PyTorch

welcome to the manual for installing the pytorch python library in Linux(Voidlinux) we should look here To install pytorch you should get right version for:

Nvidia driver
cuda or cudatoolkit or pytorch-cuda
cuDNN
pytorch
cuDNN

1. Find Versions

nvidia driver version

simple way is install the latest version of driver, and look at the output of dmsg to find the recommend version of nvidia driver, for me: GeForce 820m is nvidia390

cuda version

You must choose a cuda version that compatible with your driver, and with Compute Capability

[driver] check driver here
[Compute Capability] Look here for new or (legacy)[https://developer.nvidia.com/cuda-legacy-gpus] to get the Compute Capability.

for me: - the new cuda version that support my driver nvidia390 is CUDA 9.1 (9.1.85) - the Compute Capability for nvidia 820m is 2.1 firmi, that compatible with $v 2.1 < c u d a < 8.0$

so i will try with cuda 9.0 and cuda 8.0

cuDNN

go to cuDNN website, and search for cuDNN version that comatible with you version pf cuda that already get already, you can get something like that.

Download cuDNN v8.7.0 (November 28th, 2022), for CUDA 11.x
Download cuDNN v7.6.5 (November 5th, 2019), for CUDA 9.0 if you your cuda v9.0 you should use cuDNN v7.6.5

PyTorch

to get the right version for pytorch that conforme with cudatoolkit go to this website for me i get:

conda install pytorch==1.1.0 torchvision==0.3.0 cudatoolkit=9.0 -c pytorch

2. install

install `nvidia390`

sudo xbps-install -S nvidia390

cuda 10.1

wget -c https://developer.nvidia.com/compute/cuda/10.1/Prod/local_installers/cuda_10.1.105_418.39_linux.run

sudo sh cuda_10.1.105_418.39_linux.run --silent --toolkit --override --tmpdir=$PWD/dir3

TODO:

fixed gcc12 unssuprted by --override flag
install gcc12 by xbps

conda

Requirement:

python3.7.5
pytorch=1.3.1
torchvision=0.4.2
cudatoolkit=9.2.148
cudnn=7.6.3

conda create --yes -n env5 python=3.7
conda install \
        pytorch=1.3.1 \
        torchvision=0.4.2 \
        cudatoolkit=9.2.148 \
        cudnn=7.6.3 \
      -c pytorch

conda install \
        pytorch=1.1.0 \
        torchvision=0.3.0 \
        cudatoolkit=9.0 \
        cudnn=7.5 \
      -c pytorch

Linux docs