Also, you don't want to build torch from source nowadays (unless you have a very powerful system).
If you still want to build things, you may set cudaCapabilities that correspond to your GPU.
https://developer.nvidia.com/cuda/gpus

It will significantly speed-up compilation compared to targeting the default ones.

Hi all, I'm a little confused why cuda_cudart 13.x won't evaluate on AArch64? 12.9 is fine, and I think jetpack-nixos uses 13 through cuda-legacy - but the manifest seems similar to the one in Nixpkgs anyway?

$ nix-instantiate -I nixpkgs=channel:nixos-unstable -E 'with import <nixpkgs> { localSystem = "aarch64-linux"; config = { allowUnfree = true; cudaForwardCompat = false; cudaCapabilities = [ "8.7" ]; }; }; cudaPackages_13_2.cuda_cudart'

       error: Refusing to evaluate package 'cuda13.2-cuda_cudart-13.2.75' in /nix/store/5nf8c4ail3r7sshsjl5bshwylyi57iyc-source/pkgs/development/cuda-modules/packages/cuda_cudart.nix:98 because it is not available on the requested hostPlatform:
         hostPlatform.system = "aarch64-linux"
         package.meta.platforms = [
           "aarch64-linux"
           "x86_64-linux"
         ]
         package.meta.badPlatforms = [
           "aarch64-linux"
         ]

It works when I bump cudaCapabilities to 12.1. Weird.

$ nix-instantiate -E 'with import <nixpkgs> { localSystem = "aarch64-linux"; config = { allowUnfree = true; cudaForwardCompat = false; cudaCapabilities = [ "12.1" ]; }; }; cudaPackages_13_2.cuda_cudart'
warning: you did not specify '--add-root'; the result might be removed by the garbage collector
/nix/store/4dqyz4cicqfhdq131g2vrhbvxvbz2w9g-cuda13.2-cuda_cudart-13.2.75.drv

In reply to @ethancedwards8:matrix.org
Does anyone have any suggestions/recommended platforms for me to rent a GPU with a full linux system that I can put nixos on so that I can start learning cuda on nixos? Unfortunately I do not have an NVIDIA gpu to learn on, but was hoping it would be possible to rent

Ah yep I see that, sorry I thought I searched the PRs 😅

Thanks and congrats on the new job!

In reply to @ethancedwards8:matrix.org
Preferably less than like $2/hour for compute. I am okay with spinning it up/down when using it if I can keep a persistent block device