Has anyone found a solution to the false parsing from GPT-OSS models ?

[Alekyuk]

Hi everyone.

I tried many, many things but can't seem to find a way to use GPT-OSS-120b (model originally downloaded with LM Studio) so that it stops outputting the harmony tags like <|channel|>.

I use it with Open WebUI and Codex and the problem happens with the two.
I tried everything that I read on this pull request : https://github.com/ikawrakow/ik_llama.cpp/pull/723

Here is the command that I used :

llama-server
--model ~/.lmstudio/models/lmstudio-community/gpt-oss-120b-GGUF/gpt-oss-120b-MXFP4-00001-of-00002.gguf
--no-warmup
--host 0.0.0.0
--port 8080
--ctx-size 10192
--threads 8
--threads-batch 40
--batch-size 512
--ubatch-size 128
--gpu-layers 200
--n-cpu-moe 27
--parallel 2
--reasoning-format auto
--temp 0.4
--top-k 0
--top-p 1.0
--repeat-penalty 1.1
--presence-penalty 0.4
--jinja
--alias "GPT-OSS-120b"

Example of the issue :
<|channel|>analysis <|channel|>analysis<|message|>The user just says "Hey". Likely a greeting. Should respond accordingly. Keep tone friendly. Possibly ask how can help.<|end|><|start|>assistant<|channel|>finalHello! How can I assist you today?

version :
build/bin/llama-server --version version: 3928 (16f30fcf3) built with cc (Ubuntu 11.4.0-1ubuntu1~22.04.2) 11.4.0 for x86_64-linux-gnu

I opened this discussion and not an issue as I have seen that other issues are already opened but not primarily for this model. I hope that I don't make any mistake here, this is my first time posting on a GitHub repo.

Thank you for your help.

[am17an]

Did you see this #15396?

[Alekyuk]

I am deeply sorry.

As I stated in my post, this was the very first time that I posted anything on GitHub discussions, and obviously I did a mistake ! I did not realize that I was posting on llama.cpp's repo ... instead of ik_llama.cpp's ! I have no problem using GPT-OSS-120b on llama.cpp so I will close this discussion.

Sorry !