Using qwen3_coder or qwen3_xml ?
#65
by ericli1018 - opened
As the title suggests, this problem has been bothering me for a long time. When I use qwen3_coder, I often have to stop halfway through a task, but the probability is lower with qwen3_xml. However, I'm not sure if this chat-template can work properly with qwen3_xml, so I'd like to ask for everyone's opinion on which one to choose to use this template correctly.
afaik it's qwen3_xml as tool calling is xml trained on qwen3. but thats just what i've read once while there are plenty examples using qwen3_coder
Both parsers target Qwen XML, but qwen3_xml is the updated, preferred parser in current vLLM versions. Older vLLM builds only had qwen3_coder, which had stricter regex handling on multi-line parameter blocks.
- If you are on a recent vLLM release, use --tool-call-parser qwen3_xml.
- If you are on an older vLLM release, use --tool-call-parser qwen3_coder.
- If you explicitly opt into Hermes JSON via --chat-template-kwargs '{"tool_call_format": "json"}', switch the parser to --tool-call-parser hermes.
froggeric changed discussion status to closed