If you build cudaPackages.cudatoolkit you'll see that it's actually a symlink farm which includes a lot of stuff you don't need. In particular:

❯ readlink -f result/nvvm/lib64/libnvvm.so.4.0.0
/nix/store/fby2d6b4jgfb8awwjhzdrd13r8vx7ilw-cuda_nvcc-12.2.140-bin/nvvm/lib64/libnvvm.so.4.0.0

...it would've been enough for you to use ${getBin cudaPackages.cuda_nvcc}/nvvm/lib64/libnvvm.so which is much smaller (whoops, that's a bug, we definitely didn't want this to end up in .bin, so expect this to change)

I would think runtime for jit, since there is a @cuda.jit operator on my python code. This is now my shell.nix, removed extra buildInputs, and used the wrapper, which had to be python39 to even get back to where I was:

  config = {
    allowUnfree = true;
    cudaSupport = true;
  };
};

let
  pythonEnv = pkgs.python39.withPackages(ps: with ps; [ numba numpy ]);
in
pkgs.mkShell {
   name = "cuda-env-shell";
   buildInputs = [
      pythonEnv
     libstdcxx5
     cudaPackages.cudatoolkit
   ];
   shellHook = ''
      export CUDA_PATH=${pkgs.cudaPackages.cudatoolkit}
      export CUDA_HOME=${pkgs.cudatoolkit}
      export EXTRA_CCFLAGS="-I/usr/include"
      export NUMBAPRO_NVVM=${pkgs.cudatoolkit}/nvvm/lib64/libnvvm.so
      export NUMBAPRO_LIBDEVICE=${pkgs.cudatoolkit}/nvvm/libdevice
      export LD_LIBRARY_PATH=${pkgs.cudatoolkit}/nvvm/lib64/libnvvm.so:$CUDA_HOME
      # source venv/bin/activate
      python test2.py
   '';          
}

If I go into python repl I can import numpy and numba, but I still get the same error when running an actual cuda script. This was the python code:

from numba import cuda
import numpy as np

@cuda.jit
def cudakernel0(array):
    for i in range(array.size):
        array[i] += 0.5

array = np.array([0, 1], np.float32)
print('Initial array:', array)

print('Kernel launch: cudakernel0[1, 1](array)')
cudakernel0[1, 1](array)

print('Updated array:',array)

I can run:

from numba import cuda
print(cuda.detect())

Also I doubt the linking of nvvm.so helps, since numba is now relying on the CUDA_HOME env var (https://numba.pydata.org/numba-doc/latest/cuda/overview.html).
What do you mean to use ${getBin...}, where could I use that?

export LD_LIBRARY_PATH=${pkgs.cudatoolkit}/nvvm/lib64/libnvvm.so:$CUDA_HOME

I'm pretty sure gilbc treats LD_LIBRARY_PATH as a list of directories only (so you'd want LD_LIBRARY_PATH=${pkgs.cudatoolkit}/nvvm/lib64/ instead)

Also I doubt the linking of nvvm.so helps, since numba is now relying on the CUDA_HOME env var

The error you posted earlier suggests that they just dlopen("libnvvm.so", ...)

In reply to @ss:someonex.net

export LD_LIBRARY_PATH=${pkgs.cudatoolkit}/nvvm/lib64/libnvvm.so:$CUDA_HOME

I'm pretty sure gilbc treats LD_LIBRARY_PATH as a list of directories only (so you'd want LD_LIBRARY_PATH=${pkgs.cudatoolkit}/nvvm/lib64/ instead)

That was true yes, it seems it does find it now. However it claims:

No supported GPU compute capabilities found. Please check your cudatoolkit version matches your CUDA version

nvcc --version gives 12.2, and nvidia-smi gives 12.4. This should be compatible as far as I'm aware. I'll see if I can get more info on this, unless you know that this is not supposed to be compatible and I need to adjust one of the versions.

In reply to @jeroenvb3:matrix.org

That was true yes, it seems it does find it now. However it claims:

No supported GPU compute capabilities found. Please check your cudatoolkit version matches your CUDA version

nvcc --version gives 12.2, and nvidia-smi gives 12.4. This should be compatible as far as I'm aware. I'll see if I can get more info on this, unless you know that this is not supposed to be compatible and I need to adjust one of the versions.

In reply to @ss:someonex.net
Could you gist a reproducing code and a full log?

https://gitlab.com/jeroenvb3/cuda-setup

I added the nix config, the python code, the output, and my nvidia driver info. If you need more system info please tell but I'd think this is all that will be unique. Thanks!

In reply to @brodriguesco:matrix.org

hello , we could use some testers on NixOS that have an nvidia GPU

would really appreciate if you could review this PR: https://github.com/NixOS/nixpkgs/pull/328980

especially following the instructions in the documentation: https://github.com/NixOS/nixpkgs/blob/7c2dcd41c35833809f05108c75ca0a6e3df242ce/doc/languages-frameworks/r.section.md#torch-with-gpu-acceleration-torch-gpu-accel

question is, do we need nixglhost even on NixOs ? some first testing seems to indicate that yes