poolside

/

Laguna-M.1

like 139

Follow

Poolside 1.35k

One-line change proposing that the template's enable_thinking default match the other artifacts in this repo.

Three things here say reasoning is on by default:

1. generation_config.json: "default_chat_template_kwargs": {"enable_thinking": true}
2. The model card: "Reasoning is on by default; pass enable_thinking=False to skip the <think> block."
3. poolside/Laguna-S-2.1's template already uses default(true).

But chat_template.jinja has {%- set enable_thinking = enable_thinking | default(false) -%}.

This is invisible on vLLM if the server passes --default-chat-template-kwargs as the model card instructs. Stacks without that mechanism read the template default directly, TensorRT-LLM doesn't consume default_chat_template_kwargs, so a request omitting chat_template_kwargs renders a thinking-off prompt while every other artifact says thinking is on, and the reasoning parser missclassifies the response.
Context: NVIDIA/TensorRT-LLM#17146.