mirror of
https://github.com/Comfy-Org/comfy-kitchen.git
synced 2026-11-04 23:37:42 +00:00
[GH-ISSUE #140] MiniMax ref2va Noise output with comfy-kitchen attention #45
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#45
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 @paju1986 on GitHub (Aug 27, 2026).
Original GitHub issue: https://github.com/Comfy-Org/comfy-kitchen/issues/140
Hi.
Im on Rocm 7.2 with Amd Radeon AI PRO r9700.
When i try to use comfy-kitchen attention i get just pure noise video when i push resolution to 0.9 megapixels and video duration of 15seconds.
But if i put 5seconds instead the video works fine. Something its causing it to output pure noise above certeing point of duration.
If i use normal pytorch attention the video works fine at 15seconds again.
Im using the v2v workflow for editing video.
Also Im using the fp8_scaled pruned model.
I don't get any log in the console af any kind of error, it looks like a normal generation.
Regards
@0xDELUXA commented on GitHub (Aug 28, 2026):
There are some changes in comfy-kitchen that aren’t in ComfyUI yet, so you could try building the latest commit from source or wait for a release.
Also, upgrade ROCm/PyTorch to the latest TheRock nightly.
@kulminaator commented on GitHub (Aug 28, 2026):
@paju1986 - which fp8 model are you exactly using?
@paju1986 commented on GitHub (Aug 29, 2026):
Im using: https://huggingface.co/Comfy-Org/MiniMax-H3/blob/main/diffusion_models/minimax_h3_ref2va_pruned_fp8_scaled.safetensors
By the way i switched to the int8 version and this one works. so it looks like it has something to do with being fp8
@paju1986 commented on GitHub (Sep 2, 2026):
updated pytorch to version ro 2.14-rocm7.14 the problem seems worse now even the int8 gives me garbage ouput
[INFO] Checkpoint files will always be loaded safely.
[INFO] Total VRAM 32624 MB, total RAM 63426 MB
[INFO] pytorch version: 2.14.0+rocm7.14
[INFO] Set: torch.backends.cudnn.enabled = False for better AMD performance.
[INFO] AMD arch: gfx1201
[INFO] ROCm version: (7, 14)
[INFO] Set vram state to: NORMAL_VRAM
[INFO] Device: cuda:0 AMD Radeon AI PRO R9700 : native
[INFO] Device: cuda:1 AMD Radeon AI PRO R9700 : native
[INFO] Using pytorch attention
[INFO] Using Comfy Kitchen attention
[INFO] aimdo: /project/src/control.c:276:INFO:comfy-aimdo inited for GPU: AMD Radeon AI PRO R9700 (VRAM: 32624 MB)
[INFO] aimdo: /project/src/control.c:276:INFO:comfy-aimdo inited for GPU: AMD Radeon AI PRO R9700 (VRAM: 32624 MB)
[INFO] DynamicVRAM support detected and enabled
[INFO] Python version: 3.12.12 | packaged by Anaconda, Inc. | (main, Oct 21 2025, 20:16:04) [GCC 11.2.0]
[INFO] ComfyUI version: 0.34.3
[INFO] comfy-aimdo version: 0.4.15
[INFO] comfy-kitchen version: 0.2.31
[INFO] comfyui-frontend-package version: 1.49.6
[INFO] comfyui-workflow-templates version: 0.11.54
[INFO] comfyui-embedded-docs version: 0.5.10
[INFO] comfy-kitchen version: 0.2.31
[INFO] comfy-aimdo version: 0.4.15
@kulminaator commented on GitHub (Sep 3, 2026):
Do you use any kind of turbo lora too?
Without any turbos i see already 0.8 taking immense amounts of time. I have noticed some output issues when the machine has ran into severe memory issues (due to the size of resolution + length). But in my sample the video just has cut out to black past some point.
I upgraded to rocm10 on the weekend. I have not spotted this issue since but i have also not played around that much.
@paju1986 commented on GitHub (Sep 3, 2026):
Hi.
I realized this problem also happends with Krea2, but its working fine for a while then suddently it starts to output noise.
If i do a vram clean it seems to go away
@asagi4 commented on GitHub (Sep 3, 2026):
Happens here too, with Krea2 and Klein 9b int8 quants. fp16 seems usually fine for klein9b at least.
I don't think it's ck-attn specific, but it's definitely a problem with ComfyUI somewhere.
@paju1986 commented on GitHub (Sep 9, 2026):
Sorry you are correct i tried without ck attention and the problem persists, must be something else in comfyui
@paju1986 commented on GitHub (Sep 13, 2026):
Upgrading pytorch to 2.15 with rocm 10.2 fixed the issues for me
@0xDELUXA commented on GitHub (Sep 13, 2026):
I see. Then I think this issue can be closed, as issues opened with older ROCm versions won't be fixed retroactively.