muzzy/GLM-5.2-FP8-DFlash-GGUF

Update: now getting 15% acceptance rate on ik_llama with:

--spec-type dflash:n_max=1,p_min=0.0

This was converted with the ik_llama.cpp conversion script.

python3 convert_hf_to_gguf.py --target-model-dir /ai/GLM-5.2-FP8/ /ai/GLM-5.2-FP8-DFlash/ --outtype bf16

Original README:

This is an attempted conversion of https://huggingface.co/UCloud-org/GLM-5.2-FP8-DFlash to gguf format. Both models function in ik_llama, but I cannot get either to get higher than 5% acceptance. I'm not sure which tokenizer to use, I think is the problem.