Hugging Face: MFlux-Community

[ianscrivener]

With the GitHub glitch I spent last week writing a model conversion system and building out the HuggingFace MFlux model collection;

https://huggingface.co/mflux-community

MFlux Models: 198
Source Models: 27
Model Families: 9

All the models have been generated on CUDA. I tried various cloud backends - GitHub Actions (fails), Blacksmith (fails), Modal (expensive), RunPod (OK)... before settling on HuggingFace GPU as the best solution.

Now that MFlux is back online with GitHub, I'll reach out to HuggingFace to see if they can sponsor us a Team Plan and the GPU for ongoing model conversion.

I'll bundle the MFlux-Conv app soon and ppush to GH... though currently I have had to use my personal HF account... as Org accounts can't have GPU compute - a GH Space with a Docker container - unless they are on the GH Team Plan ($$).

BTW: MFlux CUDA can convert models (bf16 and quants)... and can generate images in BF16 only. Currently MFlux on CUDA throws an error trying to generate with quantized models... as a CUDA doesn't support Apple affine quantizations by fault. I believe there is a fix.

[ianscrivener]

@anthonywu @plz12345 @fxd0h

[plz12345]

What are the most important ones that are missing, at present?

[ianscrivener]

See #667

These recent models cannot be converted & quantized;

• Z-Image-Turbo-ControlNet-Union-2.1
• Microsoft Lens

As well as these older models;

• Flux.1-Dev-ControlNet-Canny
• Flux.1-Dev-ControlNet-Upscaler
• Flux.1-Dev-Depth
• Flux.1-Dev-Fill
• Flux.1-Dev-Kontext
• Flux.1-Dev-Fill-CatVTON
• Flux.1-Dev-Redux
• Flux.1-Schnell-ControlNet-Canny

Do these need to be converted & quantized?

• SeedVR2-3B
• SeedVR2-7B
• depth-pro

[ianscrivener]

All good now thanks to Anthony's #680

These models now have a full range of quants on HF;

1. Flux.1-Dev-ControlNet-Canny
2. Flux.1-Dev-ControlNet-Upscaler
3. Flux.1-Dev-Depth
4. Flux.1-Dev-Fill
5. Flux.1-Dev-Kontext
6. Flux.1-Dev-Fill-CatVTON
7. Flux.1-Schnell-ControlNet-Canny