vLLM XML tool_call emitted inside think block loses tool calls

Parent: Mac local LLMs: Chat templates, reasoning and tool calling · Published reference · snapshot 2026-10-05

Old path: the reasoning parser moved all text before `</think>` into `reasoning`, and the tool parser read only `content`, so the call was lost.

These notes link each claim to its source. A source may be a research report hosted on this site rather than the primary document. A published reference means the content is available; it does not certify independent review or accuracy.Read the editorial policy and follow the sources before relying on a claim.

Facts

Old path: the reasoning parser moved all text before `</think>` into `reasoning`, and the tool parser read only `content`, so the call was lost. [source]
Current path: a single `Qwen3Parser` engine in `vllm/parser/qwen3.py` handles reasoning and tool calls in one state machine. A `<tool_call>` token seen in the REASONING state is a transition to the tool-call state that emits a reasoning-end event and a tool-call-start event. [source]
In the same grammar, a bare `<function=` in the content state starts a call without `<tool_call>`, a duplicate `</think>` in the content state is dropped, and a second `<tool_call>` without a closing tag is accepted. [source]
Both names `qwen3_coder` and `qwen3_xml` now resolve to the same adapter class, `Qwen3EngineToolParser`. [source]
The engine reads `enable_thinking` from `chat_template_kwargs`, defaulting to true. When thinking is off, it starts in the content state and `extract_reasoning` returns no reasoning. [source]
Parameter values are taken from `<parameter=NAME>VALUE</parameter>`; one leading and one trailing newline are trimmed; the arguments object is built with `json.dumps(..., ensure_ascii=False)`. [source]
2026-04-06: PR 39055 opened to promote XML calls from reasoning to content in `qwen3_reasoning_parser`, non-streaming only. [source]
2026-04-25: a commenter judged it superseded by PR 40783, a broader Qwen3 reasoning-parser fix; the later comments named streaming as covered there. [source]
2026-06-23: PR 39055 was closed unmerged. Maintainer bbrowning wrote that the Qwen3 parser was rewritten from scratch in a new streaming parser engine, released in vLLM 0.24, which "specifically handles" extracting tool calls from thinking sections. [source]
vLLM v0.30.0 (2026-09-22) is the latest release listed; its notes mention `is_reasoning_end` derived from the grammar (#56200) and full-history reasoning scans eliminated (#55223). [source]
The old files `vllm/reasoning/qwen3_reasoning_parser.py` and `vllm/tool_parsers/qwen3coder_tool_parser.py` return 404 on main; the reasoning and tool-parser directories now list `qwen3_engine_reasoning_parser.py` and `qwen3_engine_tool_parser.py`. [source]
Backport result on the old code (Qwen3.6-35B-A3B-FP8, `enable_thinking=false`, vLLM 0.19.2rc1, n=20): the PR 39055 helper raised the clean-tool-call rate from about 20 percent to about 40 percent; the author explained that with thinking off the model mostly omits `</think>`, so a second fix for the missing-`</think>` case was needed. [source]
On 0.18.x, streaming stayed broken even with PR 39055 because `serving.py` switched to the tool parser only when `is_reasoning_end(output_token_ids)` saw a `</think>`. A commenter proposed a per-request promoted flag overriding `is_reasoning_end`. [source]
Review catch: promoted calls must be appended after existing content, because the old `qwen3_coder` parser used `content.find("<tool_call>")` and dropped text that came after a leading call. [source]
The grammar treats the first `<tool_call>` in reasoning as the end of reasoning even if the model was only discussing the tag; the source shows the transition but no guard. [source]
vLLM's own convention now: model-specific fixes live in `vllm/parser/<model>.py`, adapters in `tool_parsers/` stay thin, and the legacy `ToolParser` interface is frozen for out-of-tree plugins. [source]
Whether a server should recover a call from reasoning. vLLM (grammar transition) and Ollama (injects `</think>`) recover it. aldehir's earlier statement about sglang, held in reasoning-parser-versus-tool-parser-interaction.md, takes the opposite reference behaviour. Not re-argued here. [source]
Whether vLLM 0.24 shipped the Qwen3 engine as default for `qwen3` and `qwen3_coder` names, or behind a flag; the release notes read did not say, and main shows it as the registered class. [source]
The state of PR 40783 (merged or closed); the cached page was truncated before the merge fields. [source]
Whether any Apple Silicon vLLM variant (vllm-metal, vllm-mlx) uses the engine parser; vllm-mlx has its own parser per existing dossier. [source]
vLLM PR 39055 "Fix Qwen3.5 reasoning tool calls embedded inside think" was closed on 2026-06-23 with merged = false. [source]
Maintainer bbrowning closed it by saying the Qwen3 parser was rewritten in a streaming parser engine released in vLLM 0.24 that extracts tool calls from thinking sections. [source]
`vllm/parser/qwen3.py` defines a transition from REASONING on `<tool_call>` to the tool-call state emitting reasoning-end and tool-call-start events. [source]
`vllm/parser/qwen3.py` accepts a `<function=` with no `<tool_call>` wrapper, drops a duplicate `</think>`, and accepts consecutive calls without a closing `</tool_call>`. [source]
`qwen3_coder` and `qwen3_xml` both map to `Qwen3EngineToolParser` in `vllm/tool_parsers/__init__.py`. [source]
`Qwen3EngineToolParser` sets `structural_tag_model = "qwen_3_coder"`. [source]
The Qwen3 engine reads `enable_thinking` from `chat_template_kwargs` (default true) and returns no reasoning when it is false. [source]
vLLM's `tool_parsers/AGENTS.md` says new models should implement a unified `Parser` in `vllm/parser/` and expose it through a thin adapter, and that the legacy `ToolParser` interface must not change. [source]
On vLLM main the files `qwen3_reasoning_parser.py` and `qwen3coder_tool_parser.py` no longer exist (404). [source]
A backport of PR 39055 raised clean tool calls from about 20 to about 40 percent (n=20) on Qwen3.6-35B-A3B-FP8 with thinking disabled. [source]
A commenter reported 0.18.x streaming stayed broken because the switch to the tool parser depended on `is_reasoning_end` seeing `</think>`. [source]
vLLM v0.30.0 was released 2026-09-22 and its notes list `is_reasoning_end` derived from the grammar. [source]
An agent stack on vLLM 0.24 or later no longer needs a tool-call-in-think workaround for Qwen3.5 and Qwen3.6; one on 0.18 to 0.23 still does. [source]

Corrections and disagreements

CONTRADICTS: reasoning-parser-versus-tool-parser-interaction.md, which states for vLLM that "reasoning extraction runs first and the tool parser reads only content; a call left in reasoning is lost" and that PR 39055's merge state was not confirmed. PR 39055 closed unmerged on 2026-06-23, and on current main a call inside reasoning is recovered by grammar. The statement holds only for vLLM before the engine parser (0.18 to 0.23). [source]
CONTRADICTS: tool-call-parser-and-chat-template-mismatch.md line 74-75, which treats `qwen3_coder` and `qwen3_xml` as different parsers with different robustness. On main they share one engine. Third-party comparisons of their regex versus expat internals describe the older code. [source]

Children

No children recorded.

← the whole tree · 3D view· how to read this page