Trying to get comfyUI to run on a AMD Radeon AI Pro R9700 with 32 gb vram #11730
Replies: 16 comments 29 replies
|
ComfyUI crashes during model loading? In this case it is generally a problem with pagefile size. Try to increase it or set it to "managed by the system" |
|
Thanks for the idea, but it crashes at starting to do the job. At the clip text encode node. Or at the prompt schedule node. The models loads fine. |
|
Hm, i think my problem is much more fundamental. Since i haven't installed Rocm yet. I thought everything needed comes with Comfy Portable already. I can though just find advice to install Rocm for a venv and with cloning the repo. Which will not work with the portable. And this is even Linux. I am at Windows. https://rocm.docs.amd.com/en/7.10.0-preview/rocm-for-ai/pytorch-comfyui.html |
|
If you can wait, wait for at least ROCm 7.2. In 7.1.1. there are really nasty memory bugs and other stupid things. Even 6xx was more stable. I don't know why everyone is pushing ROCm support (under windows!) as "offical" - it's not even pre alpha. |
|
Ah this sounds not good :D Waiting, well i have some images to generate. Would it make sense to install a dual boot environment and use comfy under ubuntu? And what is the official steps to get comfy working together with my card under Windows? I would even go with an old Rocm version if this would make Comfy at least somehow functional again. |
|
Okay, the desktop version shows the same behaviour than the portable. It pauses and stops without any error message. |
|
Found windows 7.1.1 hip here: https://www.amd.com/en/resources/support-articles/release-notes/RN-AMDGPU-WINDOWS-PYTORCH-7-1-1.html?utm_source=chatgpt.com Trying. |
|
... and you need to install / compile pytorch wheels. Thats me finally out. I don't want to ruin my system. Geez AMD. Didn't realize that they are technically this far behind. Okay, now let's have a look if Linux is a solution. |
|
What driver are you using? Be sure to disable auto updating because once Adrenalin installs the standard driver Comfy won't work anymore. Good luck! |
|
Oh, this driver works for you? I have still the default driver applied. i didn't want to toy around with it. But now I will give it a try tomorrow. Many thanks :) |
|
The special driver was indeed the solution. I can now generate images with Z Image, and i can generate videos with Wan 2.2. More is not needed, for now. I uninstalled everything with Display Driver Uninstaller in safe mode. And have set up the new driver then. The speed increase of the AMD card is a bit disappointing compared to the 4060 Ti. A Wan 2.2 workflow that took just over 10 minutes with the old card now takes 8.5 minutes. I had expected more. But at least I now have enough RAM to avoid constantly running into an out-of-memory (OOM) situation. And Z Image runs four times faster. Here i have indeed a big speed boost. And so far it runs stable enough. So i think i will skip my plans with the dual boot system and trying Ubuntu. Maybe at a later point. Many thanks for the help :)
I cannot find out how. Adrenaline does not have a checkbox from what i can see. How do i? |
|
have you considered Docker? its great for reducing variables. let me know if it works for ye. |
|
Honestly this turns into a nightmare. I have now installed rocm and comfy on ubuntu. I don't see any errors. rocminfo shows everything correct. The device is there. Console looks okay. Comfy starts. But when it starts to calculate then Python crashes. Tested with z image and a wan image workflow. The console then shows a memory access error. What wonders me here is, i have rocm installed, i use the rocm comfyui version, but it still talks about Cuda? CLIP/text encoder model load device: cuda:0, offload device: cpu, current: cuda:0, dtype: torch.float16 In case somebody has a clue here, i am open to any suggestion. |
|
...you really should wait a few days, believe me I´ve spend hours for nothing: |



Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
Hey friends,
I have migrated from Nvidia to an AMD card. An AMD Radeon AI Pro R9700 with 32 gb vram. I am on Windows 11.
I have then downloaded the newest portable AMD version of ComfyUI, 0.8.1. But none of my workflows works anymore. Comfy simply quits. The content is never generated. And i don't get any error message. Just a Pess any Key to continue in the console.
I was able to reduce the problem to Python and Torch: Example error message:
0x00007FFA0F1AB520, G:\comfyamd\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\lib\torch_cpu.dll(0x00007FFA0F1A0000) + 0xB520 byte(s), ?count_nonzero@Tensor@at@@qeba?AV12@V?$optional@_J@std@@@z() + 0x30 byte(s)
0x00007FFA4832BA14, G:\comfyamd\ComfyUI_windows_portable\python_embeded\Lib\site-packages\torch\lib\torch_python.dll(0x00007FFA48000000) + 0x32BA14 byte(s), ?release@?$THPPointer@U_frame@@@@QEAAPEAU_frame@@xz() + 0x1C2BA4 byte(s)
0x00007FFAA41879E4, G:\comfyamd\ComfyUI_windows_portable\python_embeded\python312.dll(0x00007FFAA4150000) + 0x379E4 byte(s), PyThread_acquire_lock_timed() + 0x5F0 byte(s)
0x00007FFAA41D2018, G:\comfyamd\ComfyUI_windows_portable\python_embeded\python312.dll(0x00007FFAA4150000) + 0x82018 byte(s), PyObject_Vectorcall() + 0x688 byte(s)
The bat that i have used to end at least in an error message was made by ChatGPT:
@echo off
REM -----------------------------------------------
REM 1. Umgebungsvariablen für AMD / DirectML stabil setzen
set TORCH_DML_DISABLE_GRAPH_FUSION=1
set TORCH_DML_ENABLE_DYNAMIC_TENSORS=0
REM -----------------------------------------------
REM 2. Python-Pfad auf Embedded Python
set PYTHON_EXE=.\python_embeded\python.exe
REM 3. ComfyUI Pfad
set COMFY_MAIN=ComfyUI\main.py
REM 4. ComfyUI starten mit sicheren Flags
%PYTHON_EXE% %COMFY_MAIN% --windows-standalone-build --force-fp32 --disable-smart-memory --dont-upcast-attention
pause
So here my question, how can i get ComfyUI to work with my AMD card? What is your advice?
Kind regards
Reiner
All reactions