DLL load failed while importing cuda_utils #19
Comments
|
I even tried it on a separate clean Comfyui build, same error. ComfyUI Error ReportError Details
Stack TraceSystem Information
Devices
LogsAttached WorkflowPlease make sure that workflow does not contain any sensitive information such as API keys or passwords. Additional Context(Please add any additional context or steps to reproduce the error here) |
|
Triton/sageattention install is lot more involved than just installing the pip packages/wheels. You are missing some part of the required tools for compilation (visual studio + cuda-toolkit) or something in their configs are not correct. You can refer to this guide https://purz.notion.site/Get-Windows-Triton-working-for-Mochi-6a0c055e21c84cfba7f1dd628e624e97 Alternatively you if you can install flash_attn you don't need triton/sageattention (but it will be slower). |
|
I followed that guide to install triton / sageattention in both of these setups. |
Something is wrong with that still though, I've seen that error fixed by clearing the . triton cache, which can be found at C:\Users\username\.triton |
|
Clearing the triton cache resulted in a new error. ComfyUI Error ReportError Details
Stack TraceSystem Information
Devices
LogsAttached WorkflowPlease make sure that workflow does not contain any sensitive information such as API keys or passwords. Additional Context(Please add any additional context or steps to reproduce the error here) |
|
I even tested Mochi with sageattention. It works but this does not. |
|
"new error" also occurred for me |
|
https://github.com/woct0rdho/triton-windows/releases/tag/v3.0.0-windows.post1 |
|
Nope. I even uninstalled visual build tools, pip uninstalled all the requirements, then reinstalled visual build tools using what this shows to use reinstalled pytorch / cuda / cuda toolkit / all the requirements.txt / triton / sageattention, making sure all were for the same version of python / cuda that im using and that the PATHs were pointing to the right places. Still not working. Sage attention works for mochi though. But I get this error below when trying to use HunyuanVideoWrapper: ComfyUI Error ReportError Details
Stack TraceSystem Information
Devices
LogsAttached WorkflowPlease make sure that workflow does not contain any sensitive information such as API keys or passwords. Additional Context(Please add any additional context or steps to reproduce the error here) |
|
Someone else had this issue too, not sure what exactly finally fixed it but at one point they had Triton 3.0.0, so make sure it's Triton version 3.1.0 and sageattention version 1.0.6 (or the new 2.0.0 beta) |
|
I fixed this issue by deleting the Triton cache in Users/Triton, but after that, I encountered an error saying that my GPU is too old. I have a Quadro RTX 6000 with 24GB of VRAM. |
|
Same problem here. Edit: Seems like a combination of installing cuda 12.4, copying libs and include folder of python 3.12 to python_embeded (instead of python 3.11.9) and deleting the .triton cache folder again fixed it. Edit: Nope, python 3.12 files don't work in the end, it throws a bunch of other errors. Last edit: Fixed it, this wrapper only works with python 3.11, at least for the 3090. Have to download a really old comfyui version with the 3.11 python_embeded files and drop the content in the embeded folder of my comfy install so comfy uses python 3.11.9 files, then install sage and triton again and that's it. |
|
Same problem here. |
|
|
|
I tried for days to get this to work. It never did. Got the same Triton/Sageattention/CUDA DLL errors everyone has reported. Comfy now has Hunyuan support baked in. I can confirm it works. Here's what I did:
That's pretty much it. It now runs without a hitch. If you're still trying to get Hunyuan to run and are fighting the Triton/Sageattention/CUDA DLL errors, you can safely give up that merry go round and get it working. |
You know you could just not use sageattention or torch.compile...? Then you don't need to do any of that as it's only to make things faster, not a requirement for either the wrapper or Comfy native implementation. ComfyUI natively now finally also has support for sageattention, but to run it with that, it's the same hoops to jump through. |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
I have triton and sageattention installed without errors yet I still just get this error when trying to inference.
ComfyUI Error Report
Error Details
Stack Trace
System Information
Devices
Logs
Attached Workflow
Please make sure that workflow does not contain any sensitive information such as API keys or passwords.
The text was updated successfully, but these errors were encountered: