-
Notifications
You must be signed in to change notification settings - Fork 6.5k
New issue
Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.
By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.
Already on GitHub? Sign in to your account
--use-sage-attention flag causing CUDA illegal memory access #6125
Comments
What's your GPU? can you post the full log? |
On RunPod RTX4090, full log: /workspace/ComfyUI# python main.py --use-sage-attention --port 4444 ComfyUI-Manager: installing dependencies done.** ComfyUI startup time: 2024-12-20 00:31:07.073952 Prestartup times for custom nodes: Total VRAM 24118 MB, total RAM 1031845 MB Loading: ComfyUI-Manager (V2.55.5)ComfyUI Version: v0.3.8-4-g3cacd3f | Released on '2024-12-19'[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/model-list.json
|
Same problem. Python crashes with --use-sage-attention every time. And no problem without it. |
Do you get the same crash with sage attention 1? |
Sageattention 1.0.6 works as expected python main.py --use-sage-attention --port 6006 ComfyUI-Manager: installing dependencies. (GitPython)WARNING: Running pip as the 'root' user can result in broken permissions and conflicting behaviour with the system package manager, possibly rendering your system unusable.It is recommended to use a virtual environment instead: https://pip.pypa.io/warnings/venv. Use the --root-user-action option if you know what you are doing and want to suppress this warning. [notice] A new release of pip is available: 24.2 -> 24.3.1 ComfyUI-Manager: installing dependencies done.** ComfyUI startup time: 2024-12-20 13:50:22.608414 Prestartup times for custom nodes: Traceback (most recent call last): ComfyUI-Manager: installing dependencies done.** ComfyUI startup time: 2024-12-20 13:50:34.786604 Prestartup times for custom nodes: Total VRAM 24111 MB, total RAM 176863 MB Loading: ComfyUI-Manager (V2.55.5)ComfyUI Version: v0.3.8-7-g418eb70 | Released on '2024-12-20'[ComfyUI-Manager] default cache updated: https://raw.githubusercontent.com/ltdrdata/ComfyUI-Manager/main/custom-node-list.json
|
I'm using sage attention 2.0 too. This might be the reason? |
sageattention 1 works, but sageattention2 brings speed improvements, so hoping this issue is resolvable for 2. |
Yes, sageattention 1 works, I've got a speed improvement from about 3.3s/it to 2.75s/it. hope for 2 support. |
Same here, work well with the Kijai's wrapper, hope for v2 support in comfy |
I have sageattention installed, and can create clear videos using the Kijai nodes, but if I try to use comfy nodes I always get a very blurry image. It doesn't crash or error, but I'm using the same model/clip/vae in both workflows. Edit: I just had the same crash and error with Kijai's nodes after bumping up the number of frames from 97 to 145. Could this be a memory related issue? |
Expected Behavior
Expected Hunyuan video output, but ComfyUI crashes instead.
Actual Behavior
Hunyuan Wrapper from Kijai working fine, but for some reason sageattention isn't working correctly with comfyUI
Steps to Reproduce
Runpod RTX4090 with python 3.11/cu124 template:
Install torch for cu124
Install sage attention 2.0.0
start ComfyUI using "python main.py --use-sage-attention"
Use the ComfyUI Example workflow
ComfyUI crashes at "SamplerCustomAdvanced" node
Debug Logs
Other
No response
The text was updated successfully, but these errors were encountered: