Raise 400 on model mismatch when transformers serve is pinned

[qgallouedec]

When transformers serve is launched with a positional model argument, the server silently overwrites the "model" field in every incoming request with the pinned model id. This is surprising: a client that asks for model B receives a response generated by model A, with no indication that the requested model was ignored.

This PR changes _resolve_model in src/transformers/cli/serving/utils.py to return 400 Bad Request when the client-supplied model disagrees with the pinned one. Requests that omit model or send an empty value still fall back to the pinned model, so existing well-behaved clients are unaffected.

Reproducer

Start the server pinned to model A:

transformers serve meta-llama/Llama-3.2-1B-Instruct

Then query with a different model, either via curl:

curl http://localhost:8000/v1/chat/completions \
  -H "Authorization: Bearer x" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "Qwen/Qwen2.5-0.5B-Instruct",
    "messages": [{"role": "user", "content": "Hi!"}]
  }'

or via the OpenAI Python client:

# repro.py
from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="x")

print(client.chat.completions.create(
    model="Qwen/Qwen2.5-0.5B-Instruct",
    messages=[{"role": "user", "content": "Hi!"}],
))

Before:

ChatCompletion(id='5b793e5a-189b-49ac-a022-7f74d60c563a', choices=[Choice(finish_reason='stop', index=0, logprobs=None, message=ChatCompletionMessage(content='How can I assist you today?', refusal=None, role='assistant', annotations=None, audio=None, function_call=None, tool_calls=None))], created=1776194005, model='meta-llama/Llama-3.2-1B-Instruct@main', object='chat.completion', service_tier=None, system_fingerprint=None, usage=CompletionUsage(completion_tokens=8, prompt_tokens=37, total_tokens=45, completion_tokens_details=None, prompt_tokens_details=None))

After:

Traceback (most recent call last):
  File "/fsx/qgallouedec/transformers/repro.py", line 5, in <module>
    resp = client.chat.completions.create(
        model="Qwen/Qwen2.5-0.5B-Instruct",
        messages=[{"role": "user", "content": "Hi!"}],
    )
  File "/fsx/qgallouedec/miniconda3/envs/trl/lib/python3.13/site-packages/openai/_utils/_utils.py", line 286, in wrapper
    return func(*args, **kwargs)
  File "/fsx/qgallouedec/miniconda3/envs/trl/lib/python3.13/site-packages/openai/resources/chat/completions/completions.py", line 1204, in create
    return self._post(
           ~~~~~~~~~~^
        "/chat/completions",
        ^^^^^^^^^^^^^^^^^^^^
    ...<47 lines>...
        stream_cls=Stream[ChatCompletionChunk],
        ^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
    )
    ^
  File "/fsx/qgallouedec/miniconda3/envs/trl/lib/python3.13/site-packages/openai/_base_client.py", line 1297, in post
    return cast(ResponseT, self.request(cast_to, opts, stream=stream, stream_cls=stream_cls))
                           ~~~~~~~~~~~~^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
  File "/fsx/qgallouedec/miniconda3/envs/trl/lib/python3.13/site-packages/openai/_base_client.py", line 1070, in request
    raise self._make_status_error_from_response(err.response) from None
openai.BadRequestError: Error code: 400 - {'detail': "Server is pinned to 'meta-llama/Llama-3.2-1B-Instruct'; requested 'Qwen/Qwen2.5-0.5B-Instruct'."}

Notes

• AI assistance was used to draft this change

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[SunMarc]

Fine for me ! cc @LysandreJik for confirmation