Kijai

/

MiniMax-H3-experimental

like 140

upgrade your pytorch version to 2.11+

Int8 conv = 53 sec per step
w4a8 = 146 sec per step
am i doing something wrong??

comfy v0.31.0
comfy-kitchen 0.2.28
pytorch : 2.9.1 cu130

What is your system specs?
Im in the same boat. Int8 with all the speed shenanigans takes ~47s/it while w4a8 takes ~150+s/it.

Turns out my RTX 3060 6GB cant do w4a8 mixed precision very well (and my vram is painfully low) and the math overhead to make it work is large (i only have 16gb ram) so it hits my pagefile and gen times tank hella.

Int8 conv = 53 sec per step
w4a8 = 146 sec per step
am i doing something wrong??

comfy v0.31.0
comfy-kitchen 0.2.28
pytorch : 2.9.1 cu130

Are you using a lora? that messed with speed a lot for me. Like it did not like the turbo lora at all.

issue #18, will be fixed later