qwen3-strict-think-end

Source: fixes/qwen3-strict-think-end/

End Qwen3 reasoning only at </think>, so a tool-call example the model quotes while thinking is not returned as a real tool call. Opt-in.

Affects	Qwen3-family models served with --reasoning-parser qwen3 and tool calling, for example Qwen3.6-27B
Releases	v0.26.0, v0.30.0
Patches	vllm/parser/qwen3.py
Upstream	vllm#59821 (open), issue vllm#58147
Test	none (parser state-table change; covered by the end-to-end check below)

Problem

vLLM's engine-based Qwen3 parser treats a <tool_call> that appears while the model is still reasoning as an implicit end of reasoning: its grammar has a (REASONING, TOOL_START) transition straight into a tool call. On the vLLM commit v0.26.0 tracks, Qwen3Parser.is_reasoning_end also returns True on an unpaired <tool_call> token. This was added for models that call tools without closing </think>.

Symptom

While thinking, a model sometimes quotes the tool-format example from its chat template (<tool_call><function=example_function_name>...). The parser ends the reasoning there and returns the quoted example as a tool call, followed by the model's real call. Clients see an extra call to a function that does not exist, or a duplicated call.

Who is affected

Qwen3-family models that always close </think> before calling a tool and whose reasoning can mention the tool-call format. Models that rely on the implicit end (they call tools without closing </think>) need the stock behaviour, which is why this fix is opt-in.

The fix

The implicit-end transition is removed from the parser's state table, so only </think> ends the reasoning. On v0.26.0 the separate unpaired-<tool_call> check in is_reasoning_end is skipped as well; on v0.30.0 that check is derived from the state table and needs no second edit. Applying the fix enables the behaviour; there is no runtime switch.

Verify

python3 fixes/qwen3-strict-think-end/apply.py --check   # 0 = applied, 1 = not applied, 2 = source drifted

End to end: send a tool-calling request with reasoning enabled many times at the model's sampling temperature and count calls to example_function_name or duplicated calls. With the fix there are none, provided the model closes </think> before its calls.