mirror of
https://github.com/Comfy-Org/comfy-kitchen.git
synced 2026-11-04 23:37:42 +00:00
[GH-ISSUE #209] [ERROR] Comfy Kitchen attention is unavailable. Install a Comfy Kitchen build with attention support to use --use-ck-attention. gfx1031 #61
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#61
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 @Tpal12 on GitHub (Sep 27, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/209
Hi, I am using:
Windows 10
Python 3.12.10
PyTorch : 2.13.0+rocm10.0.0
ComfyUI : 0.37.0
comfy-kitchen: 0.2.35
and a gfx1031 GPU
I tried using the
--use-ck-attentionflag but received the following message: "Comfy Kitchen attention is unavailable. Install a Comfy Kitchen build with attention support to use --use-ck-attention".I attempted to build comfy-kitchen with the
--hipand--archbuild parameters, but it doesn't seem to be installed or used correctly in my venv.What am I missing here?
@0xDELUXA commented on GitHub (Sep 28, 2026):
--use-ck-attentionwas never meant to support RDNA2, since it requires WMMA support, which RDNA2 doesn't have. Building comfy-kitchen for gfx1031 won't make this attention backend available. You'll need to use a different attention backend instead, e.g. SageAttention via Triton.@Tpal12 commented on GitHub (Sep 28, 2026):
Oh, then I misunderstood that part:
and non-WMMA kernels (quantizers, INT8 dequantizers, RoPE and the fused RMSNorm+RoPE, AdaLN and RMS-AdaLN, the AWQ GEMV) that also run on RDNA2. It does not link or call hipBLAS/hipBLASLt; every matmul is compiled from the sources in comfy_kitchen/backends/hip/.
What a shame but at least sage 2 is working fine
THX Mate
@0xDELUXA commented on GitHub (Sep 28, 2026):
RDNA2 was released before WMMA was a thing on AMD consumer GPUs. There isn't really a way to get better performance on it than Sage via Triton.
This issue can be closed.
@LuXuxue commented on GitHub (Sep 29, 2026):
triton-windows's readme said: "RDNA2 does not have matrix cores, but it has V_DOT4_I32_I8/V_DOT8_I32_I4 for fast int8/int4"
https://github.com/Comfy-Org/comfy-kitchen/pull/200 did some work to support rdna2, and i use ai writed a prototype support for gfx1035 (which can run almost as fast as sage triton on my gfx1035, but it may not working on other cards)
@0xDELUXA commented on GitHub (Sep 29, 2026):
I still don't really see what positive changes leaving this issue open would bring for ck-atten on RDNA2.
@Tpal12 commented on GitHub (Sep 29, 2026):
Ohh yeah sry.
Just one last question
Do in need a dedicated triton for windows installation and --enable-triton-backend flag for comfyui and sage attention2?
Or is comfy kitchen triton in venv\Lib\site-packages\comfy_kitchen\backends\triton a full package of triton?
Maybe a stupid question, since it looks like kernels only (the ones i missunderstood), but still ^^
@0xDELUXA commented on GitHub (Sep 29, 2026):
comfy-kitchen's Triton backend needs
pip install triton-windowsto work, but I think you're better off using Sage separately. Here's a great node for RDNA2 users btw: https://github.com/patientx/ComfyUI-INT8-Fast-ROCM.@Tpal12 commented on GitHub (Sep 29, 2026):
Yeah already using this great nodes :)
also:
-https://github.com/patientx/sageattention-autotune
-kjnodes sage attention patch with fp16-triton setting/mode
and --enable-triton-backend flag
soo is that a useful setup or overkill/missmatch cause the patientx nodes already doing this stuff in the background?
his readme is very helpul but i still try to find the combinations/ do's and dont's