Eval bug: reasoning content isn't extracted into "reasoning_content" field for GLM-4.5/GLM-4.6

[briancw]

Name and Version

version: 6692 (ca71fb9)
built with cc (Ubuntu 13.3.0-6ubuntu2~24.04) 13.3.0 for x86_64-linux-gnu

Operating systems

Linux

GGML backends

CUDA

Hardware

3x Nvidia P40

Models

GLM-4.5-Air, GLM-4.5, GLM-4.6

Problem description & steps to reproduce

Reasoning content isn't being extracted to the "reasoning_content" field when using the GLM 4.5 or 4.6 models. It appears that the chat format is being auto detected as "Hermes 2 Pro" from server logs.

First Bad Commit

No response

Relevant log output

srv  params_from_: Chat format: Hermes 2 Pro

[abc-nix]

When using unsloth GLM-4.5-(Air)-GGUF with llama-server, I get the reasoning_content properly parsed when using the options ''--jinjaand--reasoning-format auto`.

With the new GLM-4.6-GGUF, I need to apply the old chat template used for GLM-4.5 to get the reasoning parsing to work.

[liquidsnakeblue]

I've been troubleshooting the same issue. I am not able to get the reasoning tokens separated out from the content output no matter what I try.

[allozaur]

Hey, the reasoning content parsing and processing improvements are being worked on within #16394, hopefully this would address also this issue

cc @ServeurpersoCom

[briancw]

Seems to be working perfectly now for GLM-4.5, GLM-4.5-Air, and GLM-4.6. Well done guys! This is going to make using those models with a handful of local tools much smoother.