Skip to main content

[CI/Build] Support for `bad_words` at the `/v1/completions` endpoint

https://github.com/vllm-project/vllm/pull/46793

Purpose​

The chat autocomplete endpoint ( /v1/chat/completions ) exposes the bad_words field and passes it to SamplingParams, but the existing autocomplete endpoint ( ) does not do so. This is a functional disparity. Users of the autocomplete endpoint cannot hide specific words.SamplingParamsbad_words

This PR adds the bad_words field to reflect the existing chat implementation (the sampling parameter block and the field) and connects it via CompletionRequest.to_sampling_params()chat_completion/protocol.pybad_words=self.bad_wordsto_sampling_params

This is not a duplicate. bad_words  I searched for open PRs regarding (title/body), bad_words CompletionRequest, and completion bad_words parity, but the existing bad_words PRs concern tokenizer conversion/caching bugs, and there are no PRs that add the field to the autocomplete endpoint.

Test Plan​

python -m pytest tests/entrypoints/openai/completion/test_completion.py \
  -k "bad_words" -v
# lint
ruff check vllm/entrypoints/openai/completion/protocol.py \
  tests/entrypoints/openai/completion/test_completion.py
ruff format --check vllm/entrypoints/openai/completion/protocol.py \
  tests/entrypoints/openai/completion/test_completion.py

I added two CPU-specific unit tests.

  • test_completion_request_bad_words_to_sampling_paramsbad_wordsSamplingParams —

    is passed to

    .

  • test_completion_request_bad_words_default_empty — It is set to an empty list by default (same as in chat).

Test Results​

2 passed
ruff: All checks passed! / 2 files already formatted