silveroxides

/

FLUX.2-dev-fp8_scaled

like 73

Great work!

I wanted to enquire about how you do fp8 scaled conversion and if you know fp8 matmul as well.

thank you

I currently use my own WIP module. Not everything works though.
ofc I know about fp8 matmul.

oops forgot link https://github.com/silveroxides/convert_to_quant

Does it make sense to download the int8 version if I have an RTX 4090, and will I get any performance boost?

Does it make sense to download the int8 version if I have an RTX 4090, and will I get any performance boost?

No it makes no sense at all haha. You have a 4000 series card. You want fp8 and if you want to ensure fast generation speed, use the --fast fp8_matmul launch argument in ComfyUI

Does it make sense to download the int8 version if I have an RTX 4090, and will I get any performance boost?

No it makes no sense at all haha. You have a 4000 series card. You want fp8 and if you want to ensure fast generation speed, use the --fast fp8_matmul launch argument in ComfyUI

there no fp8_matmul flag in comfyui

Does it make sense to download the int8 version if I have an RTX 4090, and will I get any performance boost?

No it makes no sense at all haha. You have a 4000 series card. You want fp8 and if you want to ensure fast generation speed, use the --fast fp8_matmul launch argument in ComfyUI

there no fp8_matmul flag in comfyui

it's --fast fp8_matrix_mult