mirror of
https://github.com/Comfy-Org/comfy-kitchen.git
synced 2026-11-04 23:37:42 +00:00
[GH-ISSUE #16] RTX30xx FP8 feather matmul triton kernel? #9
Labels
No labels
pull-request
No milestone
No project
No assignees
1 participant
Notifications
Due date
No due date set.
Dependencies
No dependencies set.
Reference
Comfy-Org/comfy-kitchen#9
Loading…
Add table
Add a link
Reference in a new issue
No description provided.
Delete branch "%!s()"
Deleting a branch is permanent. Although the deleted branch may continue to exist for a short time before it actually gets removed, it CANNOT be undone in most cases. Continue?
Originally created by @phazei on GitHub (Jan 7, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/16
I saw this was released https://github.com/SuriyaaMM/feather not too long ago.
I was wondering, is that something that would ideally be implemented here? Would it basically speed up all fp8 operations presuming the operation was completed long before the memory was copied from VRAM to the registers? Seems like it's in line with the kernels here.
@Ph0rk0z commented on GitHub (Jan 7, 2026):
That kernel doesn't work but it would be REALLY REALLY nice is comfy were to let us old gpu peasants have a fallback path that allows casting to FP16 or patched triton for compiles. Maybe even borrow from https://github.com/silveroxides/ComfyUI-QuantOps and support int8 weights.
@silveroxides commented on GitHub (Jan 7, 2026):
@Ph0rk0z Thanks for the mention. That int8 implementation is actually from a PR on ComfyUI repo. https://github.com/Comfy-Org/ComfyUI/pull/10864
Just thought I would mention that. Have you tried any int8 models with my custom node? How was the inference speed?
@Ph0rk0z commented on GitHub (Jan 8, 2026):
Tried chroma, self quatized. Speed in triton mode is decent but it fails to compile on 2080 and compiles then crashes second gen on 3090. I do need torch.compile to work though, because that pushes me into it/s territory from s/it.
I also did T5XXL and int8 is good. I am using it over FP8. Unfortunately in triton mode (where there is noticeable speedup) produces NaN in the embedding causing sprinkled static image.
Honestly would prefer to be able to change all my models from FP8 -> int8. The fidelity is like GGUF without the lora slowdown. In torch 2.9, I can only compile plain FP8 E4 so miss out on the better scaled quantizations.The turning card ends up being the image-gen GPU while 3090s mainly run some LLM and have had to try everything under the sun to make it fast.
@woct0rdho commented on GitHub (Jan 24, 2026):
I've tried to make it run fully on GPU and add it as ComfyUI ops, see https://github.com/woct0rdho/ComfyUI-FeatherOps , and the discussion in https://github.com/SuriyaaMM/feather/issues/1 . It surely speedups GEMV, but in my tests it cannot speedup large GEMM in Z-Image and larger models.
@monnky commented on GitHub (Mar 13, 2026):
I use RTX 3060 Nvidia Ampere //
why triton always disabled ?? how to enable ?? @comfyanonymous
Found comfy_kitchen backend triton: {
'available': True,
'disabled': True,
'unavailable_reason': None,
@silveroxides commented on GitHub (Mar 13, 2026):
Merge this PR to your local installation and set the launch argument --enable-triton-backend
But you only get any use from it if you use a comfy-kitchen that actually has triton backend implementation for supported matmul.
https://github.com/Comfy-Org/ComfyUI/pull/12730