How does this work on non-NixOS when someone runs another version of the nvidia driver?

I have no means to test it, so I don't know.

As a comparison, tensorflow also seems to work only on linux:

  tensorflow = callPackage ../development/python-modules/tensorflow rec {
    cudaSupport = pkgs.config.cudaSupport or false;
    inherit (pkgs.linuxPackages) nvidia_x11;
    cudatoolkit = pkgs.cudatoolkit9;
    cudnn = pkgs.cudnn_cudatoolkit9;
  };

Should I add the possibility to build the derivation with nvidia_x11=null? That way libcuda.so should be resolved to a nvidia driver that is not installed with nixos. Following the example from tensorflow i'd still leave nvidia_x11 = pkgs.linuxPackages.nvidia_x11 as a default.

This is good enough for now. Please get rid of the merge, rebase onto master, and make sure you have a commit per package you change.

All done.

Hello! Have you tried the addOpenGLRunpath way, instead of linking directly to the hardware-dependent driver?

cc NixOS/cuda-maintainers

@SomeoneSerge note this is a merge from several years ago