no <think> tag between the response

#17
by wangxhhhhhhh - opened

after setting "chat_template_kwargs": {"enable_thinking": True}, the identifier still does not appear in the answer

The tag will be detected and filtered, and the content between and will be separated out and returned as the reasoning_content field. This is the parser logic in SGLang and VLLM.

Sadly the reasoning parser doesnt currently work properly,
I have the Problem with GLM-4.6V-Flash in Vllm (tested 0.13.0 and nightly)
on 13 it outputs thinking and response in the content field
on nightly everything gets put in the reasoning_content field with an empty content
i have max length set at 50k tokens and this is happening for responses <2000 tokens aswell.
currently trying to find out how to debug it, but it seems the model is not outputting the correct end reasoning tokens

Sign up or log in to comment