[Bug]: trtllm-serve /v1/chat/completions rejects the standard OpenAI field prompt_cache_key with 400 #20367

System Info

TensorRT-LLM main @ 15f060c, trtllm-serve (any model, any GPU).

Reproduction

curl -s localhost:8000/v1/chat/completions -H 'Content-Type: application/json' -d '{
  "model": "<model>",
  "messages": [{"role": "user", "content": "hi"}],
  "prompt_cache_key": "session-1234"
}'

or with the OpenAI Python client: client.chat.completions.create(..., prompt_cache_key="session-1234").

Expected behavior

The request is served. prompt_cache_key is part of OpenAI's Chat Completions request body (a hint for prompt-cache routing), and clients built on it may set it on every request of a session.

Actual behavior

The request fails validation with extra_forbidden ("Extra inputs are not permitted") for prompt_cache_key, because ChatCompletionRequest inherits extra="forbid" from OpenAIBaseModel and has no such field. ResponsesRequest already accepts it (it switched to extra="allow", and the comment there names prompt_cache_key as one of the fields real clients send), so the same client works against /v1/responses but not against /v1/chat/completions.

Additional notes

KV cache reuse in TensorRT-LLM needs no client hint, and cache_salt is the field that isolates cached prefixes, so the field can be accepted and ignored, the same way parallel_tool_calls and user are. PR to follow.