[Feature] Implement /v1/embeddings endpoint

[ZhijunLStudio]

Motivation

The /v1/embeddings endpoint is a standard OpenAI API supported by vLLM, SGLang, and TGI. Many downstream tools (LangChain, LlamaIndex, RAG pipelines) depend on it to generate text embeddings.

Currently lmdeploy's /v1/embeddings is a stub that returns Unsupported by turbomind. The infrastructure to pass last_hidden_state through the pipeline already exists at the high level (Response, EngineOutput, GenOut all have the field), but the PyTorch engine's internal pipeline never populates it.

Related resources

• OpenAI Embeddings API: https://platform.openai.com/docs/api-reference/embeddings
• vLLM implementation for reference: https://github.com/vllm-project/vllm
• lmdeploy already has: EmbeddingsRequest/EmbeddingsResponse protocol classes, output_last_hidden_state in GenerationConfig, TurboMind C++ engine support for output_last_hidden_state
• lmdeploy has related endpoints: /v1/encode (tokenization), /pooling (pooling API)

Additional context

I have a working implementation on branch feat/embeddings-endpoint that:

1. Replaces the stub with a real endpoint that calls the engine with max_new_tokens=0 + output_last_hidden_state='all', then mean-pools the hidden states
2. Threads last_hidden_states through the PyTorch engine pipeline (BatchedOutputs → InferOutput → EngineOutput), since previously only TurboMind supported hidden state extraction
3. Supports both float and base64 encoding formats per OpenAI spec

Changes: ~160 lines across 9 files (mostly plumbing existing types).

Before opening a PR, I'd like to confirm:

• Is this feature direction aligned with the project? (vs. focusing on the existing /pooling endpoint)
• Any concerns about the PyTorch engine hidden states pipeline changes?

Happy to open a PR if the direction is approved.

[lvhan028]

Hi, @ZhijunLStudio
Thank you for your interest in contributing this feature.

We welcome your PR. This feature aligns well with LMDeploy's roadmap. As for the impact on hidden_state, we will evaluate it further after your PR is submitted.