field containing the generated image.
-
-## Example Usage
-
-You can also invoke the SDXL + ControlNet model from Python using the `baseten` SDK:
-
-```python
-import baseten
-
-model = baseten.deployed_model_version_id("MODEL_VERSION_ID") # you can get this from the Baseten web UI
-
-image = open("cat.png", "rb").read()
-image_b64 = base64.b64encode(image).decode("utf-8")
-
-request = {
-  "prompt": "A painting of a cat",
-  "image": "data:image/png;base64," + image_b64
-}
-
-response = model.predict(request)
-```
-
-The response will contain a base64 encoded image that you can save:
-
-```python
-import base64
-
-img = base64.b64decode(response["result"])
-
-with open("generated.png", "wb") as f:
-  f.write(img)
-```
-
-You can also invoke the model via REST API:
-
-```bash
-curl -X POST "https://app.baseten.co/model_versions/VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H "Authorization: Api-Key {API_KEY}" \
-     -d '{"prompt": "A painting of a cat",
-           "image": "data:image/png;base64,..."}'
-```
-
-The API will return a JSON response containing the generated image encoded in base64.
diff --git a/stable-diffusion/sdxl-controlnet/config.yaml b/stable-diffusion/sdxl-controlnet/config.yaml
deleted file mode 100644
index 3ef4081ae..000000000
--- a/stable-diffusion/sdxl-controlnet/config.yaml
+++ /dev/null
@@ -1,29 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: aerial view, a futuristic research complex in a bright foggy jungle, hard
-      lighting
-  pretty_name: Stable Diffusion ControlNet
-  tags:
-  - image-generation
-model_name: SDXL ControlNet
-python_version: py39
-requirements:
-- accelerate
-- transformers
-- safetensors
-- opencv-python
-- diffusers
-resources:
-  accelerator: A10G
-  cpu: 3500m
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/sdxl-lightning/README.md b/stable-diffusion/sdxl-lightning/README.md
deleted file mode 100644
index 91f932108..000000000
--- a/stable-diffusion/sdxl-lightning/README.md
+++ /dev/null
@@ -1,87 +0,0 @@
-# SDXL Lightning Truss
-
-SDXL Turbo is a significant improvement over SDXL allowing for image generation in a single step while maintaining image quality. This model generates images with a latency of less than 1 second!
-
-This model is packaged using [Truss](https://trussml.com), the simplest way to serve AI/ML models in production.
-
-## Deploy SDXL Lightning
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion/sdxl-lightning
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-diffusion/sdxl-lightning` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-Model inference runs well on an A100.
-
-## Invoking SDXL Turbo
-
-This model only takes two inputs called `prompt` and `num_steps` and outputs a single image with the dimensions(512x512) encoded as a base 64 string.
-
-- `prompt` (required): Text describing the desired image
-
-It returns a JSON object with the `result` field containing the generated image as a base 64 string.
-
-Here is an example of how you can invoke the model using Python:
-
-```python
-import base64
-import requests
-import os
-
-# Replace the empty string with your model id below
-model_id = ""
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-data = {
-  "prompt": "a picture of a rhino wearing a suit",
-}
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/production/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=data
-)
-
-# Get output image
-res = res.json()
-img_b64 = res.get("result")
-img = base64.b64decode(img_b64)
-
-# Save the base64 string to a PNG
-img_file = open("sdxl-output-1.png", "wb")
-img_file.write(img)
-img_file.close()
-os.system("open sdxl-output-1.png")
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://model-<model-id>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "An illustration of a rocket taking off an alien planet, vector art"
-         }'
-```
diff --git a/stable-diffusion/sdxl-lightning/config.yaml b/stable-diffusion/sdxl-lightning/config.yaml
deleted file mode 100644
index 847f14942..000000000
--- a/stable-diffusion/sdxl-lightning/config.yaml
+++ /dev/null
@@ -1,24 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: A tree in a field under the night sky
-  pretty_name: SDXL Lightning
-  tags:
-  - image-generation
-model_name: SDXL Lightning
-python_version: py310
-requirements:
-- torch==2.0.1
-- transformers==4.35.2
-- diffusers==0.23.1
-- hf_transfer==0.1.4
-- xformers==0.0.22
-- accelerate==0.24.1
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/stable-diffusion/sdxl-lora-swapping/README.md b/stable-diffusion/sdxl-lora-swapping/README.md
deleted file mode 100644
index ee1fadadb..000000000
--- a/stable-diffusion/sdxl-lora-swapping/README.md
+++ /dev/null
@@ -1,68 +0,0 @@
-# Stable Diffusion XL with LoRA Swapping
-
-This Truss provides an example on how to hot-swap LoRAs with Stable Diffusion XL. This is useful when you have a bunch of customers with different fine-tunes but you want to use the same GPU instance.
-
-On Baseten, you can expect LoRA downloading + loading to take ~4s for a 120 MB LoRA and ~2s for a standard 20 MB LoRA. Generation time will be ~6s on an A100.
-
-## Deploying the Truss
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd sdxl-lora-swapping
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `sdxl-lora-swapping` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Using the model
-
-The model takes a JSON payload with two main fields:
-
-- `prompt`: Text describing the desired image. Make sure the prompt includes key phrases to activate the LoRA if needed. For example, many pixel style LoRAs require the words "pixel style" to be present in the prompt.
-- `lora`: Dict with two keys, `repo_id` and `weights`. An example `lora` dict would like this:
-
-```lora.json
-{
-    "prompt": "pixel art, an baby giraffe",
-    "lora": {"repo_id": "nerijs/pixel-art-xl", "weights": "pixel-art-xl.safetensors"}
-}
-```
-
-There are many ojptional fields including negative prompt and size that you can find in the `predict` function of the `model.py`.
-
-It returns a JSON object with the `result` field containing the generated image.
-
-## Example Usage
-
-You can use the Truss CLI to generate an image:
-
-```
-truss predict --model PRIMARY_MODEL_ID -f lora.json
-```
-
-Here, `lora.json` should contain the data payload for the request.
-
-You can also invoke the model via REST API:
-
-```bash
-curl -X POST "https://app.baseten.co/model_versions/VERSION_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H "Authorization: Api-Key {API_KEY}" \
-     -d '{"prompt": "pixel art, painting of a cat"}'
-```
-
-The API will return a JSON response containing the generated image encoded in base64.
diff --git a/stable-diffusion/sdxl-lora-swapping/config.yaml b/stable-diffusion/sdxl-lora-swapping/config.yaml
deleted file mode 100644
index e78e10fba..000000000
--- a/stable-diffusion/sdxl-lora-swapping/config.yaml
+++ /dev/null
@@ -1,26 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    lora:
-      repo_id: nerijs/pixel-art-xl
-      weights: pixel-art-xl.safetensors
-    prompt: pixel art, an baby giraffe
-model_name: Stable Diffusion XL with LoRA Swapping
-python_version: py311
-requirements:
-- accelerate==0.23.0
-- transformers==4.33.2
-- safetensors==0.3.3
-- opencv-python==4.8.0.76
-- diffusers==0.21.2
-resources:
-  accelerator: A100
-  cpu: 3500m
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/sdxl-lora/README.md b/stable-diffusion/sdxl-lora/README.md
deleted file mode 100644
index 2d08ea240..000000000
--- a/stable-diffusion/sdxl-lora/README.md
+++ /dev/null
@@ -1,65 +0,0 @@
-# Stable Diffusion XL Truss with LoRA
-
-This is a [Truss](https://truss.baseten.co/) for Stable Diffusion XL using the `DiffusionPipeline` from the `diffusers` library. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Stable Diffusion XL with support for Low Rank Adaptation (LoRA).
-
-## Overview
-
-Stable Diffusion XL is an enhanced version of Stable Diffusion that generates 1024x1024 images. This Truss utilizes the `stabilityai/stable-diffusion-xl-base-1.0` model.
-
-This Truss also includes support for Low Rank Adaptation (LoRA). LoRA allows you to finetune Stable Diffusion on specific prompts to improve image quality and coherence. This Truss uses the `minimaxir/sdxl-wrong-lora` weights by default, which improves image quality by reducing malformed limbs and objects.
-
-## Deploying the Truss
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd sdxl-lora-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `sdxl-lora-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## API Documentation
-
-The API has one endpoint, `/predict`, which generates images.
-
-**Parameters:**
-
-- `prompt`: Input text prompt
-- `size`: Image size, default 1024
-- `use_refiner`: Enable/disable secondary refiner model, default true
-- `high_noise_frac`: Noise level for refiner model
-- `num_inference_steps`: Number of denoising steps, default 30
-
-**Example Request:**
-
-```json
-{
-  "prompt": "A beautiful painting of a fox in the forest",
-  "use_refiner": true,
-  "high_noise_frac": 0.8,
-  "num_inference_steps": 20
-}
-```
-
-**Returns:** Base64-encoded PNG image
-
-```json
-{
-    "result": "..." // base64 encoded image
-}
-
-```
diff --git a/stable-diffusion/sdxl-lora/config.yaml b/stable-diffusion/sdxl-lora/config.yaml
deleted file mode 100644
index 1a7d477d0..000000000
--- a/stable-diffusion/sdxl-lora/config.yaml
+++ /dev/null
@@ -1,20 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: Stable Diffusion XL with LoRA
-python_version: py311
-requirements:
-- accelerate
-- transformers
-- safetensors
-- opencv-python
-- diffusers
-resources:
-  accelerator: A10G
-  cpu: 3500m
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/sdxl-turbo/README.md b/stable-diffusion/sdxl-turbo/README.md
deleted file mode 100644
index f21d80ffa..000000000
--- a/stable-diffusion/sdxl-turbo/README.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# SDXL Turbo Truss
-
-SDXL Turbo is a significant improvement over SDXL allowing for image generation in a single step while maintaining image quality. This model generates images with a latency of less than 1 second!
-
-This model is packaged using [Truss](https://trussml.com), the simplest way to serve AI/ML models in production.
-
-![turbo](https://github.com/htrivedi99/truss-examples/assets/15642666/904aac3e-f06d-4161-b2c6-cf48522779ce)
-
-
-## Deploy SDXL Turbo
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd sdxl-turbo
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `sdxl-turbo` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-### Hardware notes
-
-Model inference runs well on an T4 with 16 GB of VRAM, with invocation time averaging ~1 second.
-
-## Invoking SDXL Turbo
-
-This model only takes two inputs called `prompt` and `num_steps` and outputs a single image with the dimensions(512x512) encoded as a base 64 string.
-
-- `prompt` (required): Text describing the desired image
-- `num_steps` (optional): Number of steps the model should iterate. Must be between 1-4.
-
-It returns a JSON object with the `result` field containing the generated image as a base 64 string.
-
-Here is an example of how you can invoke the model using Python:
-
-```python
-BASE64_PREAMBLE = "data:image/png;base64,"
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-
-headers = {"Authorization": f"Api-Key <BASETEN-API-KEY>"}
-resp = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers=headers,
-    json={"prompt": "An illustration of a rocket taking off an alien planet, vector art"},
-)
-resp = resp.json()
-img = b64_to_pil(resp.get("result"))
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://model-<model-id>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "An illustration of a rocket taking off an alien planet, vector art"
-         }'
-```
diff --git a/stable-diffusion/sdxl-turbo/config.yaml b/stable-diffusion/sdxl-turbo/config.yaml
deleted file mode 100644
index d9fded05b..000000000
--- a/stable-diffusion/sdxl-turbo/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_cache:
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - '*.txt'
-  repo_id: stabilityai/sdxl-turbo
-  use_volume: false
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: A tree in a field under the night sky
-  pretty_name: SDXL Turbo
-  tags:
-  - image-generation
-model_name: SDXL Turbo
-python_version: py310
-requirements:
-- torch==2.0.1
-- transformers==4.35.2
-- diffusers==0.23.1
-- hf_transfer==0.1.4
-- xformers==0.0.22
-- accelerate==0.24.1
-resources:
-  accelerator: T4
-  cpu: '3'
-  memory: 20Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/stable-diffusion/stable-diffusion-3-medium/README.md b/stable-diffusion/stable-diffusion-3-medium/README.md
deleted file mode 100644
index 075dd0a18..000000000
--- a/stable-diffusion/stable-diffusion-3-medium/README.md
+++ /dev/null
@@ -1,83 +0,0 @@
-# Stable Diffusion 3 Medium Truss
-
-This is a truss for the brand new Stable Diffusion 3 Medium model.
-
-## Deploy Stable Diffusion 3
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion/stable-diffusion-3-medium
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-diffusion-3-medium` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using Stable Diffusion through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-## Invoking Stable Diffusion 3
-
-The output will be a dictionary with a key `data` mapping to a base64 encoded image. It's processed with this script:
-
-```python
-import requests
-import os
-import base64
-from PIL import Image
-from io import BytesIO
-
-# Replace the empty string with your model id below
-model_id = ""
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-# Function used to convert a base64 string to a PIL image
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-data = {
-  "prompt": "a little boy looking through a large magical portal, the boy sees a futuristic human civilization in that portal, extremely detailed, trending on artstation, 8k"
-}
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/production/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=data
-)
-
-# Get output image
-res = res.json()
-output = res.get("data")
-
-# Convert the base64 model output to an image
-img = b64_to_pil(output)
-img.save("output_image.png")
-os.system("open output_image.png")
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://model-{model_id}.api.baseten.co/production/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "A tree in a field under the night sky"
-         }'
-```
-
-Again, the model will return a dictionary containing the base64-encoded image, which will need to be decoded and saved.
diff --git a/stable-diffusion/stable-diffusion-3-medium/config.yaml b/stable-diffusion/stable-diffusion-3-medium/config.yaml
deleted file mode 100644
index d01dfd423..000000000
--- a/stable-diffusion/stable-diffusion-3-medium/config.yaml
+++ /dev/null
@@ -1,24 +0,0 @@
-environment_variables:
-  HF_HUB_OFFLINE: 1
-external_package_dirs: []
-model_metadata: {}
-model_cache:
-  - repo_id: stabilityai/stable-diffusion-3-medium-diffusers
-    use_volume: false
-model_name: Stable Diffusion 3 Medium
-python_version: py310
-requirements:
-  - diffusers==0.29.0
-  - transformers
-  - accelerate
-  - sentencepiece
-  - protobuf
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets:
-  hf_access_token: ""
-system_packages:
-  - ffmpeg
-  - libsm6
-  - libxext6
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/README.md b/stable-diffusion/stable-diffusion-inpainting-trt/README.md
deleted file mode 100644
index 6d8d306f6..000000000
--- a/stable-diffusion/stable-diffusion-inpainting-trt/README.md
+++ /dev/null
@@ -1,76 +0,0 @@
-# Stable Diffusion Inpainting TensorRT
-
-This is a [Truss](https://truss.baseten.co/) for Stable Diffusion v1.5 Inpainting. It has been optimized to increase performance using TensorRT.
-
-## Deploy Stable Diffusion Inpainting TensorRT
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion/stable-diffusion-inpainting-trt
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-diffusion-inpainting-trt` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using Stable Diffusion through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-## Invoking Stable Diffusion Inpainting TensorRT
-
-The model accepts a few inputs:
-- __prompt__(required): Text describing the output image.
-- __negative_prompt__(optional): Text used to steer the model away from undesired output.
-- __image__(required): The input image used for inpainting in the form of a base64 string. The image should be 512 x 512 px in size.
-- __mask__(required): The an image representing the mask or the area that you want stable diffusion to generate over. It should also be a base64 string and the same dimensions as the input image.
-
-```python
-from PIL import Image
-import base64
-import requests
-
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-def pil_to_b64(pil_img):
-    buffered = BytesIO()
-    pil_img.save(buffered, format="PNG")
-    img_str = base64.b64encode(buffered.getvalue()).decode("utf-8")
-    return img_str
-
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-
-data = {
-    "prompt": "A tiger",
-    "image": pil_to_b64(Image.open("/path/to/image/dog.png")),
-    "mask": pil_to_b64(Image.open("/path/to/mask/mask.png"))
-}
-
-headers = {"Authorization": "Api-Key <BASETEN-API-KEY>"}
-
-res = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers=headers,
-    json=data,
-)
-
-res = res.json()
-outputs = res.get("outputs")
-
-for out in outputs:
-    img = b64_to_pil(out)
-    img.show()
-```
diff --git a/stable-diffusion/stable-diffusion-inpainting-trt/config.yaml b/stable-diffusion/stable-diffusion-inpainting-trt/config.yaml
deleted file mode 100644
index 7cde4606e..000000000
--- a/stable-diffusion/stable-diffusion-inpainting-trt/config.yaml
+++ /dev/null
@@ -1,11 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_name: Stable Diffusion Inpainting TRT
-python_version: py310
-requirements: []
-requirements_file: requirements.txt
-resources:
-  accelerator: A10G
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/README.md b/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/README.md
deleted file mode 100644
index 400d9f38c..000000000
--- a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/README.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# Stable Diffusion XL TensorRT Truss for H100
-
-Stable Diffusion XL 1.0 is the largest, most capable open-source image generation model of its kind. This README covers deploying and invoking this model.
-
-This model is packaged using [Truss](https://trussml.com), the simplest way to serve AI/ML models in production.
-
-## Deploy Stable Diffusion XL TensorRT
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion/stable-diffusion-xl-1.0-trt-h100
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-diffusion-xl-1.0-trt-h100` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using SDXL through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-### Hardware notes
-
-Running inference on an H100 cuts invocation time to <2 seconds.
-
-## Invoking Stable Diffusion XL TensorRT
-
-Stable Diffusion XL TensorRT returns an image in Base 64, which is not super useful as a string in your terminal. So we included a helpful utility script to show and save the image. Pipe the model results into the script.
-
-```sh
-truss predict -d '{"prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"}' | python show.py
-```
-
-The output will be a dictionary with a key `data` mapping to a base64 encoded image. It's processed with this script:
-
-```python
-import json
-import base64
-import os, sys
-
-resp = sys.stdin.read()
-image = json.loads(resp)["data"]
-img=base64.b64decode(image)
-
-file_name = f'{image[-10:].replace("/", "")}.jpeg'
-img_file = open(file_name, 'wb')
-img_file.write(img)
-img_file.close()
-os.system(f'open {file_name}')
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://app.baseten.co/models/{MODEL_ID}/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
-         }'
-```
-
-Again, the model will return a dictionary containing the base64-encoded image, which will need to be decoded and saved.
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/config.yaml b/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/config.yaml
deleted file mode 100644
index c7608fd76..000000000
--- a/stable-diffusion/stable-diffusion-xl-1.0-trt-h100/config.yaml
+++ /dev/null
@@ -1,62 +0,0 @@
-base_image:
-  image: nvcr.io/nvidia/pytorch:23.11-py3
-  python_executable_path: /usr/bin/python
-description: Generate original images from text prompts.
-environment_variables:
-  HF_HUB_ENABLE_HF_TRANSFER: 1
-external_package_dirs: []
-model_cache:
-- repo_id: baseten/sdxl-1.0-trt-8.6.1.post1-engine-H100
-  use_volume: false
-- allow_patterns:
-  - config.json
-  - diffusion_pytorch_model.safetensors
-  repo_id: madebyollin/sdxl-vae-fp16-fix
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - sd_xl_base_1.0.safetensors
-  repo_id: stabilityai/stable-diffusion-xl-base-1.0
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - sd_xl_refiner_1.0.safetensors
-  repo_id: stabilityai/stable-diffusion-xl-refiner-1.0
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: Astronaut in a jungle, cold color palette, muted colors, detailed, 8k
-  pretty_name: Stable Diffusion XL - TensorRT
-  tags:
-  - image-generation
-model_name: Stable Diffusion XL - TensorRT
-python_version: py39
-requirements:
-- accelerate
-- colored
-- cuda-python
-- ftfy
-- nvtx
-- opencv-python==4.8.0.74
-- scipy
-- transformers==4.31.0
-- safetensors
-- hf_transfer
-- diffusers==0.23.1
-- invisible-watermark>=0.2.0
-- --extra-index-url https://pypi.ngc.nvidia.com
-- polygraphy
-- --extra-index-url https://pypi.nvidia.com
-- tensorrt==8.6.1.post1
-resources:
-  accelerator: H100
-  use_gpu: true
-runtime:
-  predict_concurrency: 1
-secrets: {}
-system_packages:
-- python3.10-venv
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt/README.md b/stable-diffusion/stable-diffusion-xl-1.0-trt/README.md
deleted file mode 100644
index 23fe732c9..000000000
--- a/stable-diffusion/stable-diffusion-xl-1.0-trt/README.md
+++ /dev/null
@@ -1,74 +0,0 @@
-# Stable Diffusion XL TensorRT Truss
-
-Stable Diffusion XL 1.0 is the largest, most capable open-source image generation model of its kind. This README covers deploying and invoking this model.
-
-This model is packaged using [Truss](https://trussml.com), the simplest way to serve AI/ML models in production.
-
-## Deploy Stable Diffusion XL TensorRT
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion/stable-diffusion-xl-1.0-trt
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-diffusion-xl-1.0-trt` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using SDXL through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-### Hardware notes
-
-Running inference on an A100 cuts invocation time to ~3.5 seconds.
-
-## Invoking Stable Diffusion XL TensorRT
-
-Stable Diffusion XL TensorRT returns an image in Base 64, which is not super useful as a string in your terminal. So we included a helpful utility script to show and save the image. Pipe the model results into the script.
-
-```sh
-truss predict -d '{"prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"}' | python show.py
-```
-
-The output will be a dictionary with a key `data` mapping to a base64 encoded image. It's processed with this script:
-
-```python
-import json
-import base64
-import os, sys
-
-resp = sys.stdin.read()
-image = json.loads(resp)["data"]
-img=base64.b64decode(image)
-
-file_name = f'{image[-10:].replace("/", "")}.jpeg'
-img_file = open(file_name, 'wb')
-img_file.write(img)
-img_file.close()
-os.system(f'open {file_name}')
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://app.baseten.co/models/{MODEL_ID}/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "Astronaut in a jungle, cold color palette, muted colors, detailed, 8k"
-         }'
-```
-
-Again, the model will return a dictionary containing the base64-encoded image, which will need to be decoded and saved.
diff --git a/stable-diffusion/stable-diffusion-xl-1.0-trt/config.yaml b/stable-diffusion/stable-diffusion-xl-1.0-trt/config.yaml
deleted file mode 100644
index facf6a925..000000000
--- a/stable-diffusion/stable-diffusion-xl-1.0-trt/config.yaml
+++ /dev/null
@@ -1,62 +0,0 @@
-base_image:
-  image: nvcr.io/nvidia/pytorch:23.11-py3
-  python_executable_path: /usr/bin/python
-description: Generate original images from text prompts.
-environment_variables:
-  HF_HUB_ENABLE_HF_TRANSFER: 1
-external_package_dirs: []
-model_cache:
-- repo_id: baseten/sdxl-1.0-trt-8.6.1.post1-engine
-  use_volume: false
-- allow_patterns:
-  - config.json
-  - diffusion_pytorch_model.safetensors
-  repo_id: madebyollin/sdxl-vae-fp16-fix
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - sd_xl_base_1.0.safetensors
-  repo_id: stabilityai/stable-diffusion-xl-base-1.0
-- allow_patterns:
-  - '*.json'
-  - '*.fp16.safetensors'
-  - sd_xl_refiner_1.0.safetensors
-  repo_id: stabilityai/stable-diffusion-xl-refiner-1.0
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  example_model_input:
-    prompt: Astronaut in a jungle, cold color palette, muted colors, detailed, 8k
-  pretty_name: Stable Diffusion XL - TensorRT
-  tags:
-  - image-generation
-model_name: Stable Diffusion XL - TensorRT
-python_version: py39
-requirements:
-- accelerate
-- colored
-- cuda-python
-- ftfy
-- nvtx
-- opencv-python==4.8.0.74
-- scipy
-- transformers==4.31.0
-- safetensors
-- hf_transfer
-- diffusers==0.23.1
-- invisible-watermark>=0.2.0
-- --extra-index-url https://pypi.ngc.nvidia.com
-- polygraphy
-- --extra-index-url https://pypi.nvidia.com
-- tensorrt==8.6.1.post1
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  predict_concurrency: 1
-secrets: {}
-system_packages:
-- python3.10-venv
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/stable-diffusion/stable-diffusion-xl-1.0/README.md b/stable-diffusion/stable-diffusion-xl-1.0/README.md
deleted file mode 100644
index 39a1d670f..000000000
--- a/stable-diffusion/stable-diffusion-xl-1.0/README.md
+++ /dev/null
@@ -1,178 +0,0 @@
-# Stable Diffusion XL Truss
-
-Stable Diffusion XL 1.0 is the largest, most capable open-source image generation model of its kind. This README covers deploying and invoking this model.
-
-This model is packaged using [Truss](https://trussml.com), the simplest way to serve AI/ML models in production.
-
-## Deploy Stable Diffusion XL
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion-xl-1.0
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-diffusion-xl-1.0` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using SDXL through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-### Hardware notes
-
-Model inference runs well on an A10 with 24 GB of VRAM, with invocation time averaging ~8 seconds. If speed is essential, running inference on an A100 cuts invocation time to ~4 seconds.
-
-## Invoking Stable Diffusion XL
-
-Stable Diffusion XL returns an image in Base 64, which is not super useful as a string in your terminal. So we included a helpful utility script to show and save the image. Pipe the model results into the script.
-
-```sh
-truss predict -d '{"prompt": "A tree in a field under the night sky"}' | python show.py
-```
-
-The output will be a dictionary with a key `data` mapping to a base64 encoded image. It's processed with this script:
-
-```python
-import json
-import base64
-import os, sys
-
-resp = sys.stdin.read()
-image = json.loads(resp)["data"]
-img=base64.b64decode(image)
-
-file_name = f'{image[-10:].replace("/", "")}.jpeg'
-img_file = open(file_name, 'wb')
-img_file.write(img)
-img_file.close()
-os.system(f'open {file_name}')
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://app.baseten.co/models/MODEL_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "A tree in a field under the night sky",
-           "use_refiner": True
-         }'
-```
-
-Again, the model will return a dictionary containing the base64-encoded image, which will need to be decoded and saved.
-
-Here is a complete example of invoking this model in Python:
-
-## Model Input
-
-```python
-import requests
-import os
-import base64
-from PIL import Image
-from io import BytesIO
-
-# Replace the empty string with your model id below
-model_id = ""
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-# Function used to convert a base64 string to a PIL image
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-data = {
-  "prompt": "a little boy looking through a large magical portal, the boy sees a futuristic human civilization in that portal, extremely detailed, trending on artstation, 8k"
-}
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/production/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=data
-)
-
-# Get output image
-res = res.json()
-output = res.get("data")
-
-# Convert the base64 model output to an image
-img = b64_to_pil(output)
-img.save("output_image.png")
-os.system("open output_image.png")
-```
-
-## Model Output
-```json
-{"data": "iVBORw0KGgoAAAANSUhEUgAABAAAAAQA..."}
-```
-
-Here is the output image for the prompt shown in the request above:
-![a_little_boy_looking_through_a_large_magical_portal,_the_boy_sees_a_futuristic_human_civilization_in](https://github.com/htrivedi99/truss-examples/assets/15642666/c534c752-29cb-4da8-b24e-fda6bef5876c)
-
-Here's another example using more SDXL configurations:
-
-```python
-import requests
-import os
-import base64
-from PIL import Image
-from io import BytesIO
-
-# Replace the empty string with your model id below
-model_id = ""
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-BASE64_PREAMBLE = "data:image/png;base64,"
-
-# Function used to convert a base64 string to a PIL image
-def b64_to_pil(b64_str):
-    return Image.open(BytesIO(base64.b64decode(b64_str.replace(BASE64_PREAMBLE, ""))))
-
-data = {
-  "prompt": "Extremely detailed and intricate scene of baby phoenix hatchling cuddled up resting on a pile of ashes surrounded by fire and smoke, rays of sunlight shine on the phoenix, in the background is a dense dark forest, settings: f/8 aperture, full shot, hyper realistic, 4k",
-  "negative_prompt": "worst quality, low quality",
-  "width": 1248,
-  "height": 832,
-  "num_inference_steps": 35,
-  "use_refiner": False,
-  "scheduler": "DPM++ 2M",
-  "guidance_scale": 14
-}
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/production/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=data
-)
-
-# Get output image
-res = res.json()
-output = res.get("data")
-
-# Convert the base64 model output to an image
-img = b64_to_pil(output)
-img.save("output_image.png")
-os.system("open output_image.png")
-```
-
-Here is the model output:
-```json
-{"data": "iVBORw0KGgoAAAANSUhEUgAABAAAAAQA..."}
-```
-
-This is the output image for the prompt above:
-![Extremely_detailed_and_intricate_scene_of_baby_phoenix_hatchling_cuddled_up_resting_on_a_pile_of_ash](https://github.com/htrivedi99/truss-examples/assets/15642666/1fbf004a-9741-4c83-90a8-9e7d51dfd4e2)
diff --git a/stable-diffusion/stable-diffusion/README.md b/stable-diffusion/stable-diffusion/README.md
deleted file mode 100644
index 55ad20941..000000000
--- a/stable-diffusion/stable-diffusion/README.md
+++ /dev/null
@@ -1,83 +0,0 @@
-# Stable Diffusion Truss
-
-This is a [Truss](https://truss.baseten.co/) for Stable Diffusion v2.1 using the `StableDiffusionPipeline` from the `diffusers` library. This README will walk you through how to deploy this Truss on Baseten to get your own instance of the Stable Diffusion.
-
-## Deploy Stable Diffusion
-
-First, clone this repository:
-
-```
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-diffusion-truss` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using Stable Diffusion through the Baseten platform! Navigate to the Baseten UI to watch the model build and deploy and invoke it via the REST API.
-
-## Invoking Stable Diffusion
-
-Stable Diffusion returns an image in Base 64, which is not super useful as a string in your terminal. So we included a helpful utility script to show and save the image. Pipe the model results into the script.
-
-```sh
-truss predict -d '{"prompt": "A tree in a field under the night sky"}' | python show.py
-```
-
-The output will be a dictionary with a key `data` mapping to a base64 encoded image. It's processed with this script:
-
-```python
-import json
-import base64
-import os, sys
-
-resp = sys.stdin.read()
-image = json.loads(resp)["data"]
-img=base64.b64decode(image)
-
-file_name = f'{image[-10:].replace("/", "")}.jpeg'
-img_file = open(file_name, 'wb')
-img_file.write(img)
-img_file.close()
-os.system(f'open {file_name}')
-```
-
-You can also invoke your model via a REST API:
-
-```
-curl -X POST "https://app.baseten.co/models/MODEL_ID/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{
-           "prompt": "A tree in a field under the night sky"
-         }'
-```
-
-Again, the model will return a dictionary containing the base64-encoded image, which will need to be decoded and saved.
-
-### Stable Diffusion API documentation
-
-This section provides an overview of the Stable Diffusion API, its parameters, and how to use it. The API consists of a single route named `predict`, which you can invoke to generate images based on the provided parameters.
-
-#### API route: `predict`
-
-The predict route is the primary method for generating images based on a given set of parameters. It takes several parameters:
-
-- **prompt**: The input text you'd like to generate an image for
-- **scheduler**: (optional, default: DDIM) The scheduler used for the diffusion process. Choose from: "ddim", "dpm", "euler", "lms", or "pndm".
-- **seed**: (optional) A random seed for deterministic results. If not provided, a random seed will be generated.
-- **negative_prompt**: (optional) A string representing the negative prompt, or prompts that indicate what you don't want to generate.
-
-The API also supports passing any parameter supported by Diffuser's `StableDiffusionPipeline`.
diff --git a/stable-diffusion/stable-diffusion/config.yaml b/stable-diffusion/stable-diffusion/config.yaml
deleted file mode 100644
index 208e248e8..000000000
--- a/stable-diffusion/stable-diffusion/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-description: Generate original images from text prompts.
-environment_variables: {}
-external_data:
-- local_data_path: unet/diffusion_pytorch_model.bin
-  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/stable-diffusion-truss/unet/diffusion_pytorch_model.bin
-- local_data_path: text_encoder/pytorch_model.bin
-  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/stable-diffusion-truss/text_encoder/pytorch_model.bin
-- local_data_path: vae/diffusion_pytorch_model.bin
-  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/stable-diffusion-truss/vae/diffusion_pytorch_model.bin
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/sd.png
-  pretty_name: Stable Diffusion
-  tags:
-  - image-generation
-model_name: Stable Diffusion
-python_version: py39
-requirements:
-- diffusers
-- transformers
-- torch
-- scipy
-- accelerate
-- pillow
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 14Gi
-  use_gpu: true
-secrets: {}
-spec_version: 2.0
-system_packages: []
diff --git a/stable-diffusion/stable-video-diffusion/README.md b/stable-diffusion/stable-video-diffusion/README.md
deleted file mode 100644
index fb9ceb9f5..000000000
--- a/stable-diffusion/stable-video-diffusion/README.md
+++ /dev/null
@@ -1,90 +0,0 @@
-# Stable Diffusion Video
-
-This Truss allows you to create small videos ( < 10 seconds) using the [Stable Video Diffusion model](https://stability.ai/news/stable-video-diffusion-open-ai-video-model). The model only requires a single image as input and converts it into a video.
-
-
-https://github.com/htrivedi99/truss-examples/assets/15642666/53d3ea1f-952e-4224-8bd4-3ff95eeac8c0
-
-
-## Deploying the Truss
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd stable-diffusion/stable-video-diffusion
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `stable-video-diffusion` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Using the model
-
-The model takes a JSON payload with one required field and 4 optional fields:
-
-- `image` (required): The input image as a base64 string. This model was trained on images that are 1024 × 576 in size. Other image dimensions work as well for the most part, but to get the best results using 1024 × 576 images is recommended.
-- `num_frames` (optional) - The total number of frames in the output clip
-- `num_steps` (optional) - Steps takes at each iteration
-- `fps` (optional) - Frames per second for the clip
-- `decoding_t` (optional) - The number of frames decoded per second. This number cannot be greater than 10 as the GPU runs out of memory at that point.
-
-It returns a JSON object with the `output` key containing the generated video as a base64 string.
-
-## Example Usage
-
-Here is how you can invoke the model using Python:
-
-```python
-def base64_to_mp4(base64_string, output_file_path):
-    binary_data = base64.b64decode(base64_string)
-    with open(output_file_path, "wb") as output_file:
-        output_file.write(binary_data)
-
-
-def image_to_base64(file_path):
-    with open(file_path, "rb") as image_file:
-        binary_data = image_file.read()
-        base64_data = base64.b64encode(binary_data)
-        base64_string = base64_data.decode("utf-8")
-
-    return base64_string
-
-data = {
-    "image": image_to_base64("path/to/image/cheetah.jpeg"),
-    "num_frames": 14,
-    "fps": 10,
-    "decoding_t": 5
-}
-headers = {"Authorization": f"Api-Key <BASETEN-API-KEY>"}
-res = requests.post("https://model-<model-id>.api.baseten.co/development/predict", headers=headers, json=data)
-res = res.json()
-
-base64_output = res.get("output")
-base64_to_mp4(base64_output, "output_video.mp4")
-```
-
-You can also invoke the model via REST API:
-
-```bash
-curl -X POST "https://model-<model-id>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H "Authorization: Api-Key {BASETEN-API-KEY}" \
-     -d '{"image": "<image-as-b64-string>"}'
-```
-
-For inspiration, here are some sample input images you can use:
-
-![image1](sample_images/cheetah.jpeg)
-![image2](sample_images/racecar.jpeg)
diff --git a/stable-diffusion/stable-video-diffusion/config.yaml b/stable-diffusion/stable-video-diffusion/config.yaml
deleted file mode 100644
index 3dc105678..000000000
--- a/stable-diffusion/stable-video-diffusion/config.yaml
+++ /dev/null
@@ -1,39 +0,0 @@
-description: Stable Video Diffusion can turn any image into a short video.
-environment_variables: {}
-external_data:
-- local_data_path: weights/svd.safetensors
-  url: https://huggingface.co/stabilityai/stable-video-diffusion-img2vid/resolve/main/svd.safetensors
-- local_data_path: weights/ViT-L-14.pt
-  url: https://openaipublic.azureedge.net/clip/models/b8cca3fd41ae0c99ba7e8951adf17d267cdb84cd88be6f7c2e0eca1737a03836/ViT-L-14.pt
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/explore/stability.png
-  cover_image_url: https://cdn.baseten.co/production/static/explore/stable-video-diffusion.png
-  tags:
-  - image-to-video
-model_name: Stable Video Diffusion
-python_version: py310
-requirements:
-- einops==0.7.0
-- fire==0.5.0
-- omegaconf==2.3.0
-- git+https://github.com/openai/CLIP.git@2dbac9065bb0b4ffc28ecf0e94758261d1ddfdb0
-- lightning==2.1.2
-- kornia==0.7.0
-- open-clip-torch==2.23.0
-- invisible-watermark==0.2.0
-- xformers==0.0.22
-- opencv-python==4.8.0.76
-- scipy==1.11.3
-- transformers==4.35.2
-- hf_transfer==0.1.4
-- git+https://github.com/Stability-AI/generative-models.git@059d8e9cd9c55aea1ef2ece39abf605efb8b7cc9
-resources:
-  accelerator: A100
-  cpu: '4'
-  memory: 16Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- libgl1-mesa-glx
-- ffmpeg
diff --git a/text-embeddings-inference/README.md b/text-embeddings-inference/README.md
deleted file mode 100644
index a52303ffa..000000000
--- a/text-embeddings-inference/README.md
+++ /dev/null
@@ -1,115 +0,0 @@
-# Notice
-This section has moved to [jina-ai/jina-embeddings-v2-base-en-TEI](https://github.com/basetenlabs/truss-examples/tree/main/11-embeddings-reranker-classification-tensorrt) with an overview over fast embeddings.
-
-# Text Embeddings Inference Truss
-
-This is a Trussless Customer Server example to deploy [text-embeddings-inference](https://github.com/huggingface/text-embeddings-inference), a high performance server that handles text-embeddings, ranranking and classification models as api.
-
-## Deployment
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. [Required for gated model] Retrieve your Hugging Face token from the [settings](https://huggingface.co/settings/tokens). Set your Hugging Face token as a Baseten secret [here](https://app.baseten.co/settings/secrets) with the key `hf_access_key`.
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd text-embeddings-inference
-```
-
-With `text-embeddings-inference` as your working directory, you can deploy the model with the following command, paste your Baseten API key if prompted.
-
-```sh
-truss push --publish
-```
-
-## Performance Optimization:
-
-The config.yaml contains a couple of variables that can be tuned, depending on:
-- which GPU is used
-- which model is deployed
-- how many concurrent requests users are sending
-
-The deployment example is for Bert-large and a Nvidia-L4. Bert-large has a maxiumum sequence length of 512 tokens per sentence.
-For Bert-large architecture & the L4, there are marginal gains above a batch-size of 16000 tokens.
-
-### Concurrent requests
-```
---max-concurrent-requests 40
-# and
-runtime:
-  predict_concurrency : 40
-```
-The following set the number of parallel `post` requests.
-In this case we allow 40 parallel requests to be handled per replica & should allow to batch requests from multiple users together, reaching high token counts. Potentially 40 single parallel requests with one sequence each could fully utilize the GPU. `1*40*512=20480`
-
-
-### Tokens per batch
-```
---max-batch-tokens 32768
-```
-
-This number of total tokens in a batch. For embedding models, this will determine the VRAM usage.
-As most of TEI's models are implemented with `nested` attention implementation, `32768 tokens` could mean `64 sentence with 512 tokens` or `512 sentences with 64 tokens`. While the first will take slightly longer to compute, the peak VRAM usage will stay roughly the same. For `llama` or `mistral` based `7b` embedding models, we recommend setting it a lower setting e.g.
-```
---max-batch-tokens 8192
-```
-
-### Client batch size
-```
---max-client-batch-size 32
-```
-Client match size determines the number of sentences in a single request.
-Increase if clients cannot send multiple concurrent requests, or if clients require to larger requests size.
-
-### Endpoint, Model Selection, and OpenAPI
-Change to /rerank or /predict if you want to use the rerank or predict endpoint.
-Embedding model.
-Example supported models: https://huggingface.co/models?pipeline_tag=feature-extraction&other=text-embeddings-inference&sort=trending
-```yaml
-  predict_endpoint: /v1/embeddings
-```
-Rerank model.
-Example models https://huggingface.co/models?pipeline_tag=text-classification&other=text-embeddings-inference&sort=trending
-```yaml
-  predict_endpoint: /rerank
-```
-Classification model:
-Example classification model: https://huggingface.co/SamLowe/roberta-base-go_emotions
-```yaml
-  predict_endpoint: /predict
-```
-
-## Call your model
-
-### curl
-
-```bash
-curl -X POST https://model-xxx.api.baseten.co/development/predict \
-        -H "Authorization: Api-Key YOUR_API_KEY" \
-        -d '{"input": "text string"}'
-```
-
-
-### request python library
-
-```python
-import os
-import requests
-
-resp = requests.post(
-    "https://model-xxx.api.baseten.co/environments/production/predict",
-    headers={"Authorization": f"Api-Key {os.environ['BASETEN_API_KEY']}"},
-    json={"input": ["text string", "second string"]},
-)
-
-print(resp.json())
-```
-
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/text-embeddings-inference/config.yaml b/text-embeddings-inference/config.yaml
deleted file mode 100644
index 82332a3ce..000000000
--- a/text-embeddings-inference/config.yaml
+++ /dev/null
@@ -1,30 +0,0 @@
-base_image:
-  # select an image: L4
-  # CPU	baseten/text-embeddings-inference-mirror:cpu-1.6
-  # Turing (T4, ...)	baseten/text-embeddings-inference-mirror:turing-1.6
-  # Ampere 80 (A100, A30)	baseten/text-embeddings-inference-mirror:1.6
-  # Ampere 86 (A10, A10G, A40, ...)	baseten/text-embeddings-inference-mirror:86-1.6
-  # Ada Lovelace (L4, ...)	baseten/text-embeddings-inference-mirror:89-1.6
-  # Hopper (H100/H100 40GB)	baseten/text-embeddings-inference-mirror:hopper-1.6
-  image: baseten/text-embeddings-inference-mirror:89-1.6
-model_metadata:
-  repo_id: BAAI/bge-base-en-v1.5
-docker_server:
-  start_command: sh -c "text-embeddings-router --port 7997 --model-id /data/local-model --max-client-batch-size 32 --max-concurrent-requests 40 --max-batch-tokens 32768"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  # change to /rerank or /predict if you want to use the rerank or predict endpoint
-  # https://huggingface.github.io/text-embeddings-inference/
-  predict_endpoint: /v1/embeddings
-  server_port: 7997
-resources:
-  accelerator: L4
-  use_gpu: true
-model_name: text-embeddings-inference trussless
-build_commands: # optional step to download the weights of the model into the image
-- git clone https://huggingface.co/BAAI/bge-base-en-v1.5 /data/local-model
-runtime:
-  predict_concurrency : 40
-environment_variables:
-  VLLM_LOGGING_LEVEL: WARNING
-  hf_access_token: null
diff --git a/tutorials/README.md b/tutorials/README.md
new file mode 100644
index 000000000..16d553c2f
--- /dev/null
+++ b/tutorials/README.md
@@ -0,0 +1,23 @@
+# Tutorials
+
+Step-by-step guides for learning how to build, configure, and deploy models with Truss. Each tutorial focuses on a single concept and includes a complete, deployable example.
+
+| Directory | Description |
+|-----------|-------------|
+| [getting-started-bert](getting-started-bert/) | Introductory tutorial packaging a BERT model for deployment |
+| [llm-basics](llm-basics/) | Serve a large language model with basic inference |
+| [llm-streaming](llm-streaming/) | Stream token-by-token responses from an LLM |
+| [image-generation](image-generation/) | Generate images using a diffusion model |
+| [speech-to-text](speech-to-text/) | Transcribe audio input to text |
+| [cached-weights](cached-weights/) | Use cached model weights for faster cold starts |
+| [dynamic-batching](dynamic-batching/) | Batch incoming requests for higher throughput |
+| [private-huggingface](private-huggingface/) | Load models from a private HuggingFace repository |
+| [system-packages](system-packages/) | Install system-level dependencies in your Truss environment |
+
+## Deploying
+
+Each tutorial can be deployed to Baseten with:
+
+```bash
+truss push
+```
diff --git a/tutorials/cached-weights/README.md b/tutorials/cached-weights/README.md
new file mode 100644
index 000000000..8f5f47bcd
--- /dev/null
+++ b/tutorials/cached-weights/README.md
@@ -0,0 +1,32 @@
+# Llama with Cached Weights
+
+A tutorial example showing how to deploy Llama with Cached Weights on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [NousResearch/Llama-2-7b-chat-hf](https://huggingface.co/NousResearch/Llama-2-7b-chat-hf) |
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "What is the meaning of life?"
+}'
+```
+
+## Configuration highlights
+
+- Model cache: **volume-mounted** for fast cold starts
diff --git a/tutorials/cached-weights/config.yaml b/tutorials/cached-weights/config.yaml
new file mode 100644
index 000000000..4e9bbb6a1
--- /dev/null
+++ b/tutorials/cached-weights/config.yaml
@@ -0,0 +1,59 @@
+# # Setting up the config.yaml
+#
+# The `config.yaml` file is where you need to include the changes to
+# actually cache the weights at build time.
+description: "A tutorial example showing how to deploy Llama with Cached Weights on Baseten."
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  example_model_input: { "prompt": "What is the meaning of life?" }
+model_name: Llama with Cached Weights
+python_version: py39
+requirements:
+  - accelerate==0.21.0
+  - safetensors==0.3.2
+  - torch==2.0.1
+  - transformers==4.34.0
+  - sentencepiece==0.1.99
+  - numpy==1.26.4
+# # Configuring the model_cache
+#
+# To cache model weights, set the `model_cache` key.
+# The `repo_id` field allows you to specify a Huggingface
+# repo to pull down and cache at build-time, and the `ignore_patterns`
+# field allows you to specify files to ignore. If this is specified, then
+# this repo won't have to be pulled during runtime.
+#
+# Check out the [guide](https://truss.baseten.co/guides/model-cache) for more info.
+model_cache:
+  - repo_id: "NousResearch/Llama-2-7b-chat-hf"
+    revision: main
+    ignore_patterns:
+      - "*.bin"
+    use_volume: true
+    volume_folder: "llama-2-7b-chat-hf"
+
+# The remaining config options are again, similar to what you would
+# configure for the model without the weight caching.
+resources:
+  cpu: "4"
+  memory: 30Gi
+  use_gpu: True
+  accelerator: A10G
+secrets: {}
+# # Deploy the model
+#
+# Deploy the model like you would other Trusses, with:
+# ```bash
+# $ truss push
+# ```
+#  <Note>
+#  The build step will take  longer than with the normal
+#  Llama Truss, since bundling the model weights is now happening during the build.
+#  The deploy step & scale-ups will happen much faster with this approach.
+#  </Note>
+#
+# You can then invoke the model with:
+# ```bash
+# $ truss predict -d '{"inputs": "What is a large language model?"}'
+# ```
diff --git a/06-high-performance-cached-weights/doc.yaml b/tutorials/cached-weights/doc.yaml
similarity index 100%
rename from 06-high-performance-cached-weights/doc.yaml
rename to tutorials/cached-weights/doc.yaml
diff --git a/trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/model/__init__.py b/tutorials/cached-weights/model/__init__.py
similarity index 100%
rename from trt-llm-engine-builder-templates/llama-3_1-8b-instruct/low_ttft/model/__init__.py
rename to tutorials/cached-weights/model/__init__.py
diff --git a/06-high-performance-cached-weights/model/model.py b/tutorials/cached-weights/model/model.py
similarity index 100%
rename from 06-high-performance-cached-weights/model/model.py
rename to tutorials/cached-weights/model/model.py
diff --git a/07-high-performance-dynamic-batching/.gitignore b/tutorials/dynamic-batching/.gitignore
similarity index 100%
rename from 07-high-performance-dynamic-batching/.gitignore
rename to tutorials/dynamic-batching/.gitignore
diff --git a/07-high-performance-dynamic-batching/.truss_ignore b/tutorials/dynamic-batching/.truss_ignore
similarity index 100%
rename from 07-high-performance-dynamic-batching/.truss_ignore
rename to tutorials/dynamic-batching/.truss_ignore
diff --git a/tutorials/dynamic-batching/README.md b/tutorials/dynamic-batching/README.md
new file mode 100644
index 000000000..1f6cc877c
--- /dev/null
+++ b/tutorials/dynamic-batching/README.md
@@ -0,0 +1,33 @@
+# TRT Whisper - Dynamic Batching
+
+A tutorial example showing how to deploy TRT Whisper - Dynamic Batching on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Model | [baseten/trtllm-whisper-a10g-large-v2-1](https://huggingface.co/baseten/trtllm-whisper-a10g-large-v2-1) |
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{"audio": "<base64_encoded_wav_audio>"}'
+```
+
+## Configuration highlights
+
+- Base image: `baseten/trtllm-server:r23.12_baseten_v0.9.0.dev2024022000`
+- Model cache: **volume-mounted** for fast cold starts
+- Predict concurrency: **256**
+- System packages: `python3.10-venv, ffmpeg`
diff --git a/tutorials/dynamic-batching/config.yaml b/tutorials/dynamic-batching/config.yaml
new file mode 100644
index 000000000..dfdee9dd6
--- /dev/null
+++ b/tutorials/dynamic-batching/config.yaml
@@ -0,0 +1,35 @@
+description: "A tutorial example showing how to deploy TRT Whisper - Dynamic Batching on Baseten."
+base_image:
+  image: baseten/trtllm-server:r23.12_baseten_v0.9.0.dev2024022000
+  python_executable_path: /usr/bin/python3
+model_metadata:
+  example_model_input: {"audio": "UklGRiQAAABXQVZFZm10IBAAAAABAAEAgD4AAIA+AAABAAgAZGF0YQAAAAA="}
+model_name: TRT Whisper - Dynamic Batching
+python_version: py311
+requirements:
+  - async-batcher==0.2.0
+  - mpi4py==3.1.5
+  - pynvml==11.5.0
+  - huggingface_hub==0.20.3
+  - tiktoken==0.6.0
+  - datasets==2.17.1
+  - kaldialign==0.9
+  - openai-whisper==20250625
+  - soundfile==0.12.1
+model_cache:
+  - repo_id: baseten/trtllm-whisper-a10g-large-v2-1
+    revision: main
+    use_volume: true
+    volume_folder: trtllm-whisper-a10g-large-v2-1
+system_packages:
+  - python3.10-venv
+  - ffmpeg
+resources:
+  accelerator: A10G
+runtime:
+  predict_concurrency: 256
+external_data:
+  - local_data_path: assets/multilingual.tiktoken
+    url: https://raw.githubusercontent.com/openai/whisper/main/whisper/assets/multilingual.tiktoken
+  - local_data_path: assets/mel_filters.npz
+    url: https://raw.githubusercontent.com/openai/whisper/main/whisper/assets/mel_filters.npz
diff --git a/ultravox/model/__init__.py b/tutorials/dynamic-batching/model/__init__.py
similarity index 100%
rename from ultravox/model/__init__.py
rename to tutorials/dynamic-batching/model/__init__.py
diff --git a/07-high-performance-dynamic-batching/model/model.py b/tutorials/dynamic-batching/model/model.py
similarity index 100%
rename from 07-high-performance-dynamic-batching/model/model.py
rename to tutorials/dynamic-batching/model/model.py
diff --git a/vllm/model/__init__.py b/tutorials/dynamic-batching/packages/__init__.py
similarity index 100%
rename from vllm/model/__init__.py
rename to tutorials/dynamic-batching/packages/__init__.py
diff --git a/07-high-performance-dynamic-batching/packages/run.py b/tutorials/dynamic-batching/packages/run.py
similarity index 100%
rename from 07-high-performance-dynamic-batching/packages/run.py
rename to tutorials/dynamic-batching/packages/run.py
diff --git a/07-high-performance-dynamic-batching/packages/tokenizer.py b/tutorials/dynamic-batching/packages/tokenizer.py
similarity index 100%
rename from 07-high-performance-dynamic-batching/packages/tokenizer.py
rename to tutorials/dynamic-batching/packages/tokenizer.py
diff --git a/07-high-performance-dynamic-batching/packages/whisper_utils.py b/tutorials/dynamic-batching/packages/whisper_utils.py
similarity index 100%
rename from 07-high-performance-dynamic-batching/packages/whisper_utils.py
rename to tutorials/dynamic-batching/packages/whisper_utils.py
diff --git a/07-high-performance-dynamic-batching/test.py b/tutorials/dynamic-batching/test.py
similarity index 100%
rename from 07-high-performance-dynamic-batching/test.py
rename to tutorials/dynamic-batching/test.py
diff --git a/tutorials/getting-started-bert/README.md b/tutorials/getting-started-bert/README.md
new file mode 100644
index 000000000..267716909
--- /dev/null
+++ b/tutorials/getting-started-bert/README.md
@@ -0,0 +1,31 @@
+# bert
+
+A tutorial example showing how to deploy bert on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | CPU |
+| Python | py310 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "text": "Hello my name is {MASK}"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/tutorials/getting-started-bert/config.yaml b/tutorials/getting-started-bert/config.yaml
new file mode 100644
index 000000000..0aa3d2768
--- /dev/null
+++ b/tutorials/getting-started-bert/config.yaml
@@ -0,0 +1,62 @@
+# # Step 2: Writing the config.yaml
+#
+# Each Truss has a config.yaml file where we can configure
+# options related to the deployment. It's in this file where
+# we can define requirements, resources, and runtime options like
+# secrets and environment variables
+#
+# ### Basic Options
+#
+# In this section, we can define basic metadata about the model,
+# such as the name, and the Python version to build with.
+description: "A tutorial example showing how to deploy bert on Baseten."
+model_name: bert
+python_version: py310
+model_metadata:
+  repo_id: "google-bert/bert-base-uncased"
+  example_model_input: { "text": "Hello my name is {MASK}" }
+
+
+# ### Set up python requirements
+#
+# In this section, we define any pip requirements that
+# we need to run the model. To run this, we need PyTorch
+# and Tranformers.
+requirements:
+  - torch==2.0.1
+  - transformers==4.33.2
+  - numpy==1.26.4
+
+# ### Configure the resources needed
+#
+# In this section, we can configure resources
+# needed to deploy this model. Here, we have no need for a GPU
+# so we leave the accelerator section blank.
+resources:
+  accelerator: null
+  cpu: '1'
+  memory: 2Gi
+  use_gpu: false
+
+# ### Other config options
+#
+# Truss also has provisions for adding other runtime options
+# packages. In this example, we don't need these, so we leave
+# this empty for now.
+secrets: {}
+system_packages: []
+environment_variables: {}
+external_package_dirs: []
+
+# # Step 3: Deploying & running inference
+#
+# Deploy the model with the following command:
+#
+# ```bash
+# $ truss push
+# ```
+#
+# And then you can performance inference with:
+# ```
+# $ truss predict -d '"Truss is awesome!"'
+# ```
diff --git a/01-getting-started-bert/doc.yaml b/tutorials/getting-started-bert/doc.yaml
similarity index 100%
rename from 01-getting-started-bert/doc.yaml
rename to tutorials/getting-started-bert/doc.yaml
diff --git a/whisper/faster-whisper-small/model/__init__.py b/tutorials/getting-started-bert/model/__init__.py
similarity index 100%
rename from whisper/faster-whisper-small/model/__init__.py
rename to tutorials/getting-started-bert/model/__init__.py
diff --git a/01-getting-started-bert/model/model.py b/tutorials/getting-started-bert/model/model.py
similarity index 100%
rename from 01-getting-started-bert/model/model.py
rename to tutorials/getting-started-bert/model/model.py
diff --git a/tutorials/image-generation/README.md b/tutorials/image-generation/README.md
new file mode 100644
index 000000000..bf9e0f088
--- /dev/null
+++ b/tutorials/image-generation/README.md
@@ -0,0 +1,32 @@
+# Stable Diffusion XL
+
+A tutorial example showing how to deploy Stable Diffusion XL on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "A tree in a field under the night sky",
+  "use_refiner": true
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg, libsm6, libxext6`
diff --git a/tutorials/image-generation/config.yaml b/tutorials/image-generation/config.yaml
new file mode 100644
index 000000000..805ebcbe6
--- /dev/null
+++ b/tutorials/image-generation/config.yaml
@@ -0,0 +1,81 @@
+# # Setting up the config yaml
+#
+# Running SDXL requires a handful of Python libraries, including
+# diffusers, transformers, and others.
+description: "A tutorial example showing how to deploy Stable Diffusion XL on Baseten."
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: stabilityai/stable-diffusion-xl-base-1.0
+  example_model_input:
+    { "prompt": "A tree in a field under the night sky", "use_refiner": true }
+model_name: Stable Diffusion XL
+python_version: py39
+requirements:
+  - transformers==4.34.0
+  - accelerate==0.23.0
+  - safetensors==0.4.0
+  - git+https://github.com/basetenlabs/diffusers.git@9a353290b1497023d4745a719ec02c50f680499a
+  - invisible-watermark>=0.2.0
+  - xformers==0.0.22
+  - numpy==1.26.4
+# ## Configuring resources for SDXL 1.0
+#
+# Note that we need an A10G to run this model.
+resources:
+  accelerator: A10G
+  cpu: 3500m
+  memory: 20Gi
+  use_gpu: true
+secrets: {}
+# ## System Packages
+#
+# Running diffusers requires `ffmpeg` and a couple other system
+# packages.
+system_packages:
+  - ffmpeg
+  - libsm6
+  - libxext6
+# ## Enabling Caching
+#
+# SDXL is a very large model, and downloading it could take up to 10 minutes. This means
+# that the cold start time for this model is long. We can solve that by using our build
+# caching feature. This moves the model download to the build stage of your model--
+# caching the model will take about 10 minutes initially but you will get ~9s cold starts
+# subsequently.
+#
+# To enable caching, add the following to the config:
+# ```yaml
+# model_cache:
+#   - repo_id: madebyollin/sdxl-vae-fp16-fix
+#     revision: main
+#     allow_patterns:
+#       - config.json
+#       - diffusion_pytorch_model.safetensors
+#     use_volume: true
+#     volumne_folder: "sdxl-vae-fp16-fix"
+#   - repo_id: stabilityai/stable-diffusion-xl-base-1.0
+#     allow_patterns:
+#       - "*.json"
+#       - "*.fp16.safetensors"
+#       - sd_xl_base_1.0.safetensors
+#     use_volume: true
+#     volumne_folder: "sdxl-base"
+#   - repo_id: stabilityai/stable-diffusion-xl-refiner-1.0
+#     allow_patterns:
+#       - "*.json"
+#       - "*.fp16.safetensors"
+#       - sd_xl_refiner_1.0.safetensors
+#     use_volume: true
+#     volumne_folder: "sdxl-refiner"
+# ```
+# # Deploy the model
+#
+# Deploy the model like you would other Trusses, with:
+# ```bash
+# $ truss push
+# ```
+# You can then invoke the model with:
+# ```bash
+# $ truss predict -d '{"prompt": "A tree in a field under the night sky", "use_refiner": true}'
+# ```
diff --git a/04-image-generation/doc.yaml b/tutorials/image-generation/doc.yaml
similarity index 100%
rename from 04-image-generation/doc.yaml
rename to tutorials/image-generation/doc.yaml
diff --git a/whisper/faster-whisper-v2/model/__init__.py b/tutorials/image-generation/model/__init__.py
similarity index 100%
rename from whisper/faster-whisper-v2/model/__init__.py
rename to tutorials/image-generation/model/__init__.py
diff --git a/04-image-generation/model/model.py b/tutorials/image-generation/model/model.py
similarity index 100%
rename from 04-image-generation/model/model.py
rename to tutorials/image-generation/model/model.py
diff --git a/tutorials/llm-basics/README.md b/tutorials/llm-basics/README.md
new file mode 100644
index 000000000..8abebb04d
--- /dev/null
+++ b/tutorials/llm-basics/README.md
@@ -0,0 +1,33 @@
+# Mistral 7B
+
+A tutorial example showing how to deploy Mistral 7B on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py311 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "What is the meaning of life?"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/tutorials/llm-basics/config.yaml b/tutorials/llm-basics/config.yaml
new file mode 100644
index 000000000..8c3fb3d37
--- /dev/null
+++ b/tutorials/llm-basics/config.yaml
@@ -0,0 +1,37 @@
+# # Setting up the config.yaml
+#
+# Running Mistral 7B requires a few libraries, such as
+# `torch`, `transformers` and a couple others.
+description: "A tutorial example showing how to deploy Mistral 7B on Baseten."
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+  repo_id: "mistralai/Mistral-7B-v0.1"
+  example_model_input: {"prompt": "What is the meaning of life?"}
+model_name: Mistral 7B
+python_version: py311
+requirements:
+- transformers==4.42.3
+- sentencepiece==0.1.99
+- accelerate==0.23.0
+- torch==2.0.1
+- numpy==1.26.4
+# ## Configure resources for Mistral
+#
+# Note that we need an A10G to run this model.
+resources:
+  accelerator: A10G
+  use_gpu: true
+secrets:
+  hf_access_token: "ENTER HF ACCESS TOKEN HERE"
+system_packages: []
+# # Deploy the model
+#
+# Deploy the model like you would other Trusses, with:
+# ```bash
+# $ truss push
+# ```
+# You can then invoke the model with:
+# ```bash
+# $ truss predict -d '{"inputs": "What is a large language model?"}'
+# ```
diff --git a/02-llm/doc.yaml b/tutorials/llm-basics/doc.yaml
similarity index 100%
rename from 02-llm/doc.yaml
rename to tutorials/llm-basics/doc.yaml
diff --git a/whisper/faster-whisper-v3/model/__init__.py b/tutorials/llm-basics/model/__init__.py
similarity index 100%
rename from whisper/faster-whisper-v3/model/__init__.py
rename to tutorials/llm-basics/model/__init__.py
diff --git a/02-llm/model/model.py b/tutorials/llm-basics/model/model.py
similarity index 100%
rename from 02-llm/model/model.py
rename to tutorials/llm-basics/model/model.py
diff --git a/tutorials/llm-streaming/README.md b/tutorials/llm-streaming/README.md
new file mode 100644
index 000000000..06cffa388
--- /dev/null
+++ b/tutorials/llm-streaming/README.md
@@ -0,0 +1,30 @@
+# LLM with Streaming
+
+A tutorial example showing how to deploy LLM with Streaming on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "prompt": "what is the meaning of life"
+}'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/tutorials/llm-streaming/config.yaml b/tutorials/llm-streaming/config.yaml
new file mode 100644
index 000000000..0a5958c2c
--- /dev/null
+++ b/tutorials/llm-streaming/config.yaml
@@ -0,0 +1,27 @@
+# # Setting up the config.yaml
+#
+# Running Falcon 7B requires torch, transformers,
+# and a few other related libraries.
+description: "A tutorial example showing how to deploy LLM with Streaming on Baseten."
+model_name: "LLM with Streaming"
+model_metadata:
+    repo_id: "tiiuae/falcon-7b-instruct"
+    example_model_input: {"prompt": "what is the meaning of life"}
+requirements:
+- torch==2.0.1
+- peft==0.4.0
+- scipy==1.11.1
+- sentencepiece==0.1.99
+- accelerate==0.21.0
+- bitsandbytes==0.41.1
+- einops==0.6.1
+- transformers==4.31.0
+- numpy==1.26.4
+# ## Configure resources for Falcon
+#
+# Note that we need an A10G to run this model.
+resources:
+  cpu: "3"
+  memory: 14Gi
+  use_gpu: true
+  accelerator: A10G
diff --git a/03-llm-with-streaming/doc.yaml b/tutorials/llm-streaming/doc.yaml
similarity index 100%
rename from 03-llm-with-streaming/doc.yaml
rename to tutorials/llm-streaming/doc.yaml
diff --git a/whisper/whisper-streaming/model/__init__.py b/tutorials/llm-streaming/model/__init__.py
similarity index 100%
rename from whisper/whisper-streaming/model/__init__.py
rename to tutorials/llm-streaming/model/__init__.py
diff --git a/03-llm-with-streaming/model/model.py b/tutorials/llm-streaming/model/model.py
similarity index 100%
rename from 03-llm-with-streaming/model/model.py
rename to tutorials/llm-streaming/model/model.py
diff --git a/tutorials/private-huggingface/README.md b/tutorials/private-huggingface/README.md
new file mode 100644
index 000000000..cdec9e4bc
--- /dev/null
+++ b/tutorials/private-huggingface/README.md
@@ -0,0 +1,31 @@
+# private-model
+
+A tutorial example showing how to deploy private-model on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | CPU |
+| Python | py39 |
+
+## Deploy
+
+> **Note:** This model requires a HuggingFace access token. Set `hf_access_token` in your Baseten secrets before deploying.
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '"It is a [MASK] world"'
+```
+
+## Configuration highlights
+
+- Engine: **Custom (Truss)**
diff --git a/tutorials/private-huggingface/config.yaml b/tutorials/private-huggingface/config.yaml
new file mode 100644
index 000000000..3be8bf95c
--- /dev/null
+++ b/tutorials/private-huggingface/config.yaml
@@ -0,0 +1,40 @@
+# # Setting up the config.yaml
+#
+# The main things that need to be set up in the config are
+# `requirements`, which need to include Hugging Face transformers,
+# and the secrets.
+description: "A tutorial example showing how to deploy private-model on Baseten."
+environment_variables: {}
+model_metadata:
+  repo_id: "baseten/docs-example-gated-model"
+  example_model_input: "It is a [MASK] world"
+model_name: private-model
+python_version: py39
+requirements:
+- torch==2.0.1
+- transformers==4.30.2
+resources:
+  cpu: "1"
+  memory: 2Gi
+  use_gpu: false
+  accelerator: null
+# To make the `hf_access_token` available in the Truss, we need to include
+# it in the config. Setting the value to `null` here means that the value
+# will be set by the Baseten secrets manager.
+secrets:
+  hf_access_token: null
+system_packages: []
+# # Deploying the model
+#
+# An important note for deploying models with secrets is that
+# you must use the `--trusted` flag to give the model access to
+# secrets stored on the remote secrets manager.
+#
+# ```bash
+# $ truss push --trusted
+# ```
+#
+# After the model finishes deploying, you can invoke it with:
+# ```bash
+# $ truss predict -d '"It is a [MASK] world"'
+# ```
diff --git a/09-private-huggingface/doc.yaml b/tutorials/private-huggingface/doc.yaml
similarity index 100%
rename from 09-private-huggingface/doc.yaml
rename to tutorials/private-huggingface/doc.yaml
diff --git a/whisper/whisper-torchserve/model/__init__.py b/tutorials/private-huggingface/model/__init__.py
similarity index 100%
rename from whisper/whisper-torchserve/model/__init__.py
rename to tutorials/private-huggingface/model/__init__.py
diff --git a/09-private-huggingface/model/model.py b/tutorials/private-huggingface/model/model.py
similarity index 100%
rename from 09-private-huggingface/model/model.py
rename to tutorials/private-huggingface/model/model.py
diff --git a/tutorials/speech-to-text/README.md b/tutorials/speech-to-text/README.md
new file mode 100644
index 000000000..b6c3c8b17
--- /dev/null
+++ b/tutorials/speech-to-text/README.md
@@ -0,0 +1,31 @@
+# Whisper
+
+A tutorial example showing how to deploy Whisper on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | A10G |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `ffmpeg`
diff --git a/tutorials/speech-to-text/config.yaml b/tutorials/speech-to-text/config.yaml
new file mode 100644
index 000000000..c202dd6d3
--- /dev/null
+++ b/tutorials/speech-to-text/config.yaml
@@ -0,0 +1,22 @@
+description: "A tutorial example showing how to deploy Whisper on Baseten."
+environment_variables: {}
+model_metadata:
+  repo_id: "openai/whisper-base"
+  example_model_input: {"url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}
+model_name: Whisper
+python_version: py39
+requirements:
+- openai-whisper==20250625
+- torch==2.0.1
+- numpy==1.26.4
+resources:
+  cpu: "4"
+  memory: 16Gi
+  use_gpu: true
+  accelerator: A10G
+secrets: {}
+system_packages:
+- ffmpeg
+external_data:
+  - url: https://baseten-public.s3.us-west-2.amazonaws.com/models/whisper/small.pt
+    local_data_path: models/small.pt
diff --git a/whisper/whisper-truss/model/__init__.py b/tutorials/speech-to-text/model/__init__.py
similarity index 100%
rename from whisper/whisper-truss/model/__init__.py
rename to tutorials/speech-to-text/model/__init__.py
diff --git a/05-speech-to-text/model/model.py b/tutorials/speech-to-text/model/model.py
similarity index 100%
rename from 05-speech-to-text/model/model.py
rename to tutorials/speech-to-text/model/model.py
diff --git a/tutorials/system-packages/README.md b/tutorials/system-packages/README.md
new file mode 100644
index 000000000..71f969772
--- /dev/null
+++ b/tutorials/system-packages/README.md
@@ -0,0 +1,32 @@
+# LayoutLM Document QA
+
+A tutorial example showing how to deploy LayoutLM Document QA on Baseten.
+
+| Property | Value |
+|----------|-------|
+| Task | Tutorial |
+| Engine | Custom (Truss) |
+| GPU | CPU |
+| Python | py39 |
+
+## Deploy
+
+```sh
+truss push
+```
+
+## Invoke
+
+```sh
+curl -X POST https://model-<model_id>.api.baseten.co/predict \
+  -H "Authorization: Api-Key YOUR_BASETEN_API_KEY" \
+  -H "Content-Type: application/json" \
+  -d '{
+  "url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png",
+  "prompt": "What is the invoice number?"
+}'
+```
+
+## Configuration highlights
+
+- System packages: `tesseract-ocr`
diff --git a/tutorials/system-packages/config.yaml b/tutorials/system-packages/config.yaml
new file mode 100644
index 000000000..e45fda563
--- /dev/null
+++ b/tutorials/system-packages/config.yaml
@@ -0,0 +1,38 @@
+# # Setting up the config.yaml file
+#
+# The main items that need to be configured in the config.yaml file are requirements
+# and `system_packages` sections.
+description: "A tutorial example showing how to deploy LayoutLM Document QA on Baseten."
+environment_variables: {}
+external_package_dirs: []
+model_metadata:
+    repo_id: "impira/layoutlm-document-qa"
+    example_model_input: {"url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png", "prompt": "What is the invoice number?"}
+model_name: LayoutLM Document QA
+python_version: py39
+# Specify the versions of the Python requirements that are needed.
+# <Note>Always pin exact versions for your Python dependencies. The ML/AI space moves fast, so you want to have an up-to-date version of each package while also being protected from breaking changes.</Note>
+requirements:
+- Pillow==10.0.0
+- pytesseract==0.3.10
+- torch==2.0.1
+- transformers==4.30.2
+- numpy==1.26.4
+resources:
+  cpu: "4"
+  memory: 16Gi
+  use_gpu: false
+  accelerator: null
+secrets: {}
+# The system_packages section is the other important bit here, you can
+# add any package that's available via `apt` on Debian.
+system_packages:
+- tesseract-ocr
+# # Deploy the model
+# ```bash
+# $ truss push
+# ```
+# You can then invoke the model with:
+# ```
+# $ truss predict -d '{"url": "https://templates.invoicehome.com/invoice-template-us-neat-750px.png", "prompt": "What is the invoice number?"}'
+# ```
diff --git a/10-using-system-packages/doc.yaml b/tutorials/system-packages/doc.yaml
similarity index 100%
rename from 10-using-system-packages/doc.yaml
rename to tutorials/system-packages/doc.yaml
diff --git a/whisper/whisper-v3-truss-base64/model/__init__.py b/tutorials/system-packages/model/__init__.py
similarity index 100%
rename from whisper/whisper-v3-truss-base64/model/__init__.py
rename to tutorials/system-packages/model/__init__.py
diff --git a/10-using-system-packages/model/model.py b/tutorials/system-packages/model/model.py
similarity index 100%
rename from 10-using-system-packages/model/model.py
rename to tutorials/system-packages/model/model.py
diff --git a/ultravox/README.md b/ultravox/README.md
deleted file mode 100644
index 773a2a2c1..000000000
--- a/ultravox/README.md
+++ /dev/null
@@ -1,97 +0,0 @@
-# Ultravox vLLM Truss
-
-This is a [Truss](https://truss.baseten.co/) for Ultravox using the vLLM OpenAI Compatible server. This Truss is designed to provide an efficient and scalable way to serve Ultravox and other models in an OpenAI compatible way using vLLM.
-
-## OpenAI Bridge Compatibility
-
-This Truss is compatible with a *custom* version of our [bridge endpoint for OpenAI ChatCompletion users](https://docs.baseten.co/api-reference/openai). This means you can easily integrate this model into your existing applications that use the OpenAI API format.
-
-```
-client = OpenAI(
-    api_key=os.environ["BASETEN_API_KEY"],
-    base_url=f"https://bridge.baseten.co/{model_id}/direct/v1"
-)
-```
-
-## Truss
-
-Truss is an open-source model serving framework developed by Baseten. It allows you to develop and deploy machine learning models onto Baseten (and other platforms like [AWS](https://truss.baseten.co/deploy/aws) or [GCP](https://truss.baseten.co/deploy/gcp)). Using Truss, you can develop a GPU model using [live-reload](https://baseten.co/blog/technical-deep-dive-truss-live-reload), package models and their associated code, create Docker containers, and deploy on Baseten.
-
-## Deployment
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples.git
-cd ultravox
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `ultravox` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## vLLM OpenAI Compatible Server
-
-This Truss demonstrates how to start [vLLM's OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html). The Truss is primarily used to start the server and then route requests to it. It currently supports ChatCompletions only.
-
-### Passing startup arguments to the server
-
-In the config any key-values under `model_metadata: arguments:` will be passed to the vLLM OpenAI-compatible server at startup.
-
-### Base Image
-
-You can use any vLLM compatible base image.
-
-## API Documentation
-
-The API follows the OpenAI ChatCompletion format. You can interact with the model using the standard ChatCompletion interface.
-
-Example usage:
-
-```python
-from openai import OpenAI
-
-client = OpenAI(
-    api_key="YOUR-API-KEY",
-    base_url="https://bridge.baseten.co/MODEL-ID/v1"
-)
-
-response = client.chat.completions.create(
-    model="fixie-ai/ultravox-v0.2",
-    messages=[{
-            "role": "user",
-            "content": [
-                {"type": "text", "text": "Summarize the following: <|audio|>"},
-                {"type": "image_url", "image_url": {"url": f"data:audio/wav;base64,{base64_wav}"}}
-            ]
-        }]
-    stream=True
-)
-
-for chunk in response:
-    print(chunk.choices[0].delta)
-```
-
-## Future Improvements
-
-We are actively working on enhancing this Truss. Some planned improvements include:
-
-- Adding support for distributed serving with Ray (https://docs.vllm.ai/en/latest/serving/distributed_serving.html)
-- Implementing model caching for improved performance
-
-Stay tuned for updates!
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/ultravox/config.yaml b/ultravox/config.yaml
deleted file mode 100644
index b56bf0c5c..000000000
--- a/ultravox/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-base_image:
-  image: vshulman/vllm-openai-fixie:latest
-  python_executable_path: /usr/bin/python3
-model_metadata:
-  arguments:
-    model: fixie-ai/ultravox-v0.2
-    audio_token_id: 128002
-environment_variables: {}
-external_package_dirs: []
-model_name: Ultravox v0.2
-python_version: py310
-runtime:
-  predict_concurrency: 512
-requirements:
-  - httpx
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets: {}
-system_packages:
-- python3.10-venv
diff --git a/vllm/README.md b/vllm/README.md
deleted file mode 100644
index 17fc3f903..000000000
--- a/vllm/README.md
+++ /dev/null
@@ -1,213 +0,0 @@
-# vLLM Truss to deploy chat completion model
-
-## What is this Truss example doing
-
-This is a general purpose [Truss](https://truss.baseten.co/) that can deploy an asynchronous vLLM engine([AsyncLLMEngine](https://docs.vllm.ai/en/latest/dev/engine/async_llm_engine.html#asyncllmengine)) of any customized configuration with [all compatible chat completion models](https://docs.vllm.ai/en/latest/models/supported_models.html). We create this example to give you the most codeless experience, so you can configure all vLLM engine parameters in `config.yaml`, without making code changes in `model.py` for most of the use cases.
-
-## Configure your Truss by modifying the config.yaml
-
-### Basic options using 1 GPU
-
-Here is the minimum config file you will need to deploy a model using vLLM on 1 GPU.
-The only parameters you need to touch are:
-- `model_name`
-- `repo_id`
-- `accelerator`
-
-```
-model_name: "Llama 3.1 8B Instruct VLLM"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-  openai_compatible: true
-  vllm_config: null
-requirements:
-  - vllm==0.5.4
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
-```
-
-### Basic options using multiple GPUs
-
-If your model needs more than 1 GPU to run using tensor parallel, you will need to change `accelerator`, and to set `tensor_parallel_size` and `distributed_executor_backend` accordingly.
-
-```
-model_name: "Llama 3.1 8B Instruct VLLM"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-  openai_compatible: false
-  vllm_config:
-    tensor_parallel_size: 4
-    max_model_len: 4096
-    distributed_executor_backend: mp
-requirements:
-  - vllm==0.5.4
-resources:
-  accelerator: A10G:4
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
-```
-
-### Use vLLM's OpenAI compatible server
-
-To use vLLM in [OpenAI compatible server](https://docs.vllm.ai/en/latest/serving/openai_compatible_server.html) mode, simply set `openai_compatible: true` under `model_metadata`.
-
-### Customize vLLM engine parameters
-
-For advanced users who want to override [vLLM engine arguments](https://docs.vllm.ai/en/latest/models/engine_args.html), you can add all arguments to `vllm_config` under `model_metadata`.
-
-#### Example 1: using model quantization
-
-```
-model_name: Mistral 7B v2 vLLM AWQ - T4
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  repo_id: TheBloke/Mistral-7B-Instruct-v0.2-AWQ
-  vllm_config:
-    quantization: "awq"
-    dtype: "float16"
-    max_model_len: 8000
-    max_num_seqs: 8
-python_version: py310
-requirements:
-  - vllm==0.5.4
-resources:
-  accelerator: T4
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages: []
-runtime:
-  predict_concurrency: 128
-```
-
-#### Example 2: using customized vLLM image
-
-You can even override with your own customized vLLM docker image to work with models that are not supported yet by vanilla vLLM.
-
-```
-model_name: Ultravox v0.2
-base_image:
-  image: vshulman/vllm-openai-fixie:latest
-  python_executable_path: /usr/bin/python3
-model_metadata:
-  repo_id: fixie-ai/ultravox-v0.2
-  vllm_config:
-    audio_token_id: 128002
-environment_variables: {}
-external_package_dirs: []
-python_version: py310
-runtime:
-  predict_concurrency: 512
-requirements:
-  - httpx
-resources:
-  accelerator: A100
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages:
-- python3.10-venv
-```
-
-## Deploy your Truss
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-3. With `vllm` as your working directory, you can deploy the model with:
-
-    ```sh
-    truss push --trusted
-    ```
-
-    Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Call your model
-
-Once your deployment is up, there are [many ways](https://docs.baseten.co/invoke/quickstart) to call your model.
-
-### curl command
-
-#### If you are NOT using OpenAI compatible server
-
-```
-curl -X POST https://model-<YOUR_MODEL_ID>.api.baseten.co/development/predict \
-     -H "Authorization: Api-Key $BASETEN_API_KEY" \
-     -d '{"prompt": "what is the meaning of life"}'
-```
-
-
-#### If you are using OpenAI compatible server
-
-```
-curl -X POST "https://model-<YOUR_MODEL_ID>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {BASETEN_API_KEY}' \
-     -d '{
-           "messages": [{"role": "user", "content": "What even is AGI?"}],
-           "max_tokens": 256
-         }'
-```
-
-To access [production metrics](https://docs.vllm.ai/en/latest/serving/metrics.html) reported by OpenAI compatible server, simply add `metrics: true` to the request.
-
-```
-curl -X POST "https://model-<YOUR_MODEL_ID>.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {BASETEN_API_KEY}' \
-     -d '{
-           "metrics": true
-         }'
-```
-
-### OpenAI SDK (if you are using OpenAI compatible server)
-
-```
-from openai import OpenAI
-import os
-
-model_id = "abcd1234" # Replace with your model ID
-deployment_id = "4321cbda" # [Optional] Replace with your deployment ID
-
-client = OpenAI(
-    api_key=os.environ["BASETEN_API_KEY"],
-    base_url=f"https://bridge.baseten.co/{model_id}/v1/direct"
-)
-
-response = client.chat.completions.create(
-  model="meta-llama/Llama-3.1-8B-Instruct",
-  messages=[
-    {"role": "user", "content": "Who won the world series in 2020?"},
-    {"role": "assistant", "content": "The Los Angeles Dodgers won the World Series in 2020."},
-    {"role": "user", "content": "Where was it played?"}
-  ],
-  extra_body={
-    "baseten": {
-      "model_id": model_id,
-      "deployment_id": deployment_id
-    }
-  }
-)
-print(response.choices[0].message.content)
-
-```
-
-For more information, see [API reference](https://docs.baseten.co/api-reference/openai).
-
-## Support
-
-If you have any questions or need assistance, please open an issue in this repository or contact our support team.
diff --git a/vllm/config.yaml b/vllm/config.yaml
deleted file mode 100644
index bd98b7e40..000000000
--- a/vllm/config.yaml
+++ /dev/null
@@ -1,19 +0,0 @@
-model_name: "Llama 3.1 8B Instruct VLLM openai compatible"
-python_version: py311
-model_metadata:
-  example_model_input: {"prompt": "what is the meaning of life"}
-  repo_id: meta-llama/Llama-3.1-8B-Instruct
-  openai_compatible: true
-  vllm_config:
-    tensor_parallel_size: 1
-    max_model_len: 4096
-    enable_prefix_caching: true
-requirements:
-  - vllm==0.5.4
-resources:
-  accelerator: A100
-  use_gpu: true
-runtime:
-  predict_concurrency: 128
-secrets:
-  hf_access_token: null
diff --git a/vllm/model/helper.py b/vllm/model/helper.py
deleted file mode 100644
index 69067e11a..000000000
--- a/vllm/model/helper.py
+++ /dev/null
@@ -1,57 +0,0 @@
-import asyncio
-import logging
-import os
-import threading
-
-import httpx
-
-logger = logging.getLogger(__name__)
-
-DEFAULT_HEALTH_CHECK_INTERVAL = 5  # seconds
-
-
-async def monitor_vllm_server_health(vllm_server_url, health_check_interval):
-    assert vllm_server_url is not None, "vllm_server_url must not be None"
-    try:
-        async with httpx.AsyncClient() as client:
-            while True:
-                response = await client.get(f"{vllm_server_url}/health")
-                if response.status_code != 200:
-                    raise RuntimeError("vLLM is unhealthy")
-                await asyncio.sleep(health_check_interval)
-    except Exception as e:
-        logging.error(
-            f"vLLM has gone into an unhealthy state due to error: {e}, restarting service now..."
-        )
-        os._exit(1)
-
-
-async def monitor_vllm_engine_health(vllm_engine, health_check_interval):
-    assert vllm_engine is not None, "vllm_engine must not be None"
-    try:
-        while True:
-            await vllm_engine.check_health()
-            await asyncio.sleep(health_check_interval)
-    except Exception as e:
-        logging.error(
-            f"vLLM has gone into an unhealthy state due to error: {e}, restarting service now..."
-        )
-        os._exit(1)
-
-
-def run_background_vllm_health_check(
-    use_openai_compatible_server=False,
-    health_check_interval=DEFAULT_HEALTH_CHECK_INTERVAL,
-    vllm_engine=None,
-    vllm_server_url=None,
-):
-    logger.info("Starting background health check loop")
-    loop = asyncio.new_event_loop()
-    if use_openai_compatible_server:
-        loop.create_task(
-            monitor_vllm_server_health(vllm_server_url, health_check_interval)
-        )
-    else:
-        loop.create_task(monitor_vllm_engine_health(vllm_engine, health_check_interval))
-    thread = threading.Thread(target=loop.run_forever, daemon=True)
-    thread.start()
diff --git a/vllm/model/model.py b/vllm/model/model.py
deleted file mode 100644
index c6b5c1e16..000000000
--- a/vllm/model/model.py
+++ /dev/null
@@ -1,220 +0,0 @@
-import logging
-import os
-import subprocess
-import time
-import uuid
-
-import httpx
-from model.helper import run_background_vllm_health_check
-from transformers import AutoTokenizer
-from vllm.engine.arg_utils import AsyncEngineArgs
-from vllm.engine.async_llm_engine import AsyncLLMEngine
-
-from vllm import SamplingParams
-
-os.environ["TOKENIZERS_PARALLELISM"] = "true"
-
-logger = logging.getLogger(__name__)
-
-
-class Model:
-    # TODO: better way to detect model server startup failure than using `MAX_FAILED_SECONDS`
-    MAX_FAILED_SECONDS = 1500  # 25 minutes; the reason this would take this long is mostly if we download a large model
-    HEALTH_CHECK_INTERVAL = 5  # seconds
-
-    def __init__(self, **kwargs):
-        self._config = kwargs["config"]
-        self.model_id = None
-        self.llm_engine = None
-        self.model_args = None
-        self.hf_secret_token = kwargs["secrets"].get("hf_access_token", None)
-        self.openai_compatible = self._config["model_metadata"].get(
-            "openai_compatible", False
-        )
-        self.vllm_base_url = None
-        os.environ["HF_TOKEN"] = self.hf_secret_token
-
-    def load(self):
-        self._model_metadata = self._config["model_metadata"]
-        self._model_repo_id = self._model_metadata["repo_id"]
-        self._vllm_config = self._model_metadata["vllm_config"]
-        if self._vllm_config is None:
-            self._vllm_config = {}
-        logger.info(f"main model: {self._model_repo_id}")
-        logger.info(f"vllm config: {self._vllm_config}")
-        if self.openai_compatible:
-            self._client = httpx.AsyncClient(timeout=None)
-            command = ["vllm", "serve", self._model_repo_id]
-            for key, value in self._vllm_config.items():
-                if value is True:
-                    command.append(f"--{key.replace('_', '-')}")
-                elif value is False:
-                    continue
-                else:
-                    command.append(f"--{key.replace('_', '-')}")
-                    command.append(str(value))
-
-            logger.info(
-                f"Starting openai compatible vLLM server with command: {command}"
-            )
-
-            self._vllm_process = subprocess.Popen(
-                command, stdout=subprocess.PIPE, stderr=subprocess.PIPE, text=True
-            )
-
-            # Wait for 10 seconds and check if command fails
-            time.sleep(10)
-
-            if self._vllm_process.poll() is None:
-                logger.info("Command to start vLLM server ran successfully")
-            else:
-                stdout, stderr = self._vllm_process.communicate()
-                if self._vllm_process.returncode != 0:
-                    logger.error(f"Command failed with error: {stderr}")
-                    raise RuntimeError(
-                        f"Command failed with code {self._vllm_process.returncode}: {stderr}"
-                    )
-
-            if self._vllm_config and "port" in self._vllm_config:
-                self._vllm_port = self._vllm_config["port"]
-            else:
-                self._vllm_port = 8000
-
-            self.vllm_base_url = f"http://localhost:{self._vllm_port}"
-
-            # Polling to check if the server is up
-            server_up = False
-            start_time = time.time()
-            while time.time() - start_time < self.MAX_FAILED_SECONDS:
-                try:
-                    response = httpx.get(f"{self.vllm_base_url}/health")
-                    logger.info(f"Checking server health: {response.status_code}")
-                    if response.status_code == 200:
-                        server_up = True
-                        break
-                except httpx.RequestError as e:
-                    seconds_passed = int(time.time() - start_time)
-                    if seconds_passed % 10 == 0:
-                        logger.info(
-                            f"Server is starting for {seconds_passed} seconds: {e}"
-                        )
-                    time.sleep(1)  # Wait for 1 second before retrying
-
-            if not server_up:
-                raise RuntimeError(
-                    "Server failed to start within the maximum allowed time."
-                )
-        else:
-            try:
-                result = subprocess.run(
-                    ["nvidia-smi"], capture_output=True, text=True, check=True
-                )
-                logger.info(result.stdout)
-            except subprocess.CalledProcessError as e:
-                logger.error(f"Command failed with code {e.returncode}: {e.stderr}")
-
-            self.model_args = AsyncEngineArgs(
-                model=self._model_repo_id, **self._vllm_config
-            )
-            self.llm_engine = AsyncLLMEngine.from_engine_args(
-                engine_args=self.model_args
-            )
-            self.tokenizer = AutoTokenizer.from_pretrained(self._model_repo_id)
-
-            try:
-                result = subprocess.run(
-                    ["nvidia-smi"], capture_output=True, text=True, check=True
-                )
-                logger.info(result.stdout)
-            except subprocess.CalledProcessError as e:
-                logger.error(f"Command failed with code {e.returncode}: {e.stderr}")
-        try:
-            run_background_vllm_health_check(
-                self.openai_compatible,
-                self.HEALTH_CHECK_INTERVAL,
-                self.llm_engine,
-                self.vllm_base_url,
-            )
-        except Exception as e:
-            raise RuntimeError(f"Failed to start background health check: {e}")
-
-    async def predict(self, model_input):
-        if "messages" not in model_input and "prompt" not in model_input:
-            raise ValueError("Prompt or messages must be provided")
-
-        stream = model_input.get("stream", False)
-        if self.openai_compatible:
-            # if the key metrics: true is present, let's return the vLLM /metrics endpoint
-            if model_input.get("metrics", False):
-                response = await self._client.get(f"{self.vllm_base_url}/metrics")
-                return response.text
-
-            # convenience for Baseten bridge
-            if "model" not in model_input and self._model_repo_id:
-                logger.info(
-                    f"model_input missing model due to Baseten bridge, using {self._model_repo_id}"
-                )
-                model_input["model"] = self._model_repo_id
-
-            if stream:
-
-                async def generator():
-                    async with self._client.stream(
-                        "POST",
-                        f"{self.vllm_base_url}/v1/chat/completions",
-                        json=model_input,
-                    ) as response:
-                        async for chunk in response.aiter_bytes():
-                            if chunk:
-                                yield chunk
-
-                return generator()
-            else:
-                response = await self._client.post(
-                    f"{self.vllm_base_url}/v1/chat/completions",
-                    json=model_input,
-                )
-                return response.json()
-        else:
-            # SamplingParams does not take/use argument 'model'
-            if "model" in model_input:
-                model_input.pop("model")
-            if "prompt" in model_input:
-                prompt = model_input.pop("prompt")
-                sampling_params = SamplingParams(**model_input)
-                idx = str(uuid.uuid4().hex)
-                messages = [
-                    {"role": "user", "content": prompt},
-                ]
-                # templatize the input to the model
-                input = self.tokenizer.apply_chat_template(
-                    messages, tokenize=False, add_generation_prompt=True
-                )
-            elif "messages" in model_input:
-                messages = model_input.pop("messages")
-                sampling_params = SamplingParams(**model_input)
-                idx = str(uuid.uuid4().hex)
-                # templatize the input to the model
-                input = self.tokenizer.apply_chat_template(
-                    messages,
-                    tokenize=False,
-                )
-            logger.info(f"Using SamplingParams: {sampling_params}")
-            # since we accept any valid vllm sampling parameters, we can just pass it through
-            vllm_generator = self.llm_engine.generate(input, sampling_params, idx)
-
-            async def generator():
-                full_text = ""
-                async for output in vllm_generator:
-                    text = output.outputs[0].text
-                    delta = text[len(full_text) :]
-                    full_text = text
-                    yield delta
-
-            if stream:
-                return generator()
-            else:
-                full_text = ""
-                async for delta in generator():
-                    full_text += delta
-                return {"text": full_text}
diff --git a/whisper/faster-whisper-small/README.md b/whisper/faster-whisper-small/README.md
deleted file mode 100644
index 1678d03d3..000000000
--- a/whisper/faster-whisper-small/README.md
+++ /dev/null
@@ -1,69 +0,0 @@
-# Whisper Small
-
-This is a [Truss](https://truss.baseten.co/) for Whisper Small. This README will walk you through how to deploy this Truss on Baseten to get your own instance of Whisper Small up and running.
-
-## Faster Whisper Small Implementation
-
-This implementation of Whisper Small uses [Faster Whisper](https://github.com/SYSTRAN/faster-whisper/tree/master), which is up to 4x faster than openai/whisper for the same accuracy while using less memory.
-
-## Deployment
-
-You can deploy this model in just a few clicks from our [model library](), or deploy the Truss, which we'll describe here.
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd whisper/faster-whisper-small
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `whisper/faster-whisper-small` as your working directory, you can deploy the model with:
-
-```sh
-truss push --trusted
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-Once your Truss is deployed, you can start using Whisper Small for inference! Navigate to the [Baseten UI](https://app.baseten.co/models) to watch the model build and deploy and invoke it via the REST API for tasks like transcription.
-
-## Example usage
-
-Here's a sample script which loads a sample .mp3 file for transcription with Whisper Small:
-
-```python
-import requests
-import os
-
-# Replace the empty string with your model id below
-model_id = ""
-
-# We recommend storing your API key as an environment variable
-baseten_api_key = os.environ["BASETEN_API_KEY"]
-
-data = {
-  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
-}
-
-# Call model endpoint
-res = requests.post(
-    f"https://model-{model_id}.api.baseten.co/production/predict",
-    headers={"Authorization": f"Api-Key {baseten_api_key}"},
-    json=data
-)
-
-# Print the output of the model
-print(res.json())
-```
-
-Here is the model output:
-```json
-{'language': 'en', 'language_probability': 0.99072265625, 'duration': 11.52, 'segments': [{'text': ' Four score and seven years ago, our fathers brought forth upon this continent a new nation', 'start': 0.0, 'end': 6.5200000000000005}, {'text': ' conceived in liberty and dedicated to the proposition that all men are created equal.', 'start': 6.5200000000000005, 'end': 11.0}]}
-```
diff --git a/whisper/whisper-streaming/README.md b/whisper/whisper-streaming/README.md
deleted file mode 100644
index aaa6139cd..000000000
--- a/whisper/whisper-streaming/README.md
+++ /dev/null
@@ -1,62 +0,0 @@
-# Whisper Streaming Truss
-
-This is a truss for [Whisper Streaming](https://github.com/ufal/whisper_streaming). This truss allows you to stream the whisper transcription chunks as they get generated instead of waiting for the entire transcription to finish.
-
-This whisper streaming model supports the following whisper models: `tiny.en`, `tiny`, `base.en`, `base`, `small.en`, `small`, `medium.en`, `medium`, `large-v1`, `large-v2`,`large-v3`, `large`.
-You can specify the model you want to use inside the `config.yaml` file under the key `whisper_model` in the `model_metadata` section.
-
-```yaml
-model_metadata:
-  whisper_model: medium
-model_name: Whisper Streaming
-```
-
-## Deploying Whisper Streaming
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `whisper/whisper-streaming` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Whisper API documentation
-
-The model accepts the following inputs:
-
-- __audio__(required): The input audio file in the form of a base64 string.
-- __chunk_size__(optional): The number of seconds of audio that will get transcribed in each chunk.
-
-
-
-## Invoking the model
-
-```python
-import requests
-import base64
-
-def wav_to_base64(file_path):
-    with open(file_path, "rb") as wav_file:
-        binary_data = wav_file.read()
-        base64_data = base64.b64encode(binary_data)
-        base64_string = base64_data.decode("utf-8")
-        return base64_string
-
-resp = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers = {"Authorization": "Api-Key BASETEN-API-KEY"},
-    json={"audio": wav_to_base64("/path/to/wav/input_audio_file.wav")},
-    stream=True
-)
-
-for content in resp.iter_content():
-    print(content.decode("utf-8"), end="", flush=True)
-```
diff --git a/whisper/whisper-streaming/config.yaml b/whisper/whisper-streaming/config.yaml
deleted file mode 100644
index 9ebae7e33..000000000
--- a/whisper/whisper-streaming/config.yaml
+++ /dev/null
@@ -1,17 +0,0 @@
-base_image:
-  image: baseten/truss-server-base:3.10-gpu-v0.4.9
-  python_executable_path: /usr/bin/python3
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  whisper_model: medium
-model_name: Whisper Streaming
-python_version: py310
-requirements: []
-requirements_file: ./requirements.txt
-resources:
-  accelerator: T4
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/whisper/whisper-torchserve/README.md b/whisper/whisper-torchserve/README.md
deleted file mode 100644
index e6f685b8b..000000000
--- a/whisper/whisper-torchserve/README.md
+++ /dev/null
@@ -1,82 +0,0 @@
-# Whisper Torchserve
-
-This truss allows you to run a whisper model using [torchserve](https://pytorch.org/serve/) as the backend on truss.
-
-
-## Deployment
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `whisper/whisper-torchserve` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Model Inputs
-
-The model takes in one input:
-- __audio__: An audio file as a base64 string
-
-## Few thing to note
-Torchserve requires a compiled `.mar` file in order to serve the model. Here is a [README](https://github.com/pytorch/serve/blob/master/model-archiver/README.md) providing a brief explanation for generating this file. Once the `.mar` file is generated it needs to get placed in the `data/model_store` directory. Also in the `data/` directory is a configuration file for torchserve called `config.properties`. That file looks something like this:
-
-```
-inference_address=http://0.0.0.0:8888
-batch_size=4
-ipex_enable=true
-async_logging=true
-
-models={\
-  "whisper_base": {\
-    "1.0": {\
-        "defaultVersion": true,\
-        "marName": "whisper_base.mar",\
-        "minWorkers": 1,\
-        "maxWorkers": 2,\
-        "batchSize": 4,\
-        "maxBatchDelay": 500,\
-        "responseTimeout": 24\
-    }\
-  }\
-}
-```
-
-Here you can specify the `batchSize` as well as the name of your mar file using `marName`. When torchserve starts, it will looks for the mar file inside the `data/model_store` directory with the `marName` defined above.
-
-## Invoking the model
-
-Here is an example in Python:
-
-```python
-import requests
-import base64
-
-def wav_to_base64(file_path):
-    with open(file_path, "rb") as wav_file:
-        binary_data = wav_file.read()
-        base64_data = base64.b64encode(binary_data)
-        base64_string = base64_data.decode("utf-8")
-        return base64_string
-
-resp = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers={"Authorization": "Api-Key BASETEN-API-KEY"},
-    json={"audio": wav_to_base64("/path/to/audio-file/60-sec.wav")},
-)
-
-print(resp.json())
-```
-
-Here is a sample output:
-
-```json
-{"output": "Let me make it clear. His conduct is unacceptable. He's unfit. And be careful of what you're gonna get. He doesn't care for the American people. It's Donald Trump first. This is what I want people to understand. These people have... I mean, she has no idea what the hell the names of those provinces are, but she wants to send our sons and daughters and our troops and our military equipment to go fight it. Look at the blank expression. She doesn't know the names of the provinces. You do this at every debate. You say, no, don't interrupt me. I didn't interrupt you."}
-```
diff --git a/whisper/whisper-torchserve/config.yaml b/whisper/whisper-torchserve/config.yaml
deleted file mode 100644
index c3ff12df3..000000000
--- a/whisper/whisper-torchserve/config.yaml
+++ /dev/null
@@ -1,24 +0,0 @@
-environment_variables: {}
-external_package_dirs: []
-model_metadata: {}
-model_name: Whisper Torchserve
-python_version: py310
-requirements:
-  - torch==2.1.0
-  - torchserve==0.9.0
-  - ffmpeg-python==0.2.0
-  - transformers==4.37.2
-  - nvgpu==0.10.0
-  - httpx==0.27.0
-resources:
-  accelerator: T4
-  use_gpu: true
-model_cache:
-  - repo_id: htrivedi99/whisper-torchserve
-    use_volume: false
-secrets: {}
-system_packages:
-  - ffmpeg
-  - openjdk-11-jdk
-runtime:
-  predict_concurrency: 128
diff --git a/whisper/whisper-torchserve/data/config.properties b/whisper/whisper-torchserve/data/config.properties
deleted file mode 100644
index 29f59e4b2..000000000
--- a/whisper/whisper-torchserve/data/config.properties
+++ /dev/null
@@ -1,21 +0,0 @@
-inference_address=http://0.0.0.0:8888
-batch_size=16
-ipex_enable=true
-async_logging=true
-
-models={\
-  "whisper_base": {\
-    "1.0": {\
-        "defaultVersion": true,\
-        "marName": "whisper_base.mar",\
-        "minWorkers": 1,\
-        "maxWorkers": 4,\
-        "batchSize": 16,\
-        "maxBatchDelay": 250,\
-        "responseTimeout": 120\
-    }\
-  }\
-}
-
-# maxBatchDelay is the amount of time to wait for the batch size to fill up. Default is 250 ms.
-# default_workers_per_model=2
diff --git a/whisper/whisper-torchserve/model/model.py b/whisper/whisper-torchserve/model/model.py
deleted file mode 100644
index 4789c5116..000000000
--- a/whisper/whisper-torchserve/model/model.py
+++ /dev/null
@@ -1,69 +0,0 @@
-import base64
-import multiprocessing
-import os
-import subprocess
-from typing import Dict
-
-import httpx
-import requests
-from huggingface_hub import snapshot_download
-
-TORCHSERVE_ENDPOINT = "http://0.0.0.0:8888/predictions/whisper_base"
-TORCHSERVE_HEALTH_ENDPOINT = "http://0.0.0.0:8888/ping"
-
-
-class Model:
-    def __init__(self, **kwargs):
-        self._data_dir = kwargs["data_dir"]
-        self._model = None
-        self.torchserver_ready = False
-
-    def start_torchserver(self):
-        subprocess.run(
-            [
-                "torchserve",
-                "--start",
-                "--model-store",
-                f"{self._data_dir}/model_store",
-                "--models",
-                "whisper_base.mar",
-                "--foreground",
-                "--no-config-snapshots",
-                "--ts-config",
-                f"{self._data_dir}/config.properties",
-            ],
-            check=True,
-        )
-
-    def load(self):
-        snapshot_download(
-            "htrivedi99/whisper-torchserve",
-            local_dir=os.path.join(self._data_dir, "model_store"),
-            max_workers=4,
-        )
-        logging.info("⚡️ Weights Downloaded Successfully!")
-
-        process = multiprocessing.Process(target=self.start_torchserver)
-        process.start()
-
-        # Need to wait for the torchserve server to start up
-        while not self.torchserver_ready:
-            try:
-                res = requests.get(TORCHSERVE_HEALTH_ENDPOINT)
-                if res.status_code == 200:
-                    self.torchserver_ready = True
-                    logging.info("🔥Torchserve is ready!")
-            except Exception:
-                logging.info("⏳Torchserve is loading...")
-                time.sleep(5)
-
-    async def predict(self, request: Dict):
-        audio_base64 = request.get("audio")
-        audio_bytes = base64.b64decode(audio_base64)
-
-        async with httpx.AsyncClient() as client:
-            res = await client.post(
-                TORCHSERVE_ENDPOINT, files={"data": (None, audio_bytes)}, timeout=120
-            )
-            transcription = res.text
-        return {"output": transcription}
diff --git a/whisper/whisper-truss/README.md b/whisper/whisper-truss/README.md
deleted file mode 100644
index d5246bc41..000000000
--- a/whisper/whisper-truss/README.md
+++ /dev/null
@@ -1,79 +0,0 @@
-# Whisper Truss
-
-[Whisper](https://github.com/openai/whisper) is a speech-to-text model by [OpenAI](https://openai.com/blog/whisper/) that transcribes audio in dozens of languages with remarkable accuracy. It is open-source under the [MIT license](https://github.com/openai/whisper/blob/main/LICENSE) and hosted on Baseten as a pre-trained model. Read the [Whisper model card](https://github.com/openai/whisper/blob/main/model-card.md) for more details.
-
-Whisper's leap in transcription quality unlocks tons of compelling use cases, including:
-
-- Moderating audio content
-- Auditing call center logs
-- Automatically generating video subtitles
-- Improving podcast SEO with transcripts
-
-## Deploying Whisper
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `whisper-truss` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking Whisper
-
-Once the model is deployed, you can invoke it with:
-
-```sh
-truss predict -d '{"url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}'
-```
-
-You can also invoke your Whisper deployment via its REST API endpoint:
-
-```bash
-curl -X POST "https://app.baseten.co/models/{MODEL_ID}/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{"url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}'
-```
-
-### Whisper API documentation
-
-#### Input
-
-This deployment of Whisper takes input as a JSON dictionary with the key `url` corresponding to a string of a URL pointing at an MP3 file. For example:
-
-```json
-{
-  "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
-}
-```
-
-#### Output
-
-The model returns a fairly lengthy dictionary. For most uses, you'll be interested in the key `language` which specifies the detected language of the audio and `text` which contains the full transcription.
-
-```json
-{
-  "language": "english",
-  "segments": [
-    {
-      "start": 0,
-      "end": 6.5200000000000005,
-      "text": " Four score and seven years ago, our fathers brought forth upon this continent a new nation"
-    },
-    {
-      "start": 6.52,
-      "end": 21.6,
-      "text": " conceived in liberty and dedicated to the proposition that all men are created equal."
-    }
-  ],
-  "text": " Four score and seven years ago, our fathers brought forth upon this continent..."
-}
-```
diff --git a/whisper/whisper-truss/config.yaml b/whisper/whisper-truss/config.yaml
deleted file mode 100644
index a42e4ae6d..000000000
--- a/whisper/whisper-truss/config.yaml
+++ /dev/null
@@ -1,27 +0,0 @@
-description: Transcribe audio files across multiple languages.
-environment_variables: {}
-external_data:
-- local_data_path: models/small.pt
-  url: https://baseten-public.s3.us-west-2.amazonaws.com/models/whisper/small.pt
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/openai.png
-  cover_image_url: https://cdn.baseten.co/production/static/whisper.png
-  example_model_input:
-    url: https://cdn.baseten.co/docs/production/Gettysburg.mp3
-  pretty_name: Whisper
-  tags:
-  - speech-recognition
-model_name: Whisper
-python_version: py39
-requirements:
-- openai-whisper==20250625
-- torch==2.0.1
-resources:
-  accelerator: A10G
-  cpu: '4'
-  memory: 16Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/whisper/whisper-truss/data/.gitkeep b/whisper/whisper-truss/data/.gitkeep
deleted file mode 100644
index e69de29bb..000000000
diff --git a/whisper/whisper-v3-truss-base64/README.md b/whisper/whisper-v3-truss-base64/README.md
deleted file mode 100644
index ddd2b34c2..000000000
--- a/whisper/whisper-v3-truss-base64/README.md
+++ /dev/null
@@ -1,75 +0,0 @@
-# Whisper v3 Truss
-
-[Whisper](https://github.com/openai/whisper) is a open-source speech-to-text model by [OpenAI](https://openai.com/blog/whisper/) that transcribes audio in dozens of languages with remarkable accuracy.
-
-Whisper v3 has the same architecture as the previous model but has made the following improvements:
-- The large-v3 model is trained on 1 million hours of weakly labeled audio and 4 million hours of pseudolabeled audio collected using large-v2
-- This version has a 10-20% lower rate of error compared to the previous version when benchmarked on `Common Voice 15` and `Fleurs` dataset
-
-
-## Deploying Whisper
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `whisper-v3-truss` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking Whisper
-
-Once the model is deployed, you can invoke it with:
-
-```python
-import requests
-import base64
-
-
-audio_base64 = base64.b64encode(open("Gettysburg.mp3", "rb").read()).decode('utf-8')
-
-resp = requests.post(
-    "https://model-{MODEL_ID}.api.baseten.co/development/predict",
-    headers={"Authorization": "Api-Key $BASETEN_API_KEY"},
-    json={'audio': audio_base64},
-)
-
-print(resp.content)
-```
-
-### Whisper API documentation
-
-#### Input
-
-This deployment of Whisper takes input as a JSON dictionary with the key `audio` corresponding to a base64 encoded audio file. Here is an example input JSON:
-
-```json
-{
-    "audio": "YmG4DdeoS0HEV..."
-}
-```
-
-#### Output
-
-The model returns a fairly lengthy dictionary. For most uses, you'll be interested in the key `language` which specifies the detected language of the audio and `text` which contains the full transcription.
-
-```json
-{
-    "language": "english",
-    "segments": [
-        {
-        "start": 0,
-        "end": 11.52,
-        "text": "Four score and seven years ago our fathers brought forth upon this continent a new nation conceived in liberty and dedicated to the proposition that all men are created equal."
-        }
-    ],
-    "text": "Four score and seven years ago our fathers brought forth upon this continent a new nation conceived in liberty and dedicated to the proposition that all men are created equal."
-}
-```
diff --git a/whisper/whisper-v3-truss-base64/config.yaml b/whisper/whisper-v3-truss-base64/config.yaml
deleted file mode 100644
index 839fb552d..000000000
--- a/whisper/whisper-v3-truss-base64/config.yaml
+++ /dev/null
@@ -1,24 +0,0 @@
-description: Transcribe audio files across multiple languages.
-environment_variables: {}
-external_data:
-- local_data_path: weights/large-v3.pt
-  url: https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/openai.png
-  cover_image_url: https://cdn.baseten.co/production/static/whisper.png
-  example_model_input:
-    url: https://cdn.baseten.co/docs/production/Gettysburg.mp3
-model_name: Whisper V3 Base64 Input
-python_version: py310
-requirements:
-- torch==2.0.1
-- openai-whisper==20250625
-resources:
-  accelerator: T4
-  cpu: '3'
-  memory: 16Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/whisper/whisper-v3-truss/README.md b/whisper/whisper-v3-truss/README.md
deleted file mode 100644
index 3894fb926..000000000
--- a/whisper/whisper-v3-truss/README.md
+++ /dev/null
@@ -1,72 +0,0 @@
-# Whisper v3 Truss
-
-[Whisper](https://github.com/openai/whisper) is a open-source speech-to-text model by [OpenAI](https://openai.com/blog/whisper/) that transcribes audio in dozens of languages with remarkable accuracy.
-
-Whisper v3 has the same architecture as the previous model but has made the following improvements:
-- The large-v3 model is trained on 1 million hours of weakly labeled audio and 4 million hours of pseudolabeled audio collected using large-v2
-- This version has a 10-20% lower rate of error compared to the previous version when benchmarked on `Common Voice 15` and `Fleurs` dataset
-
-
-## Deploying Whisper
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `whisper-v3-truss` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking Whisper
-
-Once the model is deployed, you can invoke it with:
-
-```sh
-truss predict -d '{"url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}'
-```
-
-You can also invoke your Whisper deployment via its REST API endpoint:
-
-```bash
-curl -X POST "https://model-{MODEL_ID}.api.baseten.co/development/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{"url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}'
-```
-
-### Whisper API documentation
-
-#### Input
-
-This deployment of Whisper takes input as a JSON dictionary with the key `url` corresponding to a string of a URL pointing at an MP3 file. For example:
-
-```json
-{
-    "url": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"
-}
-```
-
-#### Output
-
-The model returns a fairly lengthy dictionary. For most uses, you'll be interested in the key `language` which specifies the detected language of the audio and `text` which contains the full transcription.
-
-```json
-{
-    "language": "english",
-    "segments": [
-        {
-        "start": 0,
-        "end": 11.52,
-        "text": "Four score and seven years ago our fathers brought forth upon this continent a new nation conceived in liberty and dedicated to the proposition that all men are created equal."
-        }
-    ],
-    "text": "Four score and seven years ago our fathers brought forth upon this continent a new nation conceived in liberty and dedicated to the proposition that all men are created equal."
-}
-```
diff --git a/whisper/whisper-v3-truss/config.yaml b/whisper/whisper-v3-truss/config.yaml
deleted file mode 100644
index 5baae8918..000000000
--- a/whisper/whisper-v3-truss/config.yaml
+++ /dev/null
@@ -1,25 +0,0 @@
-description: Transcribe audio files across multiple languages.
-environment_variables: {}
-external_data:
-- local_data_path: weights/large-v3.pt
-  url: https://openaipublic.azureedge.net/main/whisper/models/e5b1a55b89c1367dacf97e3e19bfd829a01529dbfdeefa8caeb59b3f1b81dadb/large-v3.pt
-external_package_dirs: []
-model_metadata:
-  avatar_url: https://cdn.baseten.co/production/static/openai.png
-  cover_image_url: https://cdn.baseten.co/production/static/whisper.png
-  example_model_input:
-    url: https://cdn.baseten.co/docs/production/Gettysburg.mp3
-model_name: Whisper V3
-python_version: py310
-requirements:
-- torch==2.4.1
-- openai-whisper==20250625
-- ffmpeg-python==0.2.0
-resources:
-  accelerator: A10G
-  cpu: '3'
-  memory: 16Gi
-  use_gpu: true
-secrets: {}
-system_packages:
-- ffmpeg
diff --git a/whisper/whisper-v3-truss/model/__init__.py b/whisper/whisper-v3-truss/model/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/whisper/whisperx-truss/README.md b/whisper/whisperx-truss/README.md
deleted file mode 100644
index 1582c867f..000000000
--- a/whisper/whisperx-truss/README.md
+++ /dev/null
@@ -1,84 +0,0 @@
-# WhisperX Truss
-
-[WhisperX](https://github.com/m-bain/whisperX) is a model built on top of Whisper that provides fast speech recognition with word-level timestamps.
-
-## Getting Access To The Model
-The base WhisperX model does not need any special requirements, but if you want to enable speaker diarization there are a couple of models that you will need to get permission from.
-1. Go to the [speaker-diarization model](https://huggingface.co/pyannote/speaker-diarization-3.0) and fill out the required info to gain access to the model.
-
-2. Go to the [segmentation model](https://huggingface.co/pyannote/segmentation-3.0) and go through the same process.
-
-Once you have access to both of those models, make sure you have your hugging face access token on hand as you will need it to run this truss.
-
-1. Create a [HuggingFace access token](https://huggingface.co/settings/tokens)
-2. Set it as a [secret in your Baseten account](https://app.baseten.co/settings/secrets) with the name `hf_access_token`
-
-## Deploying WhisperX
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-Next, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd whisperx-truss
-```
-
-With `whisperx-truss` as your working directory, you can deploy the model with:
-
-```
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Hardware notes
-
-This whisperX model comes in various sizes: ["small", "medium", "large-v2"].
-The large-v2 model can be easily run on the T4 GPU.
-
-## API route: `predict`
-
-The `predict` route is the primary method used for audio transcription.
-
-- __audio_file__: An MP3 audio file. This file must be accessible over the internet as files from local storage are not accessible.
-
-## Invoking WhisperX
-
-Once the model is deployed, you can invoke it with:
-
-```sh
-truss predict -d '{"audio_file": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}'
-```
-
-You can also invoke your Whisper deployment via its REST API endpoint:
-
-```bash
-curl -X POST "https://app.baseten.co/models/{MODEL_ID}/predict" \
-     -H "Content-Type: application/json" \
-     -H 'Authorization: Api-Key {YOUR_API_KEY}' \
-     -d '{"audio_file": "https://cdn.baseten.co/docs/production/Gettysburg.mp3"}'
-```
-
-## Output
-
-The model returns a dictionary which contains the start and end timestamps along with the transcript of what is said in between each timestamp.
-
-```json
-{
-    "model_output":
-    [
-        {
-            "end": 10.742,
-            "start": 0.765,
-            "text": "Four score and seven years ago, our fathers brought forth upon this continent, a new nation conceived in liberty and dedicated to the proposition that all men are created equal.",
-            "speaker": "SPEAKER_00"
-        }
-    ]
-}
-```
diff --git a/whisper/whisperx-truss/config.yaml b/whisper/whisperx-truss/config.yaml
deleted file mode 100644
index d662f9201..000000000
--- a/whisper/whisperx-truss/config.yaml
+++ /dev/null
@@ -1,35 +0,0 @@
-base_image:
-  image: runpod/pytorch:2.1.1-py3.10-cuda12.1.1-devel-ubuntu22.04
-  python_executable_path: /usr/bin/python
-environment_variables: {}
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    audio_file: https://cdn.baseten.co/docs/production/Gettysburg.mp3
-model_name: whisperX
-python_version: py310
-requirements:
-- --extra-index-url https://download.pytorch.org/whl/cu121
-- git+https://github.com/m-bain/whisperx.git@734084cdf6f624bc33ed9f0cfcaa82840707ba6f
-- torch==2.2.0
-- torchaudio==2.2.0
-- transformers==4.48.3
-- torchvision==0.17.0
-- ffmpeg-python==0.2.0
-- faster-whisper==1.1.0
-- pandas==2.2.3
-- nltk==3.9.1
-- setuptools==68.0.0
-- ctranslate2==4.4.0
-- pydub==0.25.1
-resources:
-  accelerator: L4
-  cpu: '1'
-  memory: 4Gi
-  use_gpu: true
-secrets:
-  hf_access_token: null
-system_packages:
-- ffmpeg
-- libsm6
-- libxext6
diff --git a/whisper/whisperx-truss/model/__init__.py b/whisper/whisperx-truss/model/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/xtts-streaming/README.md b/xtts-streaming/README.md
deleted file mode 100644
index 795a7d817..000000000
--- a/xtts-streaming/README.md
+++ /dev/null
@@ -1,115 +0,0 @@
-# XTTS Streaming
-
-This repository packages [TTS](https://github.com/coqui-ai/TTS) as a [Truss](https://truss.baseten.co/) but with streaming.
-
-TTS is a generative audio model for text-to-speech generation. This model takes in text and a speaker's voice as input and converts the text to speech in the voice of the speaker.
-
-## Deploying XTTS
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd xtts-streaming
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `xtts-v2-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking the model
-
-Here are the following inputs for the model:
-1. `text`: The text that needs to be converted into speech
-2. `language`: Language for the text
-3. `chunk_size`: Integer size of each chunk being streamed
-
-Here are two examples of streaming the audio. This first example write all of the streamed chunks to an audio file.
-
-```python
-import wave
-import requests
-
-channels = 1  # mono=1, stereo=2
-sampwidth = 2  # Sample width in bytes, typical values: 2 for 16-bit audio, 1 for 8-bit audio
-framerate = 24000  # Sampling rate, in samples per second (Hz)
-
-
-resp = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers={"Authorization": "Api-Key BASETEN-API-KEY"},
-    json={"text": "Kurt watched the incoming Pelicans. The blocky jet-powered craft were so distant they were only specks against the setting sun. He hit the magnification on his faceplate and saw lines of fire tracing their reentry vectors. They would touch down in three minutes."},
-    stream=True
-)
-
-with wave.open("dat2-wav.wav", 'wb') as wav_file:
-    wav_file.setnchannels(channels)
-    wav_file.setsampwidth(sampwidth)
-    wav_file.setframerate(framerate)
-
-    # Iterate through streamed content and write audio chunks directly
-    for chunk in resp.iter_content(chunk_size=None):  # Use server's chunk size
-        if chunk:
-            wav_file.writeframes(chunk)
-```
-
-If you want to stream the audio directly as it gets generated here is another option:
-
-```python
-import pyaudio
-
-FORMAT = pyaudio.paInt16  # Audio format (e.g., 16-bit PCM)
-CHANNELS = 1              # Number of audio channels
-RATE = 24000              # Sample rate
-
-# Initialize PyAudio
-p = pyaudio.PyAudio()
-
-# Open a stream for audio playback
-stream = p.open(format=p.get_format_from_width(2), channels=CHANNELS, rate=RATE, output=True)
-
-# Make a streaming HTTP request to the server
-original_text = "Kurt watched the incoming Pelicans. The blocky jet-powered craft were so distant they were only specks against the setting sun. He hit the magnification on his faceplate and saw lines of fire tracing their reentry vectors. They would touch down in three minutes."
-
-
-resp = requests.post(
-    "https://model-<model-id>.api.baseten.co/development/predict",
-    headers={"Authorization": "Api-Key BASETEN-API-KEY"},
-    json={"text": "Kurt watched the incoming Pelicans. The blocky jet-powered craft were so distant they were only specks against the setting sun. He hit the magnification on his faceplate and saw lines of fire tracing their reentry vectors. They would touch down in three minutes."},
-    stream=True
-)
-
-# Create a buffer to hold multiple chunks
-buffer = b''
-buffer_size_threshold = 2**20
-
-# Stream and play the audio data as it's received
-for chunk in resp.iter_content(chunk_size=4096):
-    if chunk:
-        buffer += chunk
-        if len(buffer) >= buffer_size_threshold:
-            print(f"Writing buffer of size: {len(buffer)}")
-            stream.write(buffer)
-            buffer = b''  # Clear the buffer
-        # stream.write(chunk)
-
-if buffer:
-    print(f"Writing final buffer of size: {len(buffer)}")
-    stream.write(buffer)
-
-# Close and terminate the stream and PyAudio
-stream.stop_stream()
-stream.close()
-p.terminate()
-```
diff --git a/xtts-streaming/config.yaml b/xtts-streaming/config.yaml
deleted file mode 100644
index 95acd151f..000000000
--- a/xtts-streaming/config.yaml
+++ /dev/null
@@ -1,14 +0,0 @@
-base_image:
-  image: htrivedi05/xtts-streaming
-  python_executable_path: /opt/conda/bin/python
-environment_variables:
-  COQUI_TOS_AGREED: '1'
-external_package_dirs: []
-model_metadata: {}
-model_name: XTTS Streaming - High Performance
-resources:
-  accelerator: H100
-  cpu: '3'
-  memory: 10Gi
-  use_gpu: true
-secrets: {}
diff --git a/xtts-streaming/model/__init__.py b/xtts-streaming/model/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/xtts-v2-truss/README.md b/xtts-v2-truss/README.md
deleted file mode 100644
index dae2c255e..000000000
--- a/xtts-v2-truss/README.md
+++ /dev/null
@@ -1,73 +0,0 @@
-# XTTS V2 Truss
-
-This repository packages [TTS](https://github.com/coqui-ai/TTS) as a [Truss](https://truss.baseten.co/).
-
-TTS is a generative audio model for text-to-speech generation. This model takes in text and a speaker's voice as input and converts the text to speech in the voice of the speaker.
-
-## Deploying XTTS
-
-First, clone this repository:
-
-```sh
-git clone https://github.com/basetenlabs/truss-examples/
-cd xtts-v2-truss
-```
-
-Before deployment:
-
-1. Make sure you have a [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
-2. Install the latest version of Truss: `pip install --upgrade truss`
-
-With `xtts-v2-truss` as your working directory, you can deploy the model with:
-
-```sh
-truss push
-```
-
-Paste your Baseten API key if prompted.
-
-For more information, see [Truss documentation](https://truss.baseten.co).
-
-## Invoking the model
-
-Here are the following inputs for the model:
-1. `text`: The text that needs to be converted into speech
-2. `speaker_voice`: A short audio clip of a voice in the format of a base64 string
-3. `language`: Abbreviation of the supported languages for TTS
-
-Here is an example of how to invoke this model:
-
-```python
-
-import base64
-import sys
-
-def wav_to_base64(file_path):
-  with open(file_path, "rb") as wav_file:
-    binary_data = wav_file.read()
-    base64_data = base64.b64encode(binary_data)
-    base64_string = base64_data.decode("utf-8")
-    return base64_string
-
-def base64_to_wav(base64_string, output_file_path):
-  binary_data = base64.b64decode(base64_string)
-  with open(output_file_path, "wb") as wav_file:
-    wav_file.write(binary_data)
-
-voice = wav_to_base64("/path/to/wav/file/samuel_jackson_voice.wav")
-text = "Listen up, people. Life's a wild ride, and sometimes you gotta grab it by the horns and steer it where you want to go. You can't just sit around waiting for things to happen – you gotta make 'em happen. Yeah, it's gonna get tough, but that's when you dig deep, find that inner badass, and come out swinging. Remember, success ain't handed to you on a silver platter; you gotta snatch it like it owes you money. So, lace up your boots, square those shoulders, and let the world know that you're here to play, and you're playing for keeps"
-data = {"text": text, "speaker_voice": voice, "language": "en"}
-res = requests.post("https://model-<model-id>.api.baseten.co/development/predict", headers=headers, json=data)
-res = res.json()
-output = base64_to_wav(res.get('output'), "test_output.wav")
-```
-
-The output of the model is a base64 string as well, so you can convert it to a wav file using the `base64_to_wav` function.
-
-Here is the input file for Samuel Jackson's voice:
-![speaker voice](https://github.com/htrivedi99/truss-examples/assets/15642666/6ae79c53-f63a-4e0a-b6fc-d397fee6162e)
-
-Here is the output file for the text to speech:
-![output](https://github.com/htrivedi99/truss-examples/assets/15642666/2a2f9a80-860d-4ef1-a530-7892a8bd874e)
-
-In this README, the audio files are in the MP4 format because github does not allow the WAV format. However, for the truss itself using the WAV format is recommended.
diff --git a/xtts-v2-truss/config.yaml b/xtts-v2-truss/config.yaml
deleted file mode 100644
index 7ac58ab63..000000000
--- a/xtts-v2-truss/config.yaml
+++ /dev/null
@@ -1,21 +0,0 @@
-environment_variables:
-  COQUI_TOS_AGREED: "1"
-external_package_dirs: []
-model_metadata:
-  example_model_input:
-    language: en
-    speaker_voice: Claribel Dervla
-    text: Kurt watched the incoming Pelicans. The blocky jet-powered craft were so distant they were only specks against the setting sun. He hit the magnification on his faceplate and saw lines of fire tracing their reentry vectors. They would touch down in three minutes.
-  tags:
-  - text-to-speech
-model_name: XTTS V2
-python_version: py310
-requirements:
-  - git+https://github.com/htrivedi99/TTS.git
-resources:
-  accelerator: T4
-  cpu: '3'
-  memory: 10Gi
-  use_gpu: true
-secrets: {}
-system_packages: []
diff --git a/xtts-v2-truss/model/__init__.py b/xtts-v2-truss/model/__init__.py
deleted file mode 100644
index e69de29bb..000000000
diff --git a/z-ai/glm-4-5-air-fp8/README.md b/z-ai/glm-4-5-air-fp8/README.md
deleted file mode 100644
index cd46514a1..000000000
--- a/z-ai/glm-4-5-air-fp8/README.md
+++ /dev/null
@@ -1,42 +0,0 @@
-# GLM 4.5 Air FP8 (SGLang)
-
-This example shows how to deploy [GLM 4.5 Air in FP8](https://huggingface.co/zai-org/GLM-4.5-Air-FP8) using SGLang on Baseten. This model requires 4xH100 to access the full context window.
-
-## Deploying the model
-
-This model can be deployed to Baseten using Truss:
-
-```
-pip install --upgrade truss
-truss push --publish z-ai/glm-4-5-air-fp8
-```
-
-## Calling the model
-
-This model is OpenAI compatible and can be called using the OpenAI client.
-
-```python
-import os
-from openai import OpenAI
-
-# https://model-XXXXXXX.api.baseten.co/environments/production/sync/v1
-model_url = ""
-
-client = OpenAI(
-    base_url=model_url,
-    api_key=os.environ.get("BASETEN_API_KEY"),
-)
-
-stream = client.chat.completions.create(
-    model="baseten",
-    messages=[
-        {"role": "system", "content": "You are a helpful assistant."},
-        {"role": "user", "content": "Write FizzBuzz."}
-    ],
-    stream=True,
-)
-
-for chunk in stream:
-    if chunk.choices[0].delta.content is not None:
-        print(chunk.choices[0].delta.content, end="")
-```
diff --git a/z-ai/glm-4-5-air-fp8/config.yaml b/z-ai/glm-4-5-air-fp8/config.yaml
deleted file mode 100644
index c1408600b..000000000
--- a/z-ai/glm-4-5-air-fp8/config.yaml
+++ /dev/null
@@ -1,27 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write FizzBuzz in Python"
-    stream: true
-    model: "baseten-sglang"
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: GLM 4.5 Air FP8
-base_image:
-  image: lmsysorg/sglang:v0.4.9.post6-cu126
-docker_server:
-  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-Air-FP8 --tp-size 4   --tool-call-parser glm45    --reasoning-parser glm45    --speculative-algorithm EAGLE   --speculative-num-steps 3    --speculative-eagle-topk 1    --speculative-num-draft-tokens 4   --mem-fraction-static 0.7   --disable-shared-experts-fusion   --served-model-name glm-4.5-air-fp8   --host 0.0.0.0   --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:4
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/z-ai/glm-4-5-fp8/README.md b/z-ai/glm-4-5-fp8/README.md
deleted file mode 100644
index 1b3e3a4ab..000000000
--- a/z-ai/glm-4-5-fp8/README.md
+++ /dev/null
@@ -1,42 +0,0 @@
-# GLM 4.5 FP8 (SGLang)
-
-This example shows how to deploy [GLM 4.5 in FP8](https://huggingface.co/zai-org/GLM-4.5-FP8) using SGLang on Baseten. This model requires 8xH100 to run and H200 or B200 to access the full context window.
-
-## Deploying the model
-
-This model can be deployed to Baseten using Truss:
-
-```
-pip install --upgrade truss
-truss push --publish z-ai/glm-4-5-fp8
-```
-
-## Calling the model
-
-This model is OpenAI compatible and can be called using the OpenAI client.
-
-```python
-import os
-from openai import OpenAI
-
-# https://model-XXXXXXX.api.baseten.co/environments/production/sync/v1
-model_url = ""
-
-client = OpenAI(
-    base_url=model_url,
-    api_key=os.environ.get("BASETEN_API_KEY"),
-)
-
-stream = client.chat.completions.create(
-    model="baseten",
-    messages=[
-        {"role": "system", "content": "You are a helpful assistant."},
-        {"role": "user", "content": "Write FizzBuzz."}
-    ],
-    stream=True,
-)
-
-for chunk in stream:
-    if chunk.choices[0].delta.content is not None:
-        print(chunk.choices[0].delta.content, end="")
-```
diff --git a/z-ai/glm-4-5-fp8/config.yaml b/z-ai/glm-4-5-fp8/config.yaml
deleted file mode 100644
index f2c1177dc..000000000
--- a/z-ai/glm-4-5-fp8/config.yaml
+++ /dev/null
@@ -1,27 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write FizzBuzz in Python"
-    stream: true
-    model: "baseten-sglang"
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: GLM 4.5 FP8
-base_image:
-  image: lmsysorg/sglang:v0.4.9.post6-cu126
-docker_server:
-  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.5-FP8 --tp-size 4   --tool-call-parser glm45    --reasoning-parser glm45    --speculative-algorithm EAGLE   --speculative-num-steps 3    --speculative-eagle-topk 1    --speculative-num-draft-tokens 4   --mem-fraction-static 0.7   --disable-shared-experts-fusion   --served-model-name glm-4.5-fp8   --host 0.0.0.0   --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:8
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/z-ai/glm-4-6-fp8/config.yaml b/z-ai/glm-4-6-fp8/config.yaml
deleted file mode 100644
index 140d2e125..000000000
--- a/z-ai/glm-4-6-fp8/config.yaml
+++ /dev/null
@@ -1,31 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write FizzBuzz in Python"
-    stream: true
-    model: "zai-org/GLM-4.6"
-    max_tokens: 4096
-    temperature: 0.6
-  tags:
-    - openai-compatible
-model_name: GLM-4.6-FP8
-base_image:
-  image: lmsysorg/sglang:v0.5.3rc1-cu126
-# build_commands:
-#   - pip install --upgrade pip
-#   - pip uninstall -y sglang
-#   - git clone https://github.com/sgl-project/sglang.git && cd sglang && git checkout 229d2b95f19573ece9c1c5d6b357df9874e04f59 && pip install -e "python[all]"
-docker_server:
-  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.6-FP8 --tp-size 8 --tool-call-parser glm45 --reasoning-parser glm45 --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --mem-fraction-static 0.9 --disable-shared-experts-fusion --served-model-name zai-org/GLM-4.6 --host 0.0.0.0 --port 8000"
-  readiness_endpoint: /health
-  liveness_endpoint: /health
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:8
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/z-ai/glm_4_5_v/config.yaml b/z-ai/glm_4_5_v/config.yaml
deleted file mode 100644
index de546b83c..000000000
--- a/z-ai/glm_4_5_v/config.yaml
+++ /dev/null
@@ -1,28 +0,0 @@
-model_metadata:
-  example_model_input: # Loads sample request into Baseten playground
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "Write FizzBuzz in Python"
-    stream: false
-    model: "zai-org/GLM-4.5V-FP8"
-    top_p: 0.95
-    extra_body: { "top_k": 40 }
-    max_tokens: 2048
-  tags:
-    - openai-compatible
-model_name: GLM-4.5V-FP8
-base_image:
-  image: lmsysorg/sglang:v0.5.4.post1-cu129-amd64
-docker_server:
-  start_command: sh -c "python3 -m sglang.launch_server --model-path zai-org/GLM-4.5V-FP8 --tp-size 8 --tool-call-parser glm45 --reasoning-parser glm45 --mem-fraction-static 0.9 --served-model-name zai-org/GLM-4.5V --host 0.0.0.0 --port 8000 --enable-cache-report"
-  readiness_endpoint: /health_generate
-  liveness_endpoint: /health_generate
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:8
-  use_gpu: true
-runtime:
-  predict_concurrency: 32
diff --git a/z-ai/glm_4_7_flash/config.yaml b/z-ai/glm_4_7_flash/config.yaml
deleted file mode 100644
index 0997ba76b..000000000
--- a/z-ai/glm_4_7_flash/config.yaml
+++ /dev/null
@@ -1,33 +0,0 @@
-model_metadata:
-  example_model_input:
-    messages:
-      - role: system
-        content: "You are a helpful assistant."
-      - role: user
-        content: "What is the meaning of life?"
-    stream: true
-    model: zai-org/GLM-4.7-Flash
-    max_tokens: 32768
-    temperature: 0.7
-  tags:
-    - openai-compatible
-base_image:
-  image: lmsysorg/sglang:nightly-dev-20260122-e6ccb294
-
-build_commands:
-  - pip uninstall -y transformers
-  - pip install git+https://github.com/huggingface/transformers.git@76732b4e7120808ff989edbd16401f61fa6a0afa
-
-docker_server:
-  start_command: python3 -m sglang.launch_server --model-path zai-org/GLM-4.7-Flash --tp-size 2 --tool-call-parser glm47 --reasoning-parser glm45 --speculative-algorithm EAGLE --speculative-num-steps 3 --speculative-eagle-topk 1 --speculative-num-draft-tokens 4 --mem-fraction-static 0.8 --served-model-name zai-org/GLM-4.7-Flash --host 0.0.0.0 --port 8000
-  readiness_endpoint: /health_generate
-  liveness_endpoint: /health_generate
-  predict_endpoint: /v1/chat/completions
-  server_port: 8000
-resources:
-  accelerator: H100:2
-  use_gpu: true
-runtime:
-  predict_concurrency : 32
-
-model_name: GLM 4.7 Flash