{
  stdenv,
  config,
  lib,
  buildPythonPackage,
  fetchPypi,
  python,
  pythonOlder,
  pythonAtLeast,
  openssl_1_1,
  zlib,
  setuptools,
  cudaSupport ? config.cudaSupport or false,
  cudaPackages_11 ? { },
  addOpenGLRunpath,
  # runtime dependencies
  httpx,
  numpy,
  protobuf,
  pillow,
  decorator,
  astor,
  paddle-bfloat,
  opt-einsum,
}:

let
  pname = "paddlepaddle" + lib.optionalString cudaSupport "-gpu";
  version = "2.5.0";
  format = "wheel";
  pyShortVersion = "cp${builtins.replaceStrings [ "." ] [ "" ] python.pythonVersion}";
  cpuOrGpu = if cudaSupport then "gpu" else "cpu";
  allHashAndPlatform = import ./binary-hashes.nix;
  hash =
    allHashAndPlatform."${stdenv.system}"."${cpuOrGpu}"."${pyShortVersion}"
      or (throw "${pname} has no binary-hashes.nix entry for '${stdenv.system}.${cpuOrGpu}.${pyShortVersion}' attribute");
  platform = allHashAndPlatform."${stdenv.system}".platform;
  src = fetchPypi ({
    inherit
      version
      format
      hash
      platform
      ;
    pname = builtins.replaceStrings [ "-" ] [ "_" ] pname;
    dist = pyShortVersion;
    python = pyShortVersion;
    abi = pyShortVersion;
  });
in
buildPythonPackage {
  inherit
    pname
    version
    format
    src
    ;

  disabled = pythonOlder "3.9" || pythonAtLeast "3.11";

  libraryPath = lib.makeLibraryPath (
    # TODO: remove openssl_1_1 and zlib, maybe by building paddlepaddle from
    # source as suggested in the following comment:
    # https://github.com/NixOS/nixpkgs/pull/243583#issuecomment-1641450848
    [
      openssl_1_1
      zlib
    ]
    ++ lib.optionals cudaSupport (
      with cudaPackages_11;
      [
        cudatoolkit.lib
        cudatoolkit.out
        cudnn
      ]
    )
  );

  postFixup = lib.optionalString stdenv.isLinux ''
    function fixRunPath {
      p=$(patchelf --print-rpath $1)
      patchelf --set-rpath "$p:$libraryPath" $1
      ${lib.optionalString cudaSupport ''
        addOpenGLRunpath $1
      ''}
    }
    fixRunPath $out/${python.sitePackages}/paddle/fluid/libpaddle.so
  '';

  nativeBuildInputs = [ addOpenGLRunpath ];

  propagatedBuildInputs = [
    setuptools
    httpx
    numpy
    protobuf
    pillow
    decorator
    astor
    paddle-bfloat
    opt-einsum
  ];

  pythonImportsCheck = [ "paddle" ];

  # no tests
  doCheck = false;

  meta = with lib; {
    description = "PArallel Distributed Deep LEarning: Machine Learning Framework from Industrial Practice （『飞桨』核心框架，深度学习&机器学习高性能单机、分布式训练和跨平台部署";
    homepage = "https://github.com/PaddlePaddle/Paddle";
    license = licenses.asl20;
    maintainers = with maintainers; [ happysalada ];
    platforms =
      [ "x86_64-linux" ]
      ++ optionals (!cudaSupport) [
        "x86_64-darwin"
        "aarch64-darwin"
      ];
  };
}