Hugging Face's logo

Comfy-Org
/
MiniMax-H3

Diffusion Single File
comfyui
Model card Files
xet
 Community
44

Not supported for Apple silicon

#24
by Rhystic1 - opened
Discussion
Rhystic1

Please release a version that is supported on Apple silicon since INT8 and FP8 both fail due to MPS backend not having certain operators implemented and the CPU fallback is extremely slow.

Keep up the good work!
Phodira

BF16 works. You need 96gb unified memory.
Rhystic1

BF16 works. You need 96gb unified memory.

Thank you, I was already aware of it but I have 64GB of memory.
jefflin

need install use this https://github.com/pawel-mazurkiewicz/ComfyUI-AppleSilicon-FP8 my case successfully with M4MAX 64GB using pruned_int8_convrot.safetensors
odnary

On my M3 Max (32GB), I got a VAE error as soon as I ran it.
Could you tell me which model and workflow you used?
BubbaHo
•
edited 2 days ago

@jefflin > need install use this AppleSilicon-FP8 my case successfully with M4MAX 64GB using pruned_int8_convrot.safetensors
@Rhystic1 > My error was "No backend can handle 'dequantize_int8_embedding': eager: q: dtype torch.bfloat16 not in {torch.int8}"

This was a ComfyUI bug, now fixed
You can 1- re-install/update Comfy, or 2- replace the suggested lines in ops.py in "~/ComfyUI-Installs/ComfyUI/ComfyUI/Comfy", or 3- use this fixed ops.py file: https://jumpshare.com/s/vOhaXPnRoIOOyRjGAkYu
The default T2V workflow gave me 200s/it to 500s/it for a 7 second duration at 0.2 megapixels (608x352). PAINFUL, but more testing improved things...

FIRST TESTS

  • AppleSilicon-FP8 extension is essential.
  • Comfy's "pruned int8_convrot" (21GB) is 22% faster than lilcheaty's "pruned_nvfp4"(12GB), despite being twice as large.
  • Spectrum adds an additional 18-20% speedup (with minor detail changes).
  • Combined (int8_convrot & spectrum) completed a real world test in 24 minutes VS 48 minutes (no spectrum & nvfp4).
    Could also try:
  1. Other Quants - https://huggingface.co/Abiray/MiniMax-H3-GGUF;
  2. An INT4 variant - https://huggingface.co/tsolful/Minimax_H3_INT4MixedConvRot.

EDIT 2 - TURBO TESTS

  • Update the Apple-Silicon-FP8 Nodes using Manager (manually select the latest version, it kept giving me an older one)
  • Get the MiniMax-H3 Turbo custom nodes from the manager
  • Get the latest Speedup Lora from Larryvrh
    Turbo with 6-steps is the time/quality sweet spot. The above 48 minute test is now completing in 18 minutes.
    All working on M1 series.

Sign up or log in to comment