Function/tool calls are ignored or malformed
Symptom
The model answers in prose instead of calling the tool, or arguments are broken.
Likely causes
- Model not trained for tool use
- Tool template not applied (missing --jinja)
- Too many or poorly described tools
How to fix it
- Use a model known for tool support (Llama 3.1+, Qwen2.5+, Hermes-style).
- Start llama-server with --jinja; use the right --tool-call-parser in vLLM.
- Reduce the tool list; write short, precise descriptions; set temperature to 0.
- Validate arguments and re-prompt with the error.
bash
llama-server -m model.gguf --jinja -ngl 99Still stuck? Check the lessons or measure first with the benchmark lab.