Openai-whisper failing on larger models (NVIDIA / CUDA issues)

I forgot to mention this, but the cache only takes effect after you switch to your new configuration, so you’ll need to:

  • Enable the cache
  • Disable CUDA and the packages that require CUDA
  • nixos-rebuild test or nixos-rebuild switch
  • Enable CUDA
  • nixos-rebuild switch

The issue is that you’re using an old nixpkgs. In nixos-24.11, triton is at version 3.1.0 both for python 3.11 and 3.12. So you’ll either need to use the triton you have with python3.11 or upgrade your system.

1 Like