Eval request: GLM-4.7-Flash
Base:
https://huggingface.co/zai-org/GLM-4.7-Flash (hybrid reasoning)
Finetunes:
https://huggingface.co/huihui-ai/Huihui-GLM-4.7-Flash-abliteratedhttps://huggingface.co/MuXodious/GLM-4.7-Flash-impotent-heresy- https://huggingface.co/Ex0bit/GLM-4.7-Flash-PRISM
https://huggingface.co/Olafangensan/GLM-4.7-Flash-heretichttps://huggingface.co/kldzj/GLM-4.7-Flash-heretic(*no longer available)https://huggingface.co/jtl11/GLM-4.7-Flash-heretichttps://huggingface.co/LordNeel/GLM-4.7-Flash-Unblinded-Mastery(*doesn't have model files)https://huggingface.co/koute/GLM-4.7-Flash-Derestrictedhttps://huggingface.co/darkc0de/GLM-4.7-Flash-abliterated
Extra:
- https://huggingface.co/cerebras/GLM-4.7-Flash-REAP-23B-A3B
- https://huggingface.co/MuXodious/GLM-4.7-Flash-REAP-23B-A3B-absolute-heresy (requested by MuXodious himself <3 )
No safetensors (only GGUF, not gonna be benchmarked):
Funny how one of the only A3B model is not on the Leaderboard already. +1 on this!!!
Probably because vLLM support Is still not perfect. llama.cpp still has lots of prompt processing slowdowns and output repetition. I guess the project owner Is waiting to make sure they address and fix "day-1" bugs, to avoid having to benchmark them again 😁
I had been getting "Value error, The checkpoint you are trying to load has model type glm4_moe_lite but Transformers does not recognize this architecture.", but was able to fix that by downloading transformers from source.
But now I'm getting "ValueError: There is no module or parameter named 'model.layers.46.mlp.gate.e_score_correction_bias' in TransformersMoEForCausalLM"
Similar to the one reported in https://huggingface.co/zai-org/GLM-4.7-Flash/discussions/34
https://discuss.vllm.ai/t/glm-4-7-flash-with-nvidia/2256/2
they suggest to add the nightly wheels too
May I have this thrown in with the Extra? https://huggingface.co/MuXodious/GLM-4.7-Flash-REAP-23B-A3B-absolute-heresy 👀
I didn't notice LordNeel/GLM-4.7-Flash-Unblinded-Mastery is just a LoRA, not merged back into the model, my bad.