[Feature][Responses API] Support logprobs(non-stream)

[kebe7jun]

Related issue #23225

Purpose

Support logprobs for Response API (non-stream).

Stream mode is currently divided into several cases:
gpt-oss models use Harmony, as they use a custom parser, and logprobs are not implemented yet. gpt-5-mini also does not support logprobs.

➜  ~ curl https://api.openai.com/v1/responses \
  -H "Content-Type: application/json" \
  -H "Authorization: Bearer sk-proj-xxx" \
  -d '{
    "include": ["message.output_text.logprobs"],
    "model": "gpt-5-mini",
    "input": "hello, ",
    "store": true,
    "top_logprobs": 3
  }'

{
  "error": {
    "message": "logprobs are not supported with reasoning models.",
    "type": "invalid_request_error",
    "param": "include",
    "code": "unsupported_parameter"
  }
}

Other models have not yet fully implemented the Response Streaming API, so support is temporarily unavailable.

vllm/vllm/entrypoints/openai/serving_responses.py

Lines 795 to 798 in 2e2000f

	if not isinstance(context, StreamingHarmonyContext):
	raise NotImplementedError(
	"Streaming is not supported for responses API without Harmony."
	)

Test Plan

Added unit tests.

Test Result

See CI.

(Optional) Documentation Update

---

Essential Elements of an Effective PR Description Checklist

• The purpose of the PR, such as "Fix some issue (link existing issues this PR will resolve)".
• The test plan, such as providing test command.
• The test results, such as pasting the results comparison before and after, or e2e results
• (Optional) The necessary documentation update, such as updating supported_models.md and examples for a new model.

[github-actions]

👋 Hi! Thank you for contributing to the vLLM project.

💬 Join our developer Slack at https://slack.vllm.ai to discuss your PR in #pr-reviews, coordinate on features in #feat- channels, or join special interest groups in #sig- channels.

Just a reminder: PRs would not trigger full CI run by default. Instead, it would only run fastcheck CI which starts running only a small and essential subset of CI tests to quickly catch errors. You can run other CI tests on top of those by going to your fastcheck build on Buildkite UI (linked in the PR checks section) and unblock them. If you do not have permission to unblock, ping simon-mo or khluu to add you in our Buildkite org.

Once the PR is approved and ready to go, your PR reviewer(s) can run CI to test the changes comprehensively before merging.

To run CI, PR reviewers can either: Add ready label to the PR or enable auto-merge.

🚀

[heheda12345]

LGTM! Thanks for the contribution.

[heheda12345]

BTW I think there are still some todos like logprobs + harmony and streaming, so I change "fix xxx" to "related issue xxx". Can you also help on those parts?

[kebe7jun]

BTW I think there are still some todos like logprobs + harmony and streaming, so I change "fix xxx" to "related issue xxx". Can you also help on those parts?

Glad to do it.

[charlotte12l]

@kebe7jun, is non-stream + harmony supported? I saw

vllm/vllm/entrypoints/openai/serving_responses.py

Lines 210 to 215 in de533ab

	if self.use_harmony and request.is_include_output_logprobs():
	return self.create_error_response(
	err_type="invalid_request_error",
	message="logprobs are not supported with gpt-oss models",
	status_code=HTTPStatus.BAD_REQUEST,
	)

[kebe7jun]

@kebe7jun, is non-stream + harmony supported? I saw

vllm/vllm/entrypoints/openai/serving_responses.py

Lines 210 to 215 in de533ab

	if self.use_harmony and request.is_include_output_logprobs():
	return self.create_error_response(
	err_type="invalid_request_error",
	message="logprobs are not supported with gpt-oss models",
	status_code=HTTPStatus.BAD_REQUEST,
	)

no, gpt-oss models not support logprobs.