Skip to main content

[Frontend] min_p Support in the Responses API

https://github.com/vllm-project/vllm/pull/48084

Purpose​

The Responses API (/v1/responses) accepts top_k as a request field and passes it to SamplingParams, but the corresponding min_p is missing. SamplingParams and SamplingParams.from_optional support min_p, and the Completions and Chat endpoints already expose it.

This PR adds the min_p field to ResponsesRequest and sets a default value in to_sampling_params in the same way as top_k, passing it to SamplingParams.from_optional. It follows the existing top_k handling exactly. Now, /v1/responses can also use min-p sampling, just like other generation endpoints.

This is not a duplicate. I searched for public PRs and issues using terms such as responses min_p, ResponsesRequest min_p, and responses/protocol.py min_p. responses/protocol.py contains no mention of min_p, and there is no public PR to add a sampling parameter to the Responses API. PR #45839 adds a sampling parameter only to a separate Translation API, so it does not overlap.

Files Affected​

  • vllm/entrypoints/openai/responses/protocol.py
  • tests/entrypoints/openai/responses/test_sampling_params.py
  • docs/serving/online_serving/openai_compatible_server.md

Test Plan​

.venv/bin/python -m pytest \
  tests/entrypoints/openai/responses/test_sampling_params.py -v
pre-commit run --files \
  vllm/entrypoints/openai/responses/protocol.py \
  tests/entrypoints/openai/responses/test_sampling_params.py \
  docs/serving/online_serving/openai_compatible_server.md

I added a validation for min_p to test_basic_sampling_params and added a new test, test_min_p_default, to verify the default value (0.0).

Test Results​

8 passed in 1.07s

ruff check, ruff format, and mypy all passed pre-commit.