Sampling special tokens, not just <|im_end|>?

[ivanstepanovftw]

Qwen 2.5 have special token <tool_call> and </tool_call>, but I do not know if it is model not trained to generate this token, or llama.cpp rejects generating all special tokens, but <|im_end|>.

How to allow llama.cpp to output <tool_call> token, if model is trained to output this special token?

I am interested more in the server part.

[ivanstepanovftw]

Both llama-cli and llama-server can sample special token <tool_call> during inference. Sorry for inconvenience.