{ stdenv, lib, fetchFromGitHub, fetchpatch, buildPythonPackage, python,
  cudaSupport ? false, cudaPackages, magma,
  useSystemNccl ? true,
  MPISupport ? false, mpi,
  buildDocs ? false,

  # Native build inputs
  cmake, util-linux, linkFarm, symlinkJoin, which, pybind11, removeReferencesTo,
  pythonRelaxDepsHook,

  # Build inputs
  numactl,
  Accelerate, CoreServices, libobjc,

  # Propagated build inputs
  filelock,
  jinja2,
  networkx,
  openai-triton,
  sympy,
  numpy, pyyaml, cffi, click, typing-extensions,

  # Unit tests
  hypothesis, psutil,

  # Disable MKLDNN on aarch64-darwin, it negatively impacts performance,
  # this is also what official pytorch build does
  mklDnnSupport ? !(stdenv.isDarwin && stdenv.isAarch64),

  # virtual pkg that consistently instantiates blas across nixpkgs
  # See https://github.com/NixOS/nixpkgs/pull/83888
  blas,

  # ninja (https://ninja-build.org) must be available to run C++ extensions tests,
  ninja,

  linuxHeaders_5_19,

  # dependencies for torch.utils.tensorboard
  pillow, six, future, tensorboard, protobuf,

  isPy3k, pythonOlder,

  # ROCm dependencies
  rocmSupport ? false,
  gpuTargets ? [ ],
  openmp, rocm-core, hip, rccl, miopen, miopengemm, rocrand, rocblas,
  rocfft, rocsparse, hipsparse, rocthrust, rocprim, hipcub, roctracer,
  rocsolver, hipfft, hipsolver, hipblas, rocminfo, rocm-thunk, rocm-comgr,
  rocm-device-libs, rocm-runtime, rocm-opencl-runtime, hipify
}:

let
  inherit (lib) lists strings trivial;
  inherit (cudaPackages) cudatoolkit cudaFlags cudnn nccl;
in

assert cudaSupport -> stdenv.isLinux;
assert cudaSupport -> (cudaPackages.cudaMajorVersion == "11");

# confirm that cudatoolkits are sync'd across dependencies
assert !(MPISupport && cudaSupport) || mpi.cudatoolkit == cudatoolkit;
assert !cudaSupport || magma.cudaPackages.cudatoolkit == cudatoolkit;

let
  setBool = v: if v then "1" else "0";

  # https://github.com/pytorch/pytorch/blob/v1.13.1/torch/utils/cpp_extension.py#L1751
  supportedTorchCudaCapabilities =
    let
      real = ["3.5" "3.7" "5.0" "5.2" "5.3" "6.0" "6.1" "6.2" "7.0" "7.2" "7.5" "8.0" "8.6"];
      ptx = lists.map (x: "${x}+PTX") real;
    in
    real ++ ptx;

  # NOTE: The lists.subtractLists function is perhaps a bit unintuitive. It subtracts the elements
  #   of the first list *from* the second list. That means:
  #   lists.subtractLists a b = b - a

  # For CUDA
  supportedCudaCapabilities = lists.intersectLists cudaFlags.cudaCapabilities supportedTorchCudaCapabilities;
  unsupportedCudaCapabilities = lists.subtractLists supportedCudaCapabilities cudaFlags.cudaCapabilities;

  # Use trivial.warnIf to print a warning if any unsupported GPU targets are specified.
  gpuArchWarner = supported: unsupported:
    trivial.throwIf (supported == [ ])
      (
        "No supported GPU targets specified. Requested GPU targets: "
        + strings.concatStringsSep ", " unsupported
      )
      supported;

  # Create the gpuTargetString.
  gpuTargetString = strings.concatStringsSep ";" (
    if gpuTargets != [ ] then
    # If gpuTargets is specified, it always takes priority.
      gpuTargets
    else if cudaSupport then
      gpuArchWarner supportedCudaCapabilities unsupportedCudaCapabilities
    else if rocmSupport then
      hip.gpuTargets
    else
      throw "No GPU targets specified"
  );

  cudatoolkit_joined = symlinkJoin {
    name = "${cudatoolkit.name}-unsplit";
    # nccl is here purely for semantic grouping it could be moved to nativeBuildInputs
    paths = [ cudatoolkit.out cudatoolkit.lib nccl.dev nccl.out ];
  };

  # Normally libcuda.so.1 is provided at runtime by nvidia-x11 via
  # LD_LIBRARY_PATH=/run/opengl-driver/lib.  We only use the stub
  # libcuda.so from cudatoolkit for running tests, so that we don’t have
  # to recompile pytorch on every update to nvidia-x11 or the kernel.
  cudaStub = linkFarm "cuda-stub" [{
    name = "libcuda.so.1";
    path = "${cudatoolkit}/lib/stubs/libcuda.so";
  }];
  cudaStubEnv = lib.optionalString cudaSupport
    "LD_LIBRARY_PATH=${cudaStub}\${LD_LIBRARY_PATH:+:}$LD_LIBRARY_PATH ";

  rocmtoolkit_joined = symlinkJoin {
    name = "rocm-merged";

    paths = [
      rocm-core hip rccl miopen miopengemm rocrand rocblas
      rocfft rocsparse hipsparse rocthrust rocprim hipcub
      roctracer rocfft rocsolver hipfft hipsolver hipblas
      rocminfo rocm-thunk rocm-comgr rocm-device-libs
      rocm-runtime rocm-opencl-runtime hipify
    ];
  };
in buildPythonPackage rec {
  pname = "torch";
  # Don't forget to update torch-bin to the same version.
  version = "2.0.1";
  format = "setuptools";

  disabled = pythonOlder "3.8.0";

  outputs = [
    "out" # output standard python package
    "dev" # output libtorch headers
    "lib" # output libtorch libraries
  ];

  src = fetchFromGitHub {
    owner = "pytorch";
    repo = "pytorch";
    rev = "refs/tags/v${version}";
    fetchSubmodules = true;
    hash = "sha256-xUj77yKz3IQ3gd/G32pI4OhL3LoN1zS7eFg0/0nZp5I=";
  };

  patches = lib.optionals (stdenv.isDarwin && stdenv.isx86_64) [
    # pthreadpool added support for Grand Central Dispatch in April
    # 2020. However, this relies on functionality (DISPATCH_APPLY_AUTO)
    # that is available starting with macOS 10.13. However, our current
    # base is 10.12. Until we upgrade, we can fall back on the older
    # pthread support.
    ./pthreadpool-disable-gcd.diff
  ];

  postPatch = lib.optionalString rocmSupport ''
    # https://github.com/facebookincubator/gloo/pull/297
    substituteInPlace third_party/gloo/cmake/Hipify.cmake \
      --replace "\''${HIPIFY_COMMAND}" "python \''${HIPIFY_COMMAND}"

    # Replace hard-coded rocm paths
    substituteInPlace caffe2/CMakeLists.txt \
      --replace "/opt/rocm" "${rocmtoolkit_joined}" \
      --replace "hcc/include" "hip/include" \
      --replace "rocblas/include" "include/rocblas" \
      --replace "hipsparse/include" "include/hipsparse"

    # Doesn't pick up the environment variable?
    substituteInPlace third_party/kineto/libkineto/CMakeLists.txt \
      --replace "\''$ENV{ROCM_SOURCE_DIR}" "${rocmtoolkit_joined}" \
      --replace "/opt/rocm" "${rocmtoolkit_joined}"

    # Strangely, this is never set in cmake
    substituteInPlace cmake/public/LoadHIP.cmake \
      --replace "set(ROCM_PATH \$ENV{ROCM_PATH})" \
        "set(ROCM_PATH \$ENV{ROCM_PATH})''\nset(ROCM_VERSION ${lib.concatStrings (lib.intersperse "0" (lib.splitString "." hip.version))})"
  ''
  # error: no member named 'aligned_alloc' in the global namespace; did you mean simply 'aligned_alloc'
  # This lib overrided aligned_alloc hence the error message. Tltr: his function is linkable but not in header.
  + lib.optionalString (stdenv.isDarwin && lib.versionOlder stdenv.targetPlatform.darwinSdkVersion "11.0") ''
    substituteInPlace third_party/pocketfft/pocketfft_hdronly.h --replace '#if __cplusplus >= 201703L
    inline void *aligned_alloc(size_t align, size_t size)' '#if __cplusplus >= 201703L && 0
    inline void *aligned_alloc(size_t align, size_t size)'
  '';

  preConfigure = lib.optionalString cudaSupport ''
    export TORCH_CUDA_ARCH_LIST="${gpuTargetString}"
    export CC=${cudatoolkit.cc}/bin/gcc CXX=${cudatoolkit.cc}/bin/g++
  '' + lib.optionalString (cudaSupport && cudnn != null) ''
    export CUDNN_INCLUDE_DIR=${cudnn}/include
  '' + lib.optionalString rocmSupport ''
    export ROCM_PATH=${rocmtoolkit_joined}
    export ROCM_SOURCE_DIR=${rocmtoolkit_joined}
    export PYTORCH_ROCM_ARCH="${gpuTargetString}"
    export CMAKE_CXX_FLAGS="-I${rocmtoolkit_joined}/include -I${rocmtoolkit_joined}/include/rocblas"
    python tools/amd_build/build_amd.py
  '';

  # Use pytorch's custom configurations
  dontUseCmakeConfigure = true;

  BUILD_NAMEDTENSOR = setBool true;
  BUILD_DOCS = setBool buildDocs;

  # We only do an imports check, so do not build tests either.
  BUILD_TEST = setBool false;

  # Unlike MKL, oneDNN (née MKLDNN) is FOSS, so we enable support for
  # it by default. PyTorch currently uses its own vendored version
  # of oneDNN through Intel iDeep.
  USE_MKLDNN = setBool mklDnnSupport;
  USE_MKLDNN_CBLAS = setBool mklDnnSupport;

  # Avoid using pybind11 from git submodule
  # Also avoids pytorch exporting the headers of pybind11
  USE_SYSTEM_BIND11 = true;

  preBuild = ''
    export MAX_JOBS=$NIX_BUILD_CORES
    ${python.pythonForBuild.interpreter} setup.py build --cmake-only
    ${cmake}/bin/cmake build
  '';

  preFixup = ''
    function join_by { local IFS="$1"; shift; echo "$*"; }
    function strip2 {
      IFS=':'
      read -ra RP <<< $(patchelf --print-rpath $1)
      IFS=' '
      RP_NEW=$(join_by : ''${RP[@]:2})
      patchelf --set-rpath \$ORIGIN:''${RP_NEW} "$1"
    }
    for f in $(find ''${out} -name 'libcaffe2*.so')
    do
      strip2 $f
    done
  '';

  # Override the (weirdly) wrong version set by default. See
  # https://github.com/NixOS/nixpkgs/pull/52437#issuecomment-449718038
  # https://github.com/pytorch/pytorch/blob/v1.0.0/setup.py#L267
  PYTORCH_BUILD_VERSION = version;
  PYTORCH_BUILD_NUMBER = 0;

  USE_SYSTEM_NCCL = setBool useSystemNccl;                  # don't build pytorch's third_party NCCL

  # Suppress a weird warning in mkl-dnn, part of ideep in pytorch
  # (upstream seems to have fixed this in the wrong place?)
  # https://github.com/intel/mkl-dnn/commit/8134d346cdb7fe1695a2aa55771071d455fae0bc
  # https://github.com/pytorch/pytorch/issues/22346
  #
  # Also of interest: pytorch ignores CXXFLAGS uses CFLAGS for both C and C++:
  # https://github.com/pytorch/pytorch/blob/v1.11.0/setup.py#L17
  env.NIX_CFLAGS_COMPILE = toString ((lib.optionals (blas.implementation == "mkl") [ "-Wno-error=array-bounds" ]
  # Suppress gcc regression: avx512 math function raises uninitialized variable warning
  # https://gcc.gnu.org/bugzilla/show_bug.cgi?id=105593
  # See also: Fails to compile with GCC 12.1.0 https://github.com/pytorch/pytorch/issues/77939
  ++ lib.optionals (stdenv.cc.isGNU && lib.versionAtLeast stdenv.cc.version "12.0.0") [
    "-Wno-error=maybe-uninitialized"
    "-Wno-error=uninitialized"
  ]
  # Since pytorch 2.0:
  # gcc-12.2.0/include/c++/12.2.0/bits/new_allocator.h:158:33: error: ‘void operator delete(void*, std::size_t)’
  # ... called on pointer ‘<unknown>’ with nonzero offset [1, 9223372036854775800] [-Werror=free-nonheap-object]
  ++ lib.optionals (stdenv.cc.isGNU && lib.versions.major stdenv.cc.version == "12" ) [
    "-Wno-error=free-nonheap-object"
  ]));

  nativeBuildInputs = [
    cmake
    util-linux
    which
    ninja
    pybind11
    pythonRelaxDepsHook
    removeReferencesTo
  ] ++ lib.optionals cudaSupport [ cudatoolkit_joined ]
    ++ lib.optionals rocmSupport [ rocmtoolkit_joined ];

  buildInputs = [ blas blas.provider pybind11 ]
    ++ lib.optionals stdenv.isLinux [ linuxHeaders_5_19 ] # TMP: avoid "flexible array member" errors for now
    ++ lib.optionals cudaSupport [ cudnn nccl ]
    ++ lib.optionals rocmSupport [ openmp ]
    ++ lib.optionals (cudaSupport || rocmSupport) [ magma ]
    ++ lib.optionals stdenv.isLinux [ numactl ]
    ++ lib.optionals stdenv.isDarwin [ Accelerate CoreServices libobjc ];

  propagatedBuildInputs = [
    cffi
    click
    numpy
    pyyaml

    # From install_requires:
    filelock
    typing-extensions
    sympy
    networkx
    jinja2

    # the following are required for tensorboard support
    pillow six future tensorboard protobuf
  ]
  ++ lib.optionals MPISupport [ mpi ]
  ++ lib.optionals rocmSupport [ rocmtoolkit_joined ]
  # rocm build requires openai-triton;
  # openai-triton currently requires cuda_nvcc,
  # so not including it in the cpu-only build;
  # torch.compile relies on openai-triton,
  # so we include it for the cuda build as well
  ++ lib.optionals (rocmSupport || cudaSupport) [
    openai-triton
  ];

  # Tests take a long time and may be flaky, so just sanity-check imports
  doCheck = false;

  pythonImportsCheck = [
    "torch"
  ];

  nativeCheckInputs = [ hypothesis ninja psutil ];

  checkPhase = with lib.versions; with lib.strings; concatStringsSep " " [
    "runHook preCheck"
    cudaStubEnv
    "${python.interpreter} test/run_test.py"
    "--exclude"
    (concatStringsSep " " [
      "utils" # utils requires git, which is not allowed in the check phase

      # "dataloader" # psutils correctly finds and triggers multiprocessing, but is too sandboxed to run -- resulting in numerous errors
      # ^^^^^^^^^^^^ NOTE: while test_dataloader does return errors, these are acceptable errors and do not interfere with the build

      # tensorboard has acceptable failures for pytorch 1.3.x due to dependencies on tensorboard-plugins
      (optionalString (majorMinor version == "1.3" ) "tensorboard")
    ])
    "runHook postCheck"
  ];

  pythonRemoveDeps = [
    # In our dist-info the name is just "triton"
    "pytorch-triton-rocm"
  ];

  postInstall = ''
    find "$out/${python.sitePackages}/torch/include" "$out/${python.sitePackages}/torch/lib" -type f -exec remove-references-to -t ${stdenv.cc} '{}' +

    mkdir $dev
    cp -r $out/${python.sitePackages}/torch/include $dev/include
    cp -r $out/${python.sitePackages}/torch/share $dev/share

    # Fix up library paths for split outputs
    substituteInPlace \
      $dev/share/cmake/Torch/TorchConfig.cmake \
      --replace \''${TORCH_INSTALL_PREFIX}/lib "$lib/lib"

    substituteInPlace \
      $dev/share/cmake/Caffe2/Caffe2Targets-release.cmake \
      --replace \''${_IMPORT_PREFIX}/lib "$lib/lib"

    mkdir $lib
    mv $out/${python.sitePackages}/torch/lib $lib/lib
    ln -s $lib/lib $out/${python.sitePackages}/torch/lib
  '' + lib.optionalString rocmSupport ''
    substituteInPlace $dev/share/cmake/Tensorpipe/TensorpipeTargets-release.cmake \
      --replace "\''${_IMPORT_PREFIX}/lib64" "$lib/lib"

    substituteInPlace $dev/share/cmake/ATen/ATenConfig.cmake \
      --replace "/build/source/torch/include" "$dev/include"
  '';

  postFixup = lib.optionalString stdenv.isDarwin ''
    for f in $(ls $lib/lib/*.dylib); do
        install_name_tool -id $lib/lib/$(basename $f) $f || true
    done

    install_name_tool -change @rpath/libshm.dylib $lib/lib/libshm.dylib $lib/lib/libtorch_python.dylib
    install_name_tool -change @rpath/libtorch.dylib $lib/lib/libtorch.dylib $lib/lib/libtorch_python.dylib
    install_name_tool -change @rpath/libc10.dylib $lib/lib/libc10.dylib $lib/lib/libtorch_python.dylib

    install_name_tool -change @rpath/libc10.dylib $lib/lib/libc10.dylib $lib/lib/libtorch.dylib

    install_name_tool -change @rpath/libtorch.dylib $lib/lib/libtorch.dylib $lib/lib/libshm.dylib
    install_name_tool -change @rpath/libc10.dylib $lib/lib/libc10.dylib $lib/lib/libshm.dylib
  '';

  # Builds in 2+h with 2 cores, and ~15m with a big-parallel builder.
  requiredSystemFeatures = [ "big-parallel" ];

  passthru = {
    inherit cudaSupport cudaPackages;
    # At least for 1.10.2 `torch.fft` is unavailable unless BLAS provider is MKL. This attribute allows for easy detection of its availability.
    blasProvider = blas.provider;
  } // lib.optionalAttrs cudaSupport {
    # NOTE: supportedCudaCapabilities isn't computed unless cudaSupport is true, so we can't use
    #   it in the passthru set above because a downstream package might try to access it even
    #   when cudaSupport is false. Better to have it missing than null or an empty list by default.
    cudaCapabilities = supportedCudaCapabilities;
  };

  meta = with lib; {
    changelog = "https://github.com/pytorch/pytorch/releases/tag/v${version}";
    # keep PyTorch in the description so the package can be found under that name on search.nixos.org
    description = "PyTorch: Tensors and Dynamic neural networks in Python with strong GPU acceleration";
    homepage = "https://pytorch.org/";
    license = licenses.bsd3;
    maintainers = with maintainers; [ teh thoughtpolice tscholak ]; # tscholak esp. for darwin-related builds
    platforms = with platforms; linux ++ lib.optionals (!cudaSupport && !rocmSupport) darwin;
    broken = rocmSupport && cudaSupport; # CUDA and ROCm are mutually exclusive
  };
}