quantizations coming soon?
please consider quantizing this models for fp8, int8 and nvfp8 format. The demo is very promising.
Hello. Yes, we plan to add them soon.
There is a bunch of options here to explore: https://huggingface.co/docs/diffusers/en/api/quantization. I think you could also do it.
If you need NVFP4: https://pytorch.org/blog/faster-diffusion-on-blackwell-mxfp8-and-nvfp4-with-diffusers-and-torchao/
Thank you. We will try it.
int8_convrot would be great for ComfyUI. 😉
MiniMax-like Day 0 ComfyUI support could have been even more ideal (to be compatible with all the current and future optimizations/community plugins) though
And native ComfyUI code makes the checkpoint compatible with the official Comfy-Org/comfy-quants package, through which fp8/int8-convrot/nvfp4/... quantization is a ready-made breeze
Made the quantization of the Lite model using the code above (but with the kand6 branch)
https://huggingface.co/kabachuha/Kandinsky-6.0-Lite-5s-int8-convrot
In my experiments, it fully works and it's fast. (It's designed for the native integration, didn't test yet with the authors' nodes)
My Internet bandwidth is awful, so cannot really upload the big Pro file.
Update: int8 convrot Pro quantization works and tested with the fork. Still cannot upload
im currently downloading to GGUF the model for you guys