Solves llama.cpp reasoning display
#2
by FilterJoe - opened
Mid-July changes to llama.cpp broke the ability to observe Qwen reasoning displayed. This jinja template fixes that issue, something you did not explicitly mention in the description of your fixes.
The default web UI for llama-server, http://127.0.0.1:8080, shows reasoning used (by any model with reasoning/thinking), as well as an expandable/collapsible reasoning button for each round of conversation. This does not appear when using v21 froggeric template. It does appear with this version.
Thank you!
Hello thank you, some of our reasoning work started with PRs on froggeric's template. I've learned then that they need to have the template a certain way for their inference server, so I wanted to fork 21.3 to customize it mostly for llama.cpp.