[Bugfix][Reasoning] Strip grouped think markers from streaming deltas

[wuyingjun-lucky]

Summary

Fix streaming reasoning parsing when the thinking start marker is grouped into the same delta as reasoning text, including prefixed same-delta cases.

• Strip grouped start markers in BaseThinkingReasoningParser before emitting reasoning text, matching the non-streaming partition behavior.
• Apply the same normalization in KimiK2ReasoningParser before splitting on </think> or <|tool_calls_section_begin|>, while preserving any implicit reasoning prefix before <think>.
• Add regression tests for grouped start-token streaming cases, including prefixed deltas and Kimi same-delta end/content and tool-section transitions.

Why this is not a duplicate

• Not duplicating fix(reasoning): prevent streaming end-token desync in base and other parsers #39044. That PR fixes end-token / buffered stop-sequence desync; this PR fixes start-token leakage when <think> and reasoning text are grouped into one streaming delta.
• Not duplicating [Bugfix][Tool Parser] Fix Kimi-K2.5 parser accuracy, buffer limits, and token leaks #37384. That PR rewrites the Kimi-K2.5 tool parser; this change is in the reasoning parsers and covers grouped start markers before tool-section routing.

Test plan

• /home/wuyingjun/llm/vllm/.venv/bin/python -m pytest -q tests/reasoning/test_base_thinking_reasoning_parser.py tests/reasoning/test_kimi_k2_reasoning_parser.py
  • Result: 37 passed
• /home/wuyingjun/llm/vllm/.venv/bin/pre-commit run --files vllm/reasoning/basic_parsers.py vllm/reasoning/kimi_k2_reasoning_parser.py tests/reasoning/test_base_thinking_reasoning_parser.py tests/reasoning/test_kimi_k2_reasoning_parser.py
  • Result: passed
• Local GPU OpenAI server smoke test using /home/wuyingjun/code/model/Qwen3-0.6B
  • Result: server started successfully, emitted streaming reasoning chunks, and returned final content 1 + 1 = 2.

AI assistance

This PR was prepared with AI assistance. I reviewed the diff and ran the validation steps above.

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[gemini-code-assist]

Code Review

This pull request addresses an issue where reasoning start markers (e.g., ) could leak into the output when they are grouped with reasoning text in a single streaming delta. It adds comprehensive test cases for both the base and Kimi K2 reasoning parsers and updates the streaming extraction logic to strip the start token. Review feedback highlights that using removeprefix is too restrictive because the start token might not always appear at the very beginning of a delta (for instance, if preceded by whitespace). It is recommended to use more robust string operations like partition or replace to ensure the tag is correctly removed regardless of its position within the delta.

[gemini-code-assist]

The use of removeprefix is too restrictive for streaming deltas. If the model emits a delta where the start token is preceded by other characters (such as a space or a newline), removeprefix will fail to strip the tag, causing it to leak into the reasoning output. To maintain consistency with the non-streaming extract_reasoning implementation (lines 169-172) which discards the prefix and the tag, you should strip the tag regardless of its position in the delta. Using partition is a more robust approach here.

Suggested change

	return DeltaMessage(reasoning=delta_text.removeprefix(self.start_token))
	return DeltaMessage(reasoning=delta_text.partition(self.start_token)[2])

[wuyingjun-lucky]

Updated in 30d3e81. The base parser now uses partition-based stripping so grouped start tags are removed even if they are prefixed in the delta, and I added a regression test covering that case.

[gemini-code-assist]

Using removeprefix here is insufficient if the <think> tag appears in the middle of a streaming delta (e.g., after a leading space). If the tag is not at the very beginning of delta_text, it won't be removed, and will subsequently be included in the reasoning content or interfere with the end-token detection logic at line 225. Since Kimi supports implicit reasoning (where text before the tag is also reasoning), you should specifically remove the tag itself while preserving any preceding text in the delta.

Suggested change

	delta_reasoning = (
	delta_text.removeprefix(self._start_token)
	if (
	self._start_token_id in delta_token_ids
	and self._start_token_id not in previous_token_ids
	)
	else delta_text
	)
	delta_reasoning = (
	delta_text.replace(self._start_token, "", 1)
	if (
	self._start_token_id in delta_token_ids
	and self._start_token_id not in previous_token_ids
	)
	else delta_text
	)

[wuyingjun-lucky]

Updated in 30d3e81. Kimi now removes only the first marker via replace(..., 1), preserving any implicit reasoning prefix before the tag, and there is a regression test for the prefixed same-delta case.

[wuyingjun-lucky]

Addressed the latest review feedback in 30d3e81 by handling prefixed grouped <think> tags in both streaming parsers and adding regression coverage for those cases. Local validation passed:

• /home/wuyingjun/llm/vllm/.venv/bin/python -m pytest -q tests/reasoning/test_base_thinking_reasoning_parser.py tests/reasoning/test_kimi_k2_reasoning_parser.py -> 37 passed
• /home/wuyingjun/llm/vllm/.venv/bin/pre-commit run --files vllm/reasoning/basic_parsers.py vllm/reasoning/kimi_k2_reasoning_parser.py tests/reasoning/test_base_thinking_reasoning_parser.py tests/reasoning/test_kimi_k2_reasoning_parser.py -> passed

The remaining failing check is pre-run-check, which is waiting on the contributor gate: the PR needs a maintainer-applied ready or verified label before the rest of CI will run.

[mergify]

This pull request has merge conflicts that must be resolved before it can be
merged. Please rebase the PR, @wuyingjun-lucky.

https://docs.github.com/en/pull-requests/collaborating-with-pull-requests/working-with-forks/syncing-a-fork