infrastructure/machines/krz01/ollama.nix

{
  lib,
  buildGoModule,
  fetchFromGitHub,
  buildEnv,
  linkFarm,
  overrideCC,
  makeWrapper,
  stdenv,
  addDriverRunpath,
  nix-update-script,

  cmake,
  gcc11,
  clblast,
  libdrm,
  rocmPackages,
  cudaPackages,
  darwin,
  autoAddDriverRunpath,
  extraLibraries ? [ ],

  nixosTests,
  testers,
  ollama,
  ollama-rocm,
  ollama-cuda,

  config,
  # one of `[ null false "rocm" "cuda" ]`
  acceleration ? null,
}:

assert builtins.elem acceleration [
  null
  false
  "rocm"
  "cuda"
];

let
  pname = "ollama";
  version = "2024-09-10-cc35";

  src = fetchFromGitHub {
    owner = "aliotard";
    repo = "ollama";
    rev = "34827c01f7723c7f5f9f5e392fe85f5a4a5d5fc0";
    hash = "sha256-xFNuqcW7YWeyCyw5QLBnCHHTSMITR6LJkJT0CXZC+Y8=";
    fetchSubmodules = true;
  };

  vendorHash = "sha256-hSxcREAujhvzHVNwnRTfhi0MKI3s8HNavER2VLz6SYk=";

  validateFallback = lib.warnIf (config.rocmSupport && config.cudaSupport) (lib.concatStrings [
    "both `nixpkgs.config.rocmSupport` and `nixpkgs.config.cudaSupport` are enabled, "
    "but they are mutually exclusive; falling back to cpu"
  ]) (!(config.rocmSupport && config.cudaSupport));
  shouldEnable =
    mode: fallback: (acceleration == mode) || (fallback && acceleration == null && validateFallback);

  rocmRequested = shouldEnable "rocm" config.rocmSupport;
  cudaRequested = shouldEnable "cuda" config.cudaSupport;

  enableRocm = rocmRequested && stdenv.isLinux;
  enableCuda = cudaRequested && stdenv.isLinux;

  rocmLibs = [
    rocmPackages.clr
    rocmPackages.hipblas
    rocmPackages.rocblas
    rocmPackages.rocsolver
    rocmPackages.rocsparse
    rocmPackages.rocm-device-libs
    rocmPackages.rocm-smi
  ];
  rocmClang = linkFarm "rocm-clang" { llvm = rocmPackages.llvm.clang; };
  rocmPath = buildEnv {
    name = "rocm-path";
    paths = rocmLibs ++ [ rocmClang ];
  };

  cudaLibs = [
    cudaPackages.cuda_cudart
    cudaPackages.libcublas
    cudaPackages.cuda_cccl
  ];
  cudaToolkit = buildEnv {
    name = "cuda-merged";
    paths = map lib.getLib cudaLibs ++ [
      (lib.getOutput "static" cudaPackages.cuda_cudart)
      (lib.getBin (cudaPackages.cuda_nvcc.__spliced.buildHost or cudaPackages.cuda_nvcc))
    ];
  };

  metalFrameworks = with darwin.apple_sdk_11_0.frameworks; [
    Accelerate
    Metal
    MetalKit
    MetalPerformanceShaders
  ];

  wrapperOptions =
    [
      # ollama embeds llama-cpp binaries which actually run the ai models
      # these llama-cpp binaries are unaffected by the ollama binary's DT_RUNPATH
      # LD_LIBRARY_PATH is temporarily required to use the gpu
      # until these llama-cpp binaries can have their runpath patched
      "--suffix LD_LIBRARY_PATH : '${addDriverRunpath.driverLink}/lib'"
      "--suffix LD_LIBRARY_PATH : '${lib.makeLibraryPath (map lib.getLib extraLibraries)}'"
    ]
    ++ lib.optionals enableRocm [
      "--suffix LD_LIBRARY_PATH : '${rocmPath}/lib'"
      "--set-default HIP_PATH '${rocmPath}'"
    ]
    ++ lib.optionals enableCuda [
      "--suffix LD_LIBRARY_PATH : '${lib.makeLibraryPath (map lib.getLib cudaLibs)}'"
    ];
  wrapperArgs = builtins.concatStringsSep " " wrapperOptions;

  goBuild =
    if enableCuda then buildGoModule.override { stdenv = overrideCC stdenv gcc11; } else buildGoModule;
  inherit (lib) licenses platforms maintainers;
in
goBuild {
  inherit
    pname
    version
    src
    vendorHash
    ;

  env =
    lib.optionalAttrs enableRocm {
      ROCM_PATH = rocmPath;
      CLBlast_DIR = "${clblast}/lib/cmake/CLBlast";
    }
    // lib.optionalAttrs enableCuda { CUDA_LIB_DIR = "${cudaToolkit}/lib"; }
    // {
      CMAKE_CUDA_ARCHITECTURES = "35;37";
    };

  nativeBuildInputs =
    [ cmake ]
    ++ lib.optionals enableRocm [ rocmPackages.llvm.bintools ]
    ++ lib.optionals enableCuda [ cudaPackages.cuda_nvcc ]
    ++ lib.optionals (enableRocm || enableCuda) [
      makeWrapper
      autoAddDriverRunpath
    ]
    ++ lib.optionals stdenv.isDarwin metalFrameworks;

  buildInputs =
    lib.optionals enableRocm (rocmLibs ++ [ libdrm ])
    ++ lib.optionals enableCuda cudaLibs
    ++ lib.optionals stdenv.isDarwin metalFrameworks;

  patches = [
    # disable uses of `git` in the `go generate` script
    # ollama's build script assumes the source is a git repo, but nix removes the git directory
    # this also disables necessary patches contained in `ollama/llm/patches/`
    # those patches are applied in `postPatch`
    ./disable-git.patch
  ];

  postPatch = ''
    # replace inaccurate version number with actual release version
    substituteInPlace version/version.go --replace-fail 0.0.0 '${version}'

    # apply ollama's patches to `llama.cpp` submodule
    for diff in llm/patches/*; do
      patch -p1 -d llm/llama.cpp < $diff
    done
  '';

  overrideModAttrs = _: _: {
    # don't run llama.cpp build in the module fetch phase
    preBuild = "";
  };

  preBuild = ''
    # disable uses of `git`, since nix removes the git directory
    export OLLAMA_SKIP_PATCHING=true
    # build llama.cpp libraries for ollama
    go generate ./...
  '';

  postFixup =
    ''
      # the app doesn't appear functional at the moment, so hide it
      mv "$out/bin/app" "$out/bin/.ollama-app"
    ''
    + lib.optionalString (enableRocm || enableCuda) ''
      # expose runtime libraries necessary to use the gpu
      wrapProgram "$out/bin/ollama" ${wrapperArgs}
    '';

  ldflags = [
    "-s"
    "-w"
    "-X=github.com/ollama/ollama/version.Version=${version}"
    "-X=github.com/ollama/ollama/server.mode=release"
    "-X=github.com/ollama/ollama/gpu.CudaComputeMajorMin=3"
    "-X=github.com/ollama/ollama/gpu.CudaComputeMinorMin=5"
  ];

  passthru = {
    tests =
      {
        inherit ollama;
        version = testers.testVersion {
          inherit version;
          package = ollama;
        };
      }
      // lib.optionalAttrs stdenv.isLinux {
        inherit ollama-rocm ollama-cuda;
        service = nixosTests.ollama;
        service-cuda = nixosTests.ollama-cuda;
        service-rocm = nixosTests.ollama-rocm;
      };

    updateScript = nix-update-script { };
  };

  meta = {
    description =
      "Get up and running with large language models locally"
      + lib.optionalString rocmRequested ", using ROCm for AMD GPU acceleration"
      + lib.optionalString cudaRequested ", using CUDA for NVIDIA GPU acceleration";
    homepage = "https://github.com/ollama/ollama";
    changelog = "https://github.com/ollama/ollama/releases/tag/v${version}";
    license = licenses.mit;
    platforms = if (rocmRequested || cudaRequested) then platforms.linux else platforms.unix;
    mainProgram = "ollama";
    maintainers = with maintainers; [
      abysssol
      dit7ya
      elohmeier
      roydubnium
    ];
  };
}