Add tier-1 BEI-torch example templates

[bohanyao-baseten]

Summary

Three tier-1 truss templates for base_model: encoder_torch (BEI-torch backend), in a new 11-embeddings-reranker-classification-torch/ directory mirroring the existing -tensorrt/ layout:

• BEI-torch-nvidia-Nemotron-3-Embed-8B-bf16 (H100)
• BEI-torch-nvidia-Nemotron-3-Embed-1B-bf16 (L4)
• BEI-torch-nvidia-llama-nemotron-rerank-1b-v2 (L4)

Paired with basetenlabs/truss#2592 and basetenlabs/baseten#25084.

Plan: https://app.notion.com/p/3ae91d24727381d5902de1d02944ead4

Notes

• Templates are hand-written, not yet wired into ../11-embeddings-reranker-classification-tensorrt/templating/generate_templates.py. A BEI_TORCH Solution class needs to land there before tier-2/3 templates get added. Flagged as follow-up.
• Each README calls out the three-JSON-patch workaround needed on the stock NVIDIA repos (transformers-incompatible tokenizer_class: "TokenizersBackend", list-form extra_special_tokens, unregistered model_type: "llama_bidirec") until BEI-torch lands the serde alias + tokenizer-config fallback upstream. Customers point at a forked repo with those patches for now.

Test plan

• Wait for Add base_model=encoder_torch for BEI-torch backend truss#2592 and basetenlabs/baseten#25084 to merge and the BEI_TORCH_IMAGE_URI constance to be set to a real tag.
• For each template: truss push --publish against a patched HF fork, confirm MODEL_READY, check /v1/embeddings (or /rerank) returns valid outputs, run a cosine-similarity spot-check against a vLLM reference.

🤖 Generated with Claude Code

[CLAassistant]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.
_{You have signed the CLA already but the status is still pending? Let us recheck it.}