Additional parser config for vLLM

#1
by WonderRico - opened

Hello / Salut,

What are the proper parser config params to use ? for "--reasoning-parser" and "--tool-call-parser" flags

Merci!

OpenLLM France org

The chat template is Qwen3-derived and uses the Hermes format.
So --tool-call-parser hermes --enable-auto-tool-choice

And you can leave --reasoning-parser off.
Luciole-XB-Instruct-1.1 is a non-thinking model. The thinking traces were only used in the first SFT phase (that would correspond to --reasoning-parser deepseek_r1, for the interim SFT-with-thinking checkpoint), but phases two and three (SFT and DPO) were both without thinking.
However we plan to publish thinking model variants in a near future.

It's so good to still have capable non thinking models

Sign up or log in to comment