stable-diffusion-webui icon indicating copy to clipboard operation
stable-diffusion-webui copied to clipboard

Published 20 hours ago • verified publisher icon AUTOMATIC1111

Question/Help: GPU is too old

Open impetus0 opened this issue 3 years ago • 7 comments

The console returned a pretty cut and dry error: Found GPU0 NVIDIA GTX TITAN Black which is of cuda capability 3.5. PyTorch no longer supports this GPU because it is too old. The minimum cuda capability supported by this library is 3.7.

Is buying a new GPU the only option here or is there some workaround?

impetus0 avatar Oct 17 '22 22:10 impetus0

You have to manually install PyTorch from source as the PyPI packages don't include the old CUDA binaries to save space: https://github.com/pytorch/pytorch/blob/master/README.md#from-source

Sources:

https://discuss.pytorch.org/t/solved-pytorch-no-longer-supports-this-gpu-because-it-is-too-old/15444

https://discuss.pytorch.org/t/pytorch-no-longer-supports-this-gpu-because-it-is-too-old/13803

https://discuss.pytorch.org/t/solved-pytorch-no-longer-supports-this-gpu-because-it-is-too-old/15444/30?page=2

Though if the source version doesn't support your CUDA version, then you need a new GPU.

Edit:

Cuda 3.5 might be no longer supported at all: https://discuss.pytorch.org/t/pytorch-1-2-0-with-compute-capability-3-5/53916/4

ProGamerGov avatar Oct 17 '22 23:10 ProGamerGov

Thanks, albanD in your last link does say 3.5 is no longer supported, and that was back in Aug 2019. My TITAN card aged fast it feels like!

impetus0 avatar Oct 18 '22 01:10 impetus0

for running on CPU-only:

(webui-user.bat)

set COMMANDLINE_ARGS=--precision full --no-half --use-cpu all --skip-torch-cuda-test --autolaunch set CUDA_VISIBLE_DEVICES=-1

ghost avatar Oct 18 '22 20:10 ghost

Running on CPU is a good idea, thanks! Adding those lines to webui-user.bat gave me this error on run

Screenshot 2022-10-18 182752

impetus0 avatar Oct 18 '22 22:10 impetus0

Do you have the latest version? I mean I installed fresh just yesterday and "all" is a valid option, unless it's been removed since yesterday which seems unlikely.

ghost avatar Oct 18 '22 22:10 ghost

Sorry, forgot to git pull update. It works! Thanks a bunch. Are there any pros/cons to using CPU vs GPU to generate images? CPU utilization is only 60% during generation, so that's good.

impetus0 avatar Oct 19 '22 01:10 impetus0

you re welcome. my cpu was100% used and getting hot so i throttled at75% using power settings, now its cool i can leave it overnight. aside from speed (5-6 minutes at 20steps) and half precision not being supported, all is well

ghost avatar Oct 19 '22 09:10 ghost