👍️

I'll try to send some of the shells that I use a little later.
But meanwhile a few links you might find useful, if you haven't seen them yet:

• zimbatm's nixpkgs-unfree: https://discourse.nixos.org/t/announcing-nixpkgs-unfree/17505
  The idea is to use inputs.nixpkgs.url = github:numtide/nixpkgs-unfree/$branch as input, which exposes nixpkgs with the unfree already enabled and tracks NixOS/nixpkgs automatically. I'm using smth like this for the flake registry in my configuration.nix, so that nixpkgs#python3Packages.pytorch resolves into the pre-cached unfree and cuda-enabled pytorch for me, instead of throwing an error about NIXPKGS_ALLOW_UNFREE=1
• It's somewhat experimental, but there's now cachix with prebuilt cuda-enabled sci-comp packages, you can find more at https://nixos.wiki/wiki/CUDA. The cache is populated automatically for the last release, unstable, and master, much like nixpkgs-unfree does it. There's a slight delay between a branch update and the cache, but probably not more than a day
• I prefer using source-based pytorch instead of the wheel-based pytorch-bin. The way I use cuda-enabled packages is essentially just importing nixpkgs with { config = { allowUnfree = true; cudaSupport = true; }; } and using directly pytorch, jax, blender, etc. And I'd rather consume nixpkgs-unfree as a flake input, than import it myself in a project.