pyenv update
pyenv install -l
pyenv install 3.10.8
pyenv global 3.10.8
Show CUDA version:
nvidia-smi
./webui.sh --xformers
source venv/bin/activate
pip install xformers==0.0.12
./webui.sh --xformers
pip uninstall xformers==0.0.12
pip install git+https://github.com/facebookresearch/xformers.git#egg=xformers
./webui.sh --xformers
pip install cutlass
pip uninstall git+https://github.com/facebookresearch/xformers.git#egg=xformers
pip install git+https://github.com/facebookresearch/xformers.git#egg=xformers
./webui.sh --xformers
Go to https://developer.nvidia.com/cuda-gpus and expland "CUDA-Enabled GeForce and TITAN Products" section and find Compute Capability value for Geforce RTX 2060.
pip uninstall git+https://github.com/facebookresearch/xformers.git#egg=xformers
export TORCH_CUDA_ARCH_LIST=7.5
export FORCE_CUDA=1
pip install git+https://github.com/facebookresearch/xformers.git#egg=xformers
https://docs.nvidia.com/cuda/cuda-installation-guide-linux/index.html#ubuntu-installation
AUTOMATIC1111/stable-diffusion-webui#3525
wget https://developer.download.nvidia.com/compute/cuda/repos/$distro/$arch/cuda-keyring_1.0-1_all.deb
wget https://developer.download.nvidia.com/compute/cuda/repos/ubuntu2004/x86_64/cuda-keyring_1.0-1_all.deb
sudo dpkg -i cuda-keyring_1.0-1_all.deb
sudo apt-get update
sudo apt install cuda-11-7
Note: apt install cuda
would have installed 11.8 so, I think I should have used that rather than an explicit version. Installing cuda also installs/updates the drivers to the corresponding version.
When you redo the export
and pip install
steps after installing cuda
the wheel building step goes from taking a minute or so to taking a serious amount of time, i.e. tens of minutes (watch it work away with htop
).
See https://forums.developer.nvidia.com/t/fixed-suspend-resume-issues-with-the-driver-version-470/187150/2 https://forums.developer.nvidia.com/t/occassional-failure-to-resume-pci-pm-suspend-nv-pmops-suspend-0x0-0x20-nvidia-returns-5/187091
less /var/log/syslog
sudo systemctl disable nvidia-suspend.service
sudo systemctl disable nvidia-hibernate.service
sudo systemctl disable nvidia-resume.service
sudo vim /etc/modprobe.d/nvidia-power-management.conf
Changing NVreg_PreserveVideoMemoryAllocations=0
from 0 to 1 did allow it to suspend but means Blender in Cycles mode doesn't come back properly on waking.
Try setting it back to 1 and rebooting - maybe it just needs to be power cycled in a "clean" state.
And is there a package I should be installing to provide these services?
Question: https://forums.developer.nvidia.com/t/what-apt-package-contains-nvidia-suspend-service-etc/233632
Some comments on this video on enabling the Dreambooth Automatic1111 extension suggest that you need to manually install https://pypi.org/project/accelerate/