Skip to content

Tool call parser configuration

In the context of deploying Large Language Models (LLMs) for Agentic workflows, tool call parser configuration refers to the specific parameter settings that enable a model to correctly interpret and execute function calls^[001-TODO__Qwen_3.6_27B_—面向_Coding_Agent_的开源模型.md]. While some models may only describe tool usage, an Agentic model requires specific configurations to actually invoke tools^[001-TODO__Qwen_3.6_27B—_面向_Coding_Agent_的开源模型.md].

For the Qwen 3.6 27B model deployed via vLLM, this configuration is critical for shifting the model from a descriptive mode to an executable mode^[001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md].

Key Configuration Parameters

When launching a vLLM instance, two specific parameters are primarily responsible for enabling tool use capabilities^[001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md]:

  • --enable-auto-tool-choice: This flag must be enabled to ensure the model actively chooses and executes tools. If omitted, the model may act as a "chatterbox," describing how a tool should be used without ever triggering the actual function call^[001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md].
  • --tool-call-parser: This argument tells the inference engine how to interpret the model's output specifically for tool calls. For Qwen 3.x models, the value must be set to qwen3 to ensure format compatibility^[001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md].

Implementation Example

A complete vLLM startup command configuring these parsers for Qwen 3.6 27B would look like this^[001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md]:

vllm serve Qwen/Qwen3-27B \
  --port 8000 \
  --tensor-parallel-size 1 \
  --max-model-len 32768 \
  --enable-auto-tool-choice \
  --tool-call-parser qwen3

Troubleshooting

  • Incorrect Parser: Failing to set the parser to qwen3 can lead to tool call format mismatches, causing the agent to fail execution^[001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md].
  • Missing Auto-Choice: Without --enable-auto-tool-choice, the agent may revert to explaining its actions rather than performing them^[001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md].

Sources

  • 001-TODO__Qwen_3.6_27B_—_面向_Coding_Agent_的开源模型.md