[A5bhinav]

Fixes #682

/v1/completions now answers n != 1 with a 400, the same refusal /v1/chat/completions already gives, instead of a 200 with one choice. No change for n = 1.

Summary

CompletionRequest declares n, but handle_completion never read it: n=2 got HTTP 200 and a single choice, so a client asking for several samples (best-of selection, self-consistency, eval harnesses) silently got one and had no way to tell. Chat already refuses the field with 400 "Only n=1 is supported".

• server/openai_api.py: handle_completion returns create_error_response("Only n=1 is supported", param="n") when n != 1. The check sits after _completion_unsupported_reason and before anything is sent to the engine, ahead of the stream / non-stream split, so both paths refuse it, the same way the route already refuses logprobs, echo, suffix and logit_bias. The body matches chat's:

  {"error": {"message": "Only n=1 is supported", "type": "invalid_request_error", "param": "n", "code": null}}

• n <= 0, which was accepted before, is refused too.

Why now: a silent wrong answer is worse than a refusal, and this is the smallest change that makes the two routes agree. Serving n > 1 on completions is a separate, larger change in a follow-up PR.

Tests

tests/server/test_openai_api.py::test_completion_refuses_n_other_than_one, parametrized over n in {0, 2} and stream on and off: the response is a 400 with param "n", and nothing is sent to the engine (state.sent is None). With main's handle_completion, all four cases fail.

Test environment

• GPU: none. I don't have access to an NVIDIA GPU yet, so this has not been run against a live ft serve.
• CPU: Apple M4 Pro, 24 GB RAM, macOS 26.6.2
• torch 2.14.1 (CPU), Python 3.12.13
• This branch at 8fe8c42, A/B against main a2ce636.

Unit tests

PYTHONPATH=python pytest tests/server/test_openai_api.py

This branch: 15 passed, 17 failed, 1 error. Main: 11 passed, 17 failed, 1 error. The 4 extra passes are the new test; with only the test file from this branch, main fails all 4. The 17 failures and the error are all "ModuleNotFoundError: No module named 'flashlib'", which needs CUDA, and are the same on main.

End-to-end A/B against main

Not run yet: it needs a GPU, which I don't have access to yet. Once I do, I will run it and add the results here; if a maintainer or reviewer has one sooner, the repro attached to #682 should show HTTP 400 and PASS on the four completions rows on this branch, and HTTP 200 with one choice (FAIL) on main:

ft serve --model <checkpoint>
python repro_completions_n_ignored.py --base-url http://127.0.0.1:1941