froggeric

/

Qwen-Fixed-Chat-Templates

like 1.5k

Summary

Users who cannot pass chat_template_kwargs (LM Studio and similar UIs) have no working way to change the default reasoning effort. A user in peculiar-ragdoll/Qwen-Sharp-Chat-Templates#3 (same code as here) edited {%- set _initial_effort = 'medium' %} to 'xhigh' and saw no effect. That is expected: _effort_raw falls back to the literal 'medium' two lines above, and the if/elif/else that follows always overwrites _initial_effort, so that line is dead code.

Changes

• chat_template.jinja: one knob, {%- set _default_reasoning_effort = 'medium' %}, with a one-line Jinja comment, used as the fallback for _effort_raw. The dead _initial_effort = 'medium' line is removed (the else branch still sets 'medium'). Setting the knob to 'none' defaults to non-thinking mode; an explicit reasoning_effort kwarg still overrides the knob. Rendering with no kwargs is byte-identical to before.
• chat_template_oneline.txt: regenerated with scripts/minify_jinja.py (the comment is stripped there, as all comments are).
• scripts/test_v22.py: new test 24c renders a patched copy of the template with the knob set to 'xhigh' and 'none', checks the stock default injects nothing, and checks an explicit kwarg still wins.
• README.md: a short "Editing defaults without kwargs" list in the reasoning-effort section, naming the three top-of-file defaults (_default_reasoning_effort, enable_thinking, _tool_format — the last one is what #79 already tells LM Studio MLX users to edit).

The {#- … #} comment form is accepted by minja (comment_tok regex allows the - trim marker) as well as Jinja2, so llama.cpp / LM Studio / oMLX / vLLM all parse it.

Verification

python3 scripts/test_v22.py → 101/101, and again with QWEN_TEMPLATE_FILE=chat_template_oneline.txt → 101/101; python3 scripts/test_v21.py → 9/9; python3 scripts/fuzz_template.py --cases 300 → clean.