Upload scaled float8_e4m3fn version for use with ComfyUI on lower end cards.
#8
by
silveroxides
- opened
I made a scaled fp8 variant using same method we use for Chroma now.
Comparison with original weights. Original first.
Hallo, can you show how to use the scaled float8_e4m3fn version in diffusers? What parameter should I set to used this safetensors ?
Thank You
@BeraVonSodom
As stated in title, this is targeting ComfyUI. No idea how to use it with diffusers if it is even possible.
Diffusers has its own method for quantization through optimum.quanto I believe