How can I use GPU hardware ?

Hi @romain-keramitas-prl – thanks for reporting.

What OS, CPU and GPUs are you trying to use? And which library are you trying to install/run that uses GPUs? That would help use test your use case.

Hey @loreto !

I'm running on Ubuntu 20.04, with consumer CPUs / GPUs.

The CPU is an Intel® Core™ i7.
The GPU is a NVIDIA GeForce RTX 2060, with relatively recent drivers.

I'm trying to install and use Pytorch. I've created the following devbox configuration:

{
  "packages": [
    "python310",
    "poetry",
  ],
  "shell": {
    "init_hook": [
      "poetry shell",
      "export PKG_CONFIG_PATH=`pwd`/.devbox/nix/profile/default/lib/pkgconfig/",
      "export LD_LIBRARY_PATH=`pwd`/.devbox/nix/profile/default/lib/"
    ]
  },
  "nixpkgs": {
    "commit": "52e3e80afff4b16ccb7c52e9f0f5220552f03d04"
  }
}

It's unrelated, but I found that settings the two env variables was necessary to use certain packages.

I added the Pytorch library using poetry add torch, which got version 1.13.1. I tried running the following script, which just creates two arrays and tries to multiply them:

import torch

def create_arrays(n):
    x = torch.ones(n, n)
    y = torch.randn(n, n * 2)
    return x , y


def main():
    x, y = create_arrays(1000)
    x = x.to("cuda")
    y = y.to("cuda")
    z = x @ y

if __name__ == "__main__":
    main()

I got the following error, which I don't when running outside of devbox:

Traceback (most recent call last):
  File "/redacted/test_torch.py", line 16, in <module>
    main()
  File "/redacted/test_torch.py", line 11, in main
    x = x.to("cuda")
  File "/redacted/.venv/lib/python3.10/site-packages/torch/cuda/__init__.py", line 229, in _lazy_init
    torch._C._cuda_init()
RuntimeError: Found no NVIDIA driver on your system. Please check that you have an NVIDIA GPU and installed a driver from http://www.nvidia.com/Download/index.aspx

Thanks! We'll use this to debug and circle back.

I've been tinkering with this recently, and got your example to work on WSL (Not sure if you're using that, or Linux directly). I believe the key is that you need to pass the folder with your NVIDIA drivers in the LD_LIBRARY_PATH variable.

The Devbox JSON that worked on my machine was the following:

{
  "packages": [
    "python310",
    "poetry",
    "stdenv.cc.cc.lib",
    "cudatoolkit"
  ],
  "shell": {
    "init_hook": [
      "export LD_LIBRARY_PATH=`pwd`/.devbox/nix/profile/default/lib:/usr/lib/wsl/lib"
    ]
  },
  "nixpkgs": {
    "commit": "f79ac848e3d6f0c12c52758c0f25c10c97ca3b62"
  }
}

Where /usr/lib/wsl/lib is the directory where the Windows NVIDIA drivers are linked in WSL2.

Setting LD_LIBRARY_PATH to the following seems to run on both WSL2 and Ubuntu machines with NVIDIA drivers installed:

export LD_LIBRARY_PATH=`pwd`/.devbox/nix/profile/default/lib:/usr/lib/x86_64-linux-gnu:/usr/wsl/lib

#760 could include setting this variable automatically based on the user's OS.