infrastructure/machines/krz01/_configuration.nix

{
  config,
  lib,
  pkgs,
  meta,
  name,
  ...
}:

lib.extra.mkConfig {
  enabledModules = [
    # INFO: This list needs to stay sorted alphabetically
  ];

  enabledServices = [
    # INFO: This list needs to stay sorted alphabetically
    # Machine learning API machine
    "microvm-ml01"
    "microvm-router01"
    "nvidia-tesla-k80"
    "proxmox"
  ];

  extraConfig = {
    microvm = {
      host.enable = true;
    };
    dgn-hardware = {
      useZfs = true;
      zfsPools = [
        "dpool"
        "ppool0"
      ];
    };

    services.netbird.enable = true;

    # We are going to use CUDA here.
    nixpkgs.config.cudaSupport = true;
    hardware.graphics.enable = true;
    environment.systemPackages = [
      ((pkgs.openai-whisper-cpp.override { cudaPackages = pkgs.cudaPackages_11; }).overrideAttrs (old: {
        src = pkgs.fetchFromGitHub {
          owner = "ggerganov";
          repo = "whisper.cpp";
          rev = "v1.7.1";
          hash = "sha256-EDFUVjud79ZRCzGbOh9L9NcXfN3ikvsqkVSOME9F9oo=";
        };
        env = {
          WHISPER_CUBLAS = "";
          GGML_CUDA = "1";
        };
        # We only need Compute Capability 3.7.
        CUDA_ARCH_FLAGS = [ "sm_37" ];
        # We are GPU-only anyway.
        patches = (old.patches or [ ]) ++ [
          ./no-weird-microarch.patch
          ./all-nvcc-arch.patch
        ];
      }))
    ];

    services = {
      ollama = {
        enable = true;
        host = meta.network.${name}.netbirdIp;
        package = pkgs.callPackage ./ollama.nix {
          cudaPackages = pkgs.cudaPackages_11;
          # We need to thread our nvidia x11 driver for CUDA.
          extraLibraries = [ config.hardware.nvidia.package ];
        };
      };
    };

    networking.firewall.interfaces.wt0.allowedTCPPorts = [ config.services.ollama.port ];
  };

  root = ./.;
}