Comfyui not using GPU, how to fix this?
We are running in instance of Comfyui and it's using CPU and not GPU, using a 4090. How can this be fixed?
52 Replies
Check if your pod is not running in 0 GPU Mode
It is running on a 4090.
Yep
What's ur comfyui args
btw I just made better comfy ui template that not breaks on extensions instalation
#Better ComfyUI - CUDA12
I don't think that GPU utilization metric is 100% accurate and it also not real-time, better to use
nvidia-smi to check GPU utilizationWhen using the Better Comfy Setup and I connect to port 7777 it says "GPU 0 NVIDIA GeForce RTX 4090"
Yes
thats normal... whats your point
I don't know. It's still using the CPU. I'm not sure where to look for the comfy args, other then using the log on Runpod, but there is nothing there that would indicate anything useful.
whats your workflow like
check the comfy docs
Why do you think it's using the cpu??
Workflow works fine locally
Because I have the crystools Nodepack and it shows it's using the CPU
And it takes forever. On my local 4090 it's done in 10 seconds.,
create a new pod maybe that pod has a problem?
Already tried that
report it to the contact support if it does
Do you see anything in the ComfyUI logs?
This looks fine
Not sure whats with xformers not being available tho, but that shouldn't cause it not to use GPU.
It could impact generation times tho.
Does your local have xformers installed?
"pytorch version: 2.3.1+cu121" this could be a problem for xformers, there is no compatible xformers for torch 2.3.1.
try your local files to a pod if you want to replicate it
and install the same packages as your local
can't install xformers if its installed on local
have to reinstall torch
only 2.3.0 is compatible with xformers not 2.3.1
I don't think my local instance has xformers installed but I'm not sure. it's not in the start up file.
activate local venv and then:
then you can compare the versions
otherwise also
wow didn't know pip freeze exists its cool hahahah
Once I can find terminal in this new backend you got it
Huh?
which backend did you mean
It's not a jupyter notebook so finding the terminal is more difficult. THe port 7777
code-server
ohh
try to screenshot the left menu bro
maybe i can help pointing it out
No other bottom ones?
i thought there are like 3-4 buttons on the bottom left?
try pressing the bottom left blue one
search command or cli there or terminal
maybe full screen screenshot?
try exploring the left side menu section then hahah
haha I've been trying XD
Menu -> Terminal -> New Terminal
thank youuuu
where is the files button btw
I meant menu not files
oh
that top left one?
torch is 2.3.1
hamburger icon
how bout on your local?
I'm trying. It's being obnoxious.
whats up?
Trying to use stuff like "pip3 show torch" in the terminal and it's just throwing errors.
what errors?
In just the basic terminal in C it's saying it's not found but I'm pretty sure I would have to have it to make comfy work and CUDA to do it's thing.
pip list
try that
maybe your torch is in env but you didn't activate it yet?
New verison of pip is available so I have pip installed.
xformers is not need as pytorch2.0+ has already better memory optimization
btw my template requires cuda 12.1+