e5m2 for T2V & DF 540P SkyReels missing, attempted my own conversion, some direction would be appreciated
I have a 3090, and I tried downloading the others, but they won't run. Could you please include the 540P versions for T2V please?
Ah, after reading another thread I saw you said it was a torch compile issue. So I disabled my torch compile node and I can now get the 540P versions running. It's significantly slower though, by like 20-30%
Hey @Kijai ! Figured you're busy creating the next awesome node we'll all be thanking you for, so I took a crack at converting the 540p SkyReels models to FP8-E5M2 myself.
I analyzed your existing FP8 models to see which layers you quantized and which you kept as FP32, and I replicated that pattern. For the actual conversion of the FP32 weights to E5M2, I used PyTorch's direct tensor conversion: original_tensor.to(torch.float8_e5m2).
The E5M2 models work (tested with ComfyUI + CausVid) and torch.compile gives the expected speedup, which is great! However, the output quality seems noticeably lower than your E4M3FN versions.
This is my first try at ever attempting something as such. I'm guessing the difference might be down to how scaling factors are handled. My direct .to() conversion relies on PyTorch's default implicit scaling. Did you use a specific method, tool, or explicit scaling factors when you created your E4M3FN versions that might have helped minimize the quality loss? Any pointers or suggestions on how I might improve the E5M2 quality would be hugely appreciated!
If anyone wants to try out these E5M2 versions (and maybe compare quality), they're here:
https://huggingface.co/phazei/phazei-SkyReels-V2-fp8-e5m2/tree/main
Thanks again for all your work for the community!