Installing Pytorch CUDA on Manjaro

1

Hi!
I’m failing to install pytorch with cuda support on a clean Manjaro install. (Lysia, 20.0.3: kernel is 5.6.15-1-MANJARO)
The GPU is a GTX1080ti, CUDA is 10.2.89, and I have tried two ways of installing torch. The first was, as the documentation specifies: pip install --user torch torchvision. I got this through selecting [Stable, linux, pip, python, cuda 10.2] on the “getting started” guide. I included the --user flag, since installing python packages as root is discouraged on the Archwiki.

The other method I tried was with pacman, installing the python-pytorch-cuda package.
On both cases, the result of the torch.cuda.is_available() function call is False.

I’ve installed cuda-10.2.89-5 through pacman, with the opencl nvidia-340xx drivers. Could this be a problem?
I’ve read this post saying that any distro is fine for pytorch, and the “Getting started” guide lists Arch as the first linux on the list to use pytorch.

I have found no information concerning this problem at all, that’s why I’m writing here. The fact that I haven’t found any post on this makes me think that my setup is wrong, and this is not a bug, but I don’t know where else to look. Would anyone happen to have a solution to this? or maybe a suggestion into what I should be looking at?
nvtop shows the gpu working well.
Here’s the output of nvcc --version:

nvcc: NVIDIA (R) Cuda compiler driver
Copyright (c) 2005-2019 NVIDIA Corporation
Built on Wed_Oct_23_19:24:38_PDT_2019
Cuda compilation tools, release 10.2, V10.2.89

This is the output of torch.cuda.current_device()

AssertionError: 
The NVIDIA driver on your system is too old (found version 10010).
Please update your GPU driver by downloading and installing a new
version from the URL: http://www.nvidia.com/Download/index.aspx
Alternatively, go to: https://pytorch.org to install
a PyTorch version that has been compiled with your version
of the CUDA driver.

This error was reported for cuda 9.0, but it was closed, and I’m using 10.2.

Thanks!

2

For CUDA10.2.89, you would need a driver >=440.33 as given in this table.
Could you try to update the driver and reinstall PyTorch?

1 Like
3

It worked! I had to uninstall all video drivers, cuda-dependent packages, and reinstall the video-nvidia-440xx package. I did so with Manjaro’s auto-install tool: https://wiki.manjaro.org/index.php?title=Manjaro_Settings_Manager#Hardware_Detection

I think the problem was installing the opencl-nvidia package. This installs the 430xx nvidia drivers by default. I don’t remember when I did this, but it had to do with nvidia-docker, for one of my last projects.

Thanks for the tip!

1 Like