Tensor needs to be moved from one gpu to the other. about stable-diffusion HOT 5 CLOSED

When is this happening? Are you using gfpgan/esrgan as an option on generation, or in their standalone tabs?

gpu 1 is the card you're using for gfpgan/esrgan and gpu 0 is the card for the main model, correct?

Have you tried explicitly setting --gpu option as well as the other options? there might be something messed up with how the selected gpus are applied

from stable-diffusion.

The problem seems to be only with GFPGAN. It throws the error in both option on generation as well as in standalone tab mode.

esrgan is working correctly.

Your assumption is correct: gpu 1 for gfpgan/esrgan and gpu 0 for main model. I have now tried both with explicit gpu set and without:
python scripts/webui.py --optimized-turbo --gpu 0 --extra-models-gpu --esrgan-gpu 1 --gfpgan-gpu 1
python scripts/webui.py --optimized-turbo --extra-models-gpu --esrgan-gpu 1 --gfpgan-gpu 1

I tried the non-optimized variant as well, same result (esrgan works, gfpgan doesn't)
python scripts/webui.py --gpu 0 --extra-models-gpu --esrgan-gpu 1 --gfpgan-gpu 1

!!Runtime error (txt2img)!! Expected all tensors to be on the same device, but found at least two devices, cuda:0 and cuda:1! (when checking argument for argument weight in method wrapper__cudnn_convolution) exiting...calling os._exit(0)

Thank you for sharing your interface, it helps a lot.

from stable-diffusion.

Traceback led me to a facexlib problem.
Someone else also mentioned it here: https://github.com/AbdBarho/stable-diffusion-webui-docker/issues/4#issuecomment-1231927656

I guess the only way would be to manually patch facexlib or move GFPGAN to cpu only in case you don't want it on GPU.

from stable-diffusion.

Ok, fixed this way:

conda activate ldm
pip uninstall facexlib
git clone https://github.com/xinntao/facexlib
cd facexlib
git pull origin pull/19/head
python setup.py install
cd ..
python scripts/webui.py --optimized-turbo --gpu 0 --extra-models-gpu --esrgan-gpu 1 --gfpgan-gpu 1

And everything is working as intended.

from stable-diffusion.

I've forked, merged that pull request into the fork, and updated the environment.yaml so should be fixed for all now, thanks!

from stable-diffusion.