In reply to @ccicnce113424:matrix.org
@[Edward Tjörnhammar] I'm NOT rushing anything. I didn't touch the HPC-specific logic, and the PR is literally open for review. I've already replied to all your concerns in the PR comments. If you genuinely believe HPC is broken by these changes, please quote the exact lines of code on GitHub that cause the issue. Let's keep the discussion technical and factual.

Hi Samuel Ainsworth!

I'm having yet another hash mismatch on xla:

error: hash mismatch in fixed-output derivation '/nix/store/r97i94lw0gyqi8x1z7w0f65v8p8hwlyr-xla-0-unstable-2026-02-21-deps.tar.drv':
         specified: sha256-QTUqcP5t91Z4s+esxxFz2tGJAJplWXWZuYPqcC7ld+E=
            got:    sha256-go2EUhzSI2LcvYTjkrj3nbbwiLSusuj2Ijh6rlcarqE=

connor (burnt/out) (UTC-8), do you know if libnvvm is a CUDA>=13.0 thing? Or should it also be packaged for 12.9?

It should be... But only the 13.x redist manifests mention it...

connor (burnt/out) (UTC-8), do you know if libnvvm is a CUDA>=13.0 thing? Or should it also be packaged for 12.9?

It should be... But only the 13.x redist manifests mention it...

EDIT: It's under cudaPackages.cuda_nvcc/nvvm for 12.x 🫠

Hi all, this might be a stupid question but why do I have to build some packages still locally when on nixos-unstable with nixpkgs.config.cudaSupport = true; even though I've configured the binary cache in my substituters? Is this just because cuda packages have complicated versions and change very often that not all packages are built and cached on the cachix or am I doing something wrong?

This actually hinders me from updating right now because python3.13-torch-2.11.0 for python3.13-whisper-20250625 and python3.12-torch-2.11.0 for aider-chat-playwright-browser-help-bedrock-0.86.1 are eating my RAM and swap alive.

Yeah, this is the full relevant code:

    nix = {
      registry.nixpkgs.flake = inputs.nixpkgs;
      channel.enable = false; # Get rid of both unneeded state files and the nasty "Nix search path entry ... does not exist"
      settings = rec {
        use-xdg-base-directories = true;
        experimental-features = [
          "nix-command"
          "flakes"
        ];
        substituters = [
          "https://nix-community.cachix.org"
          "https://ai.cachix.org"
          "https://numtide.cachix.org"
          "https://cache.garnix.io"
          "https://cache.nixos-cuda.org"
        ];
        trusted-substituters = substituters;
        trusted-public-keys = [
          "nix-community.cachix.org-1:mB9FSh9qf2dCimDSUo8Zy7bkq5CX+/rkCWyvRCYg3Fs="
          "ai.cachix.org-1:N9dzRK+alWwoKXQlnn0H6aUx0lU/mspIoz8hMvGvbbc="
          "numtide.cachix.org-1:2ps1kLBUWjxIneOy1Ik6cQjb41X0iXVXeHigGmycPPE="
          "cache.garnix.io:CTFPyKSLcx5RMJKfLo5EEPUObbA78b0YQ2DTCJXqr9g="
          "cache.nixos-cuda.org:74DUi4Ye579gUqzH4ziL9IyiJBlDpMRn9MBN8oNan9M="
        ];
      };
      optimise = {
        automatic = true;
        dates = [ "weekly" ];
      };
    };

    nixpkgs.config.allowUnfreePredicate = (
      pkg:
      builtins.elem (lib.getName pkg) [
        "nvidia-x11"
        "cuda_cccl"
        "libcublas"
        "libcurand"
        "libcusparse"
        "libnvjitlink"
        "libcufft"
        "cudnn"
        "cuda_nvrtc"
        "cuda_cudart"
        "cuda_nvcc"
        "libnpp"
        "cuda_cupti"
        "cuda_nvml_dev"
        "cuda_nvtx"
        "libcufile"
        "libcusolver"
        "libcusparse_lt"
        "cuda_profiler_api"
        "cuda_cuobjdump"
        "cuda_nvdisasm"
        "cuda-merged"
        "cuda_gdb"
        "cuda_nvprune"
        "cuda_cuxxfilt"
        "cuda_sanitizer_api"
        "nvidia-settings"
        "obsidian"
        "mathematica"
        "spotify"
        "discord"
        "zoom"
        "steamcmd"
        "steam"
        "steam-unwrapped"
      ]
    );