MODEL_TYPE_TO_TOOL_PARSER maps qwen3_5/qwen3_5_moe to a parser whose format the template does not emit

System Info

• TensorRT-LLM: 1.3.0rc23
• GPU: NVIDIA GeForce RTX 5090 (sm_120), driver 595.71.05
• Backend: PyTorch; single GPU, no TP/PP
• Models: Qwen/Qwen3.8-27B (qwen3_5 family); same class of mismatch previously observed on Qwen3.6-35B-A3B

Who can help?

No response

Reproduction

MODEL_TYPE_TO_TOOL_PARSER (tensorrt_llm/serve/tool_parser/tool_parser_factory.py) maps model
types qwen3_5 and qwen3_5_moe to the qwen3 parser, which expects the JSON tool-call form.

The chat template for these models emits the XML-ish form instead:

<tool_call><function=name><parameter=k>v</parameter></function></tool_call>

Serve with automatic tool-parser selection and issue a request with tools defined. Measured on
a four-call scenario:

• with the auto-selected qwen3 parser: 0 of 4 tool calls parsed
• with --tool_parser qwen3_coder: 4 of 4 parsed, arguments correct including an optional enum

Expected behavior

Automatic selection maps a model family to a parser that can parse the format that family's template emits.

actual behavior

tool_calls comes back [] and content is emptied, with HTTP 200 and no error or warning. The response is indistinguishable from a model that chose not to call a tool, so a caller cannot detect the failure.

additional notes

--tool_parser qwen3_coder parses this format correctly on the same build and model, so the
parser exists and only the mapping is wrong.

Related but distinct: #17917 reports that a parser which matches its marker yet extracts no calls
returns 200 with no diagnostic — that is the silence; this issue is the mapping that triggers
it here. Fixing either alone leaves the other. #17574 and #17740 concern the streaming parser.

Questions for maintainers

1. Should qwen3_5/qwen3_5_moe map to qwen3_coder, or is the correct discriminator the
   template content rather than the model type?
2. Are other entries in MODEL_TYPE_TO_TOOL_PARSER verified by effect against their templates?
   The failure mode here is silent, so an unverified entry looks identical to a working one.

Start in tensorrt_llm/serve/tool_parser/tool_parser_factory.py and compare the qwen3 and qwen3_coder parser mappings with the tool-call format emitted by the qwen3_5 templates. Reproduce automatic parser selection with tools defined and verify that qwen3_5 and qwen3_5_moe produce parsed tool calls with their arguments rather than empty content and tool_calls.