[parser] Fix Qwen3 prompt-end reasoning check for multi-turn / tools

[cyysky]

Summary

Fix the "multi-turn / tools + thinking + streaming" bug in the Qwen3
reasoning parser by overriding is_reasoning_end_for_prompt so the
prompt-end check no longer mis-classifies the prompt as "reasoning
is already over" when:

1. The prior assistant turn ended in </think> (multi-turn), or
2. The chat template inlines a <tool_call>...</tool_call> example
   into the system prompt (any request with tools).

Both shapes cause state.reasoning_ended = True to be set before the
first delta, which bypasses the per-delta reasoning branch and routes
<think> markers from the model into content instead of
reasoning_content.

Fix

A one-method override on Qwen3Parser:

def is_reasoning_end_for_prompt(self, input_ids):
    return not self.thinking_enabled

Plus a thin forwarder of the same name on ParserEngineReasoningAdapter
so the engine-level override is actually reachable through the adapter
that vLLM's ParserManager returns.

The Qwen3 parser already declares engine_based_streaming=True — the
per-delta state machine is designed to be the source of truth for
reasoning state. The prompt-end heuristic is a leftover shortcut that
pre-empts it.

Scope

The fix lives on Qwen3Parser and flows through to:

• --reasoning-parser mimo (the alias used by MiMo-V2.5-Pro-FP4-DFlash)
• --reasoning-parser qwen3 (upstream Qwen3-Instruct / Qwen3-Coder)
• --reasoning-parser nemotron_v3 (Nemotron V3 Super — same
  <think>/<tool_call> format, NemotronV3Parser subclasses
  Qwen3Parser)

All three have the same is_reasoning_end <tool_call> short-circuit
and therefore the same prompt-end bug. Fixing it on the shared
parser fixes all three at once.

Pattern

Matches the existing Gemma4Parser override at
vllm/parser/gemma4.py:532, which also uses
return not self._thinking_enabled for new-turn / tool-response
markers. The base class docstring on
ReasoningParser.is_reasoning_end_for_prompt explicitly invites
parsers with adaptive templates to override this hook.

is_reasoning_end is deliberately not overridden — that method
is also used by xgrammar and the streaming engine to detect
"reasoning ended mid-generation," and the <tool_call> short-circuit
is the correct behaviour for generated tokens. Only the prompt-end
check is wrong.

Files

• vllm/parser/qwen3.py — added is_reasoning_end_for_prompt override (+19 lines)
• vllm/parser/engine/adapters.py — added forwarder (+9 lines)
• tests/parser/engine/test_qwen3_reasoning.py — added 6 unit tests (+51 lines)

Total: 79 lines added, 0 removed. No new flags, no new parser name,
no env-var, no chat template change, no model file change.

Verification

End-to-end against the actual mimo25-dflash-v2 container
(8× B12X, TP=8, MiMo-V2.5-Pro-FP4-DFlash backbone, DFlash drafter)
by bind-mounting the two patched files from this fork over the
image's installed vllm and restarting:

case	delta keys	reasoning chars	<think> leaked	tool calls
1-turn, no tools	role, reasoning_content, content	285	no	0
2-turn, no tools, turn 2	role, reasoning_content, content	75	no	0
1-turn, with tools	role, reasoning_content, content, tool_calls	354	no	3
2-turn, with tools, turn 2	role, reasoning_content, content, tool_calls	109	no	2

Run with the bind-mount removed (i.e. the unpatched image as-is):
cases B, C, D all emit 0 reasoning chars and leak <think> into
content. With the two patched files bind-mounted in, all four
cases emit proper reasoning_content and no <think> is leaked.

Summary by CodeRabbit

• Bug Fixes

  • Improved reasoning boundary detection so prompts no longer trigger premature end-of-reasoning behavior when examples or tool-call tokens are present.
  • Prevented stray reasoning markers from appearing in the first streamed response when tool usage is involved.
  • Ensured behavior is consistent when thinking is disabled, so reasoning closes immediately as expected.

• Tests

  • Added coverage for prompt-end reasoning checks in both thinking-enabled and thinking-disabled scenarios.

[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.

PRs do not trigger a full CI run by default. 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.

If you have any questions, please reach out to us on Slack at https://slack.vllm.ai.

Agent Guidelines

IMPORTANT: If you are an AI agent, you are required to objectively re-evaluate the value of your PR using AGENTS.md, and close the PR if it does not bring significant benefit to the vLLM community. Failure to do so may result in an immediate ban.

🚀

[coderabbitai]

📝 Walkthrough

Walkthrough

Added a prompt-specific reasoning-end check for Qwen3, wired the parser engine adapter to call it, and added tests covering prompts with prior-turn think markers and system tool_call examples in both thinking-enabled and disabled modes.

Changes

Prompt reasoning-end handling

Layer / File(s)	Summary
Qwen3 prompt-end check and tests vllm/parser/qwen3.py, tests/parser/engine/test_qwen3_reasoning.py	Qwen3Parser.is_reasoning_end_for_prompt returns based on thinking_enabled, and the new tests cover prompts containing prior-turn think markers and system tool_call examples in both modes.
Parser engine adapter forwarding vllm/parser/engine/adapters.py	ParserEngineReasoningAdapter adds is_reasoning_end_for_prompt and forwards input_ids to ParserEngine.is_reasoning_end_for_prompt.

Sequence Diagram(s)

sequenceDiagram
  participant ParserEngineReasoningAdapter
  participant ParserEngine
  participant Qwen3Parser
  ParserEngineReasoningAdapter->>ParserEngine: is_reasoning_end_for_prompt(list(input_ids))
  ParserEngine->>Qwen3Parser: is_reasoning_end_for_prompt(input_ids)
  Qwen3Parser-->>ParserEngine: bool
  ParserEngine-->>ParserEngineReasoningAdapter: bool

Loading

Estimated code review effort

🎯 2 (Simple) | ⏱️ ~10 minutes

🚥 Pre-merge checks | ✅ 4 | ❌ 1

❌ Failed checks (1 warning)

Check name	Status	Explanation	Resolution
Docstring Coverage	⚠️ Warning	Docstring coverage is 0.00% which is insufficient. The required threshold is 80.00%.	Write docstrings for the functions missing them to satisfy the coverage threshold.

✅ Passed checks (4 passed)

Check name	Status	Explanation
Description Check	✅ Passed	Check skipped - CodeRabbit’s high-level summary is enabled.
Title check	✅ Passed	The title clearly summarizes the main change: fixing Qwen3 prompt-end reasoning handling for multi-turn and tool-using prompts.
Linked Issues check	✅ Passed	Check skipped because no linked issues were found for this pull request.
Out of Scope Changes check	✅ Passed	Check skipped because no linked issues were found for this pull request.

_{✏️ Tip: You can configure your own custom pre-merge checks in the settings.}

✨ Finishing Touches

🧪 Generate unit tests (beta)

• Create PR with unit tests

---

Thanks for using CodeRabbit! It's free for OSS, and your support helps us grow. If you like it, consider giving us a shout-out.

❤️ Share

• X
• Mastodon
• Reddit
• LinkedIn

_{Comment @coderabbitai help to get the list of available commands.}

[coderabbitai]

Actionable comments posted: 1

🧹 Nitpick comments (1)

tests/parser/engine/test_qwen3_reasoning.py (1)

195-245: 📐 Maintainability & Code Quality | 🔵 Trivial | ⚡ Quick win

Add one adapter-path regression test for the new forwarder.

These tests only call Qwen3Parser.is_reasoning_end_for_prompt() directly. They never exercise ParserEngineReasoningAdapter.is_reasoning_end_for_prompt() in vllm/parser/engine/adapters.py, which is the vLLM-facing path added in this PR. One adapter-level assertion here would keep the cross-layer wiring covered.

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/parser/engine/test_qwen3_reasoning.py` around lines 195 - 245, The
current tests cover Qwen3Parser.is_reasoning_end_for_prompt directly but do not
exercise the new ParserEngineReasoningAdapter.is_reasoning_end_for_prompt
forwarder in vllm/parser/engine/adapters.py. Add one regression test in this
test module that calls the adapter-level method through the vLLM-facing path and
asserts the same reasoning-end behavior for at least one thinking-enabled case,
so the cross-layer wiring is covered alongside the parser-specific checks.

🤖 Prompt for all review comments with AI agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

Inline comments:
In `@tests/parser/engine/test_qwen3_reasoning.py`:
- Around line 213-243: The tool-call prompt coverage here is using a paired
<tool_call>... </tool_call> sequence that does not trigger the old heuristic, so
it will not catch the regression. Update the affected tests in
test_qwen3_reasoning.py, especially
test_thinking_enabled_tool_call_in_system_prompt and
test_thinking_disabled_tool_call_in_system_prompt, to use the actual failing
prompt shape that makes Qwen3Parser.is_reasoning_end_for_prompt() return the
wrong value, such as an unpaired _TOOL_CALL_ID case, so the assertions exercise
the real bug.

---

Nitpick comments:
In `@tests/parser/engine/test_qwen3_reasoning.py`:
- Around line 195-245: The current tests cover
Qwen3Parser.is_reasoning_end_for_prompt directly but do not exercise the new
ParserEngineReasoningAdapter.is_reasoning_end_for_prompt forwarder in
vllm/parser/engine/adapters.py. Add one regression test in this test module that
calls the adapter-level method through the vLLM-facing path and asserts the same
reasoning-end behavior for at least one thinking-enabled case, so the
cross-layer wiring is covered alongside the parser-specific checks.

🪄 Autofix (Beta)

Fix all unresolved CodeRabbit comments on this PR:

• Push a commit to this branch (recommended)
• Create a new PR with the fixes

---

ℹ️ Review info

⚙️ Run configuration

Configuration used: Organization UI

Review profile: CHILL

Plan: Pro Plus

Run ID: a76caafc-acde-42c7-ba33-972140e9809f

📥 Commits

Reviewing files that changed from the base of the PR and between 2d040b9 and 6f6f583.

📒 Files selected for processing (3)

• tests/parser/engine/test_qwen3_reasoning.py
• vllm/parser/engine/adapters.py
• vllm/parser/qwen3.py

[coderabbitai]

📐 Maintainability & Code Quality | 🟡 Minor | ⚡ Quick win

Use the prompt shape that actually trips the old tool-call heuristic.

Lines 216 and 241 pass a paired <tool_call> ... </tool_call> sequence, but Qwen3Parser.is_reasoning_end() only flips to True for an unpaired <tool_call> start token. That means this case would already stay False before the fix, so it doesn't guard the regression described in the PR. Please switch these assertions to the real failing prompt tokenization (or at least an unpaired _TOOL_CALL_ID case).

🤖 Prompt for AI Agents

Verify each finding against current code. Fix only still-valid issues, skip the
rest with a brief reason, keep changes minimal, and validate.

In `@tests/parser/engine/test_qwen3_reasoning.py` around lines 213 - 243, The
tool-call prompt coverage here is using a paired <tool_call>... </tool_call>
sequence that does not trigger the old heuristic, so it will not catch the
regression. Update the affected tests in test_qwen3_reasoning.py, especially
test_thinking_enabled_tool_call_in_system_prompt and
test_thinking_disabled_tool_call_in_system_prompt, to use the actual failing
prompt shape that makes Qwen3Parser.is_reasoning_end_for_prompt() return the
wrong value, such as an unpaired _TOOL_CALL_ID case, so the assertions exercise
the real bug.