Add tier-1 BEI-torch example templates

11-embeddings-reranker-classification-torch/BEI-torch-nvidia-Nemotron-3-Embed-1B-bf16/README.md

@@ -0,0 +1,34 @@
+# BEI-torch with nvidia/Nemotron-3-Embed-1B-BF16
+
+Deployment for [Nemotron-3-Embed-1B-BF16](https://huggingface.co/nvidia/Nemotron-3-Embed-1B-BF16)
+on the BEI-torch backend. Native embedding dim 2048; MRL heads at 512 / 1024 / 2048.
+
+Fits on an L4 (24 GB) at bf16.
+
+## Prerequisites
+
+1. [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
+2. Latest truss: `pip install --upgrade truss`
+3. Set `hf_access_token` as a [Baseten secret](https://app.baseten.co/settings/secrets).
+
+## Known issue: stock `nvidia/Nemotron-3-Embed-1B-BF16` fails to load
+
+Same three JSON patches apply as for the 8B model — see
+[`../BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16/README.md`](../BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16/README.md).
+Point `checkpoint_repository.repo` at a fork with those patches applied.
+
+## Deploy
+
+```sh
+git clone https://github.com/basetenlabs/truss-examples.git
+cd 11-embeddings-reranker-classification-torch/BEI-torch-nvidia-Nemotron-3-Embed-1B-bf16
+truss push --publish
+```
+
+## Call
+
+```bash
+curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
+     -H "Authorization: Api-Key $BASETEN_API_KEY" \
+     -d '{"input": "text string", "model": "model"}'
+```

11-embeddings-reranker-classification-torch/BEI-torch-nvidia-Nemotron-3-Embed-1B-bf16/config.yaml

@@ -0,0 +1,23 @@
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-torch-nvidia-Nemotron-3-Embed-1B-bf16-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 15Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_torch
+    checkpoint_repository:
+      repo: nvidia/Nemotron-3-Embed-1B-BF16
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    quantization_type: no_quant
+  runtime:
+    webserver_default_route: /v1/embeddings

11-embeddings-reranker-classification-torch/BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16/README.md

@@ -0,0 +1,50 @@
+# BEI-torch with nvidia/Nemotron-3-Embed-8B-BF16
+
+Deployment for [Nemotron-3-Embed-8B-BF16](https://huggingface.co/nvidia/Nemotron-3-Embed-8B-BF16)
+on the BEI-torch backend (text-embeddings-router + vLLM). Native embedding dim 4096; MRL heads
+at 512 / 1024 / 2048 / 4096.
+
+## Prerequisites
+
+1. [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
+2. Latest truss: `pip install --upgrade truss`
+3. Set `hf_access_token` as a [Baseten secret](https://app.baseten.co/settings/secrets).
+
+## Known issue: stock `nvidia/Nemotron-3-Embed-8B-BF16` fails to load
+
+Three JSON files in the stock repo trip HF `transformers`; fork the repo and apply these
+patches, then set `checkpoint_repository.repo` to your fork:
+
+1. In `tokenizer_config.json`: remove the `tokenizer_class: "TokenizersBackend"` key.
+2. In `tokenizer_config.json`: remove `extra_special_tokens` (or convert the list to a dict).
+3. In `config.json`: change `model_type: "llama_bidirec"` → `"llama"` and set
+   `architectures: ["LlamaBidirectionalModel"]`.
+
+The BEI-torch arch registry maps `LlamaBidirectionalModel` to a bidirectional-attention
+Llama variant; `model_type: "llama"` only exists so the transformers config parser accepts
+the file. This workaround will be unnecessary once the BEI-torch `llama_bidirec` serde
+alias and tokenizer-config fallback land upstream.
+
+## Deploy
+
+```sh
+git clone https://github.com/basetenlabs/truss-examples.git
+cd 11-embeddings-reranker-classification-torch/BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16
+truss push --publish
+```
+
+## Call
+
+```bash
+curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
+     -H "Authorization: Api-Key $BASETEN_API_KEY" \
+     -d '{"input": "text string", "model": "model"}'
+```
+
+Matryoshka embeddings (native 4096; also supports 512, 1024, 2048):
+
+```bash
+curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/v1/embeddings \
+     -H "Authorization: Api-Key $BASETEN_API_KEY" \
+     -d '{"input": "text string", "model": "model", "dimensions": 1024}'
+```

11-embeddings-reranker-classification-torch/BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16/config.yaml

@@ -0,0 +1,23 @@
+model_metadata:
+  example_model_input:
+    encoding_format: float
+    input: text string
+    model: model
+model_name: BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16-truss-example
+python_version: py39
+resources:
+  accelerator: H100
+  cpu: '1'
+  memory: 30Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_torch
+    checkpoint_repository:
+      repo: nvidia/Nemotron-3-Embed-8B-BF16
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    quantization_type: no_quant
+  runtime:
+    webserver_default_route: /v1/embeddings

11-embeddings-reranker-classification-torch/BEI-torch-nvidia-llama-nemotron-rerank-1b-v2/README.md

@@ -0,0 +1,39 @@
+# BEI-torch with nvidia/llama-nemotron-rerank-1b-v2
+
+Deployment for [llama-nemotron-rerank-1b-v2](https://huggingface.co/nvidia/llama-nemotron-rerank-1b-v2)
+on the BEI-torch backend. Reranker analog to Nemotron-3-Embed; scores a `(query, documents[])`
+pair via `/rerank`.
+
+Fits on an L4 (24 GB) at bf16.
+
+## Prerequisites
+
+1. [Baseten account](https://app.baseten.co/signup) and [API key](https://app.baseten.co/settings/account/api_keys).
+2. Latest truss: `pip install --upgrade truss`
+3. Set `hf_access_token` as a [Baseten secret](https://app.baseten.co/settings/secrets).
+
+## Known issue: stock `nvidia/llama-nemotron-rerank-1b-v2` fails to load
+
+Same three JSON patches apply as for the Nemotron-3-Embed models — see
+[`../BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16/README.md`](../BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16/README.md).
+Point `checkpoint_repository.repo` at a fork with those patches applied.
+
+## Deploy
+
+```sh
+git clone https://github.com/basetenlabs/truss-examples.git
+cd 11-embeddings-reranker-classification-torch/BEI-torch-nvidia-llama-nemotron-rerank-1b-v2
+truss push --publish
+```
+
+## Call
+
+```bash
+curl -X POST https://model-xxxxxx.api.baseten.co/environments/production/sync/rerank \
+     -H "Authorization: Api-Key $BASETEN_API_KEY" \
+     -d '{
+       "query": "what is baseten?",
+       "documents": ["Baseten is a model deployment platform.", "Bananas are yellow."],
+       "model": "model"
+     }'
+```

11-embeddings-reranker-classification-torch/BEI-torch-nvidia-llama-nemotron-rerank-1b-v2/config.yaml

@@ -0,0 +1,25 @@
+model_metadata:
+  example_model_input:
+    query: what is baseten?
+    documents:
+      - Baseten is a model deployment platform.
+      - Bananas are yellow.
+    model: model
+model_name: BEI-torch-nvidia-llama-nemotron-rerank-1b-v2-truss-example
+python_version: py39
+resources:
+  accelerator: L4
+  cpu: '1'
+  memory: 15Gi
+  use_gpu: true
+trt_llm:
+  build:
+    base_model: encoder_torch
+    checkpoint_repository:
+      repo: nvidia/llama-nemotron-rerank-1b-v2
+      revision: main
+      source: HF
+    max_num_tokens: 32768
+    quantization_type: no_quant
+  runtime:
+    webserver_default_route: /rerank

11-embeddings-reranker-classification-torch/README.md

@@ -0,0 +1,27 @@
+# BEI-torch templates
+
+Reference truss configs for the BEI-torch backend (`base_model: encoder_torch`) — the torch/vLLM
+runtime used for embedding, reranker, and classification models whose architecture the
+TRT-LLM `encoder`/`encoder_bert` paths cannot compile (LlamaBidirectional / Ministral3-Embed,
+Gemma2Embedding, jina-v3, etc.).
+
+Templates in this directory are hand-written pending integration with
+`../11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py`.
+When picking `encoder_torch` vs `encoder`/`encoder_bert`, use:
+
+- **`encoder_bert`** — BERT-family (BERT, ModernBERT, XLM-R, jina-v2). Fastest cold-start.
+- **`encoder`** — causal-arch models with FP8 support on H100/B200 (Qwen3-Embedding, SFR-Embedding).
+- **`encoder_torch`** — everything else. LlamaBidirectional (Nemotron-3-Embed), Gemma2Embedding,
+  jina-v3, custom architectures BEI-torch supports but TRT-LLM does not.
+
+## Known issue with stock NVIDIA Nemotron-3-Embed repos
+
+The stock `nvidia/Nemotron-3-Embed-*` repos ship JSON that HF `transformers` rejects on load:
+
+1. `tokenizer_config.json` has `tokenizer_class: "TokenizersBackend"` (not a real class).
+2. `tokenizer_config.json` has `extra_special_tokens` as a list; transformers expects a dict.
+3. `config.json` has `model_type: "llama_bidirec"` which is not registered with transformers.
+
+Until the BEI-torch backend lands the `llama_bidirec` serde alias and a tokenizer-config
+fallback, point `checkpoint_repository.repo` at a fork with those three fields patched.
+See the per-template README for the exact patch set.