Closed ProtoBelisarius closed 4 months ago
Try to pip uninstall onnxruntime onnxruntime-gpu
and pip install onnxruntime
onnxruntime-gpu - is only usable for CUDA
ReActor installs onnxruntime if no CUDA is detected
Perhaps you have some other node that requires onnxruntime-gpu and that's why you have it installed
Try to
pip uninstall onnxruntime onnxruntime-gpu
andpip install onnxruntime
onnxruntime-gpu - is only usable for CUDA ReActor installs onnxruntime if no CUDA is detected Perhaps you have some other node that requires onnxruntime-gpu and that's why you have it installed
I refer to my provided additional information, and say that this wont help, I have the packages onnx onnxruntime onnxruntime-gpu installed. If I uninstall gpu it wont even load/import the node when starting comfy.
Using the tip from this comment: https://github.com/Gourieff/comfyui-reactor-node/issues/119#issuecomment-1869038663 made it work, so the node works fine. Even with -gpu installed. It just doesnt switch correctly/at all. idk
onnxruntime onnxruntime-gpu It just doesnt switch correctly/at all. idk
Because only one package of ORT must be installed, they won't work together properly (they use one folder "onnxruntime" inside the "site-packages" directory, the latest installed overwrites some files of the previous one)
Using the tip from this comment: https://github.com/Gourieff/comfyui-reactor-node/issues/119#issuecomment-1869038663
Yes, this is one of the ways to solve it (for AMD), you forced ORT-GPU to use only CPU EP
Wasnt really solved, still needs a hacky patch by the user to work at all on RoCM.
if cuda is not None:
if cuda.is_available():
providers = ["ROCMExecutionProvider"] #["CUDAExecutionProvider"]
for other people stumbling upon this problem, if you change it to ROCMExecutionProvider, it works with an AMD GPU, at least on linux. (dont use window)
Go to the reactor_swapper.py file in the node directory ../ComfyUI/custom_nodes/comfyui-reactor-node/scripts/reactor_swapper.py and change line 29.
First, confirm
What happened?
Installed and Reinstalled and it still gets stuck when Im trying to run it. Insisting on wanting to use a cuda device instead of switching to cpu. How do I make it ignore the gpu and just use cpu?
Steps to reproduce the problem
basic default workflow + (try to) use ReActor with default settings after adding the node
Sysinfo
Linux(Arch), Firefox, RX 6800XT, nightly torch 2.3.0(rocm6), default nodes
Relevant console log
Additional information
Idk if relevant, but when I uninstall the onnxruntime-gpu it fails to import the node and throws this error:
`