[GH-ISSUE #75] Int8 slower on GTX 1660 Super after 0.2.19 #25

New issue

Closed

opened 2026-10-26 22:19:18 +00:00 by bsakaguchi · 2 comments

bsakaguchi commented

2026-10-26 22:19:18 +00:00

Owner

Originally created by @bakapotatolord on GitHub (Jul 14, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/75

So I updated ComfyUI which pushed comfy-kitchen to 0.2.19 and tested Krea2 Int8 ConvRot.

It seems to have slowed down the generation by nearly 3 times, went from ~6 s/it to ~18 s/it.

I manually rolled back the package version to 0.2.18 and it's back to ~6 s/it.

My GPU is GTX 1660 Super

Originally created by @bakapotatolord on GitHub (Jul 14, 2026). Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/75 So I updated ComfyUI which pushed comfy-kitchen to 0.2.19 and tested Krea2 Int8 ConvRot. It seems to have slowed down the generation by nearly 3 times, went from ~6 s/it to ~18 s/it. I manually rolled back the package version to 0.2.18 and it's back to ~6 s/it. My GPU is GTX 1660 Super

bsakaguchi closed this issue

2026-10-26 22:19:18 +00:00

bsakaguchi commented

2026-10-26 22:19:19 +00:00

Author

Owner

@comfyanonymous commented on GitHub (Jul 14, 2026):

Should be fixed now can you confirm?

@comfyanonymous commented on GitHub (Jul 14, 2026): Should be fixed now can you confirm?

bsakaguchi commented

2026-10-26 22:19:19 +00:00

Author

Owner

@bakapotatolord commented on GitHub (Jul 15, 2026):

Yeah, that seems to have fixed it. Thanks! Back to ~6 s/it. I assume the commit is for the other Turing GPUs since I don't notice any improvement in gen times prior to it?

One more thing, using lora with Krea2 Int8 ConvRot seems to double the gen time from ~6 to ~12 s/it. Using the custom node ComfyUI-INT8-Fast by BobJohnson24 seems to only increase it to ~8 s/it instead. This was an issue before the commit but thought to mention it now.

Should I raise that as a separate issue or?

@bakapotatolord commented on GitHub (Jul 15, 2026): Yeah, that seems to have fixed it. Thanks! Back to ~6 s/it. I assume the commit is for the other Turing GPUs since I don't notice any improvement in gen times prior to it? One more thing, using lora with Krea2 Int8 ConvRot seems to double the gen time from ~6 to ~12 s/it. Using the custom node ComfyUI-INT8-Fast by BobJohnson24 seems to only increase it to ~8 s/it instead. This was an issue before the commit but thought to mention it now. Should I raise that as a separate issue or?

bsakaguchi referenced this issue

2026-10-26 22:20:15 +00:00

[PR #25] [MERGED] Add arm64 runner #88