froggeric

/

Qwen-Fixed-Chat-Templates

like 1.55k

Hey everyone,

Just pushed v22.4 to the repo. This is a quick follow-up release addressing a few community PRs and polishing up tool calling, vLLM compatibility, and UI customization.

Here is what is new:

Parallel tool call token parity
Consecutive <tool_call> blocks in assistant history now separate with a single \n instead of \n\n. This matches what the model actually outputs during parallel generation and keeps tokenization consistent across turns.

Support for message.reasoning
Added extraction for message.reasoning (used by vLLM's OpenAI server endpoint and Responses API schemas), alongside reasoning_content and thinking.

Direct configuration knobs for UI users
If you use LM Studio or any UI that does not let you pass chat_template_kwargs, I set up clear variables right at the top of chat_template.jinja:

• _default_reasoning_effort = 'medium' (change to 'low', 'xhigh', or 'none')
• enable_thinking = true
• _tool_format = 'xml'
  This lets you edit your preferred defaults at the very top of the file without having to mess with the mapping logic below.

Documentation updates
Updated the README to clarify the recommended vLLM flags (--reasoning-parser qwen3 --tool-call-parser qwen3_xml) and documented that inline chat tags like <|think_off|> stay active across later turns until another tag overrides them.

Big thanks to @gdevenyi for contributing these improvements!