{
  lib,
  stdenv,
  python,
  buildPythonPackage,
  pythonRelaxDepsHook,
  fetchFromGitHub,
  which,
  ninja,
  cmake,
  packaging,
  setuptools,
  torch,
  outlines,
  wheel,
  psutil,
  ray,
  pandas,
  pyarrow,
  sentencepiece,
  numpy,
  transformers,
  xformers,
  fastapi,
  uvicorn,
  pydantic,
  aioprometheus,
  pynvml,
  openai,
  pyzmq,
  tiktoken,
  torchvision,
  py-cpuinfo,
  lm-format-enforcer,
  prometheus-fastapi-instrumentator,
  cupy,
  writeShellScript,

  config,

  cudaSupport ? config.cudaSupport,
  cudaPackages ? { },

  # Has to be either rocm or cuda, default to the free one
  rocmSupport ? !config.cudaSupport,
  rocmPackages ? { },
  gpuTargets ? [ ],
}@args:

let
  cutlass = fetchFromGitHub {
    owner = "NVIDIA";
    repo = "cutlass";
    rev = "refs/tags/v3.5.0";
    sha256 = "sha256-D/s7eYsa5l/mfx73tE4mnFcTQdYqGmXa9d9TCryw4e4=";
  };
in

buildPythonPackage rec {
  pname = "vllm";
  version = "0.6.2";
  pyproject = true;

  stdenv = if cudaSupport then cudaPackages.backendStdenv else args.stdenv;

  src = fetchFromGitHub {
    owner = "vllm-project";
    repo = pname;
    rev = "refs/tags/v${version}";
    hash = "sha256-zUkqAPPhDRdN9rDQ2biCl1B+trV0xIHXub++v9zsQGo=";
  };

  patches = [
    ./0001-setup.py-don-t-ask-for-hipcc-version.patch
    ./0002-setup.py-nix-support-respect-cmakeFlags.patch
  ];

  # Ignore the python version check because it hard-codes minor versions and
  # lags behind `ray`'s python interpreter support
  postPatch = ''
    substituteInPlace CMakeLists.txt \
      --replace-fail \
        'set(PYTHON_SUPPORTED_VERSIONS' \
        'set(PYTHON_SUPPORTED_VERSIONS "${lib.versions.majorMinor python.version}"'
  '';

  nativeBuildInputs = [
    cmake
    ninja
    pythonRelaxDepsHook
    which
  ] ++ lib.optionals rocmSupport [ rocmPackages.hipcc ];

  build-system = [
    packaging
    setuptools
    wheel
  ];

  buildInputs =
    (lib.optionals cudaSupport (
      with cudaPackages;
      [
        cuda_cudart # cuda_runtime.h, -lcudart
        cuda_cccl
        libcusparse # cusparse.h
        libcusolver # cusolverDn.h
        cuda_nvcc
        cuda_nvtx
        libcublas
      ]
    ))
    ++ (lib.optionals rocmSupport (
      with rocmPackages;
      [
        clr
        rocthrust
        rocprim
        hipsparse
        hipblas
      ]
    ));

  dependencies =
    [
      aioprometheus
      fastapi
      lm-format-enforcer
      numpy
      openai
      outlines
      pandas
      prometheus-fastapi-instrumentator
      psutil
      py-cpuinfo
      pyarrow
      pydantic
      pyzmq
      ray
      sentencepiece
      tiktoken
      torch
      torchvision
      transformers
      uvicorn
      xformers
    ]
    ++ uvicorn.optional-dependencies.standard
    ++ aioprometheus.optional-dependencies.starlette
    ++ lib.optionals cudaSupport [
      cupy
      pynvml
    ];

  dontUseCmakeConfigure = true;
  cmakeFlags = [ (lib.cmakeFeature "FETCHCONTENT_SOURCE_DIR_CUTLASS" "${lib.getDev cutlass}") ];

  env =
    lib.optionalAttrs cudaSupport { CUDA_HOME = "${lib.getDev cudaPackages.cuda_nvcc}"; }
    // lib.optionalAttrs rocmSupport {
      # Otherwise it tries to enumerate host supported ROCM gfx archs, and that is not possible due to sandboxing.
      PYTORCH_ROCM_ARCH = lib.strings.concatStringsSep ";" rocmPackages.clr.gpuTargets;
      ROCM_HOME = "${rocmPackages.clr}";
    };

  pythonRelaxDeps = true;

  pythonImportsCheck = [ "vllm" ];

  meta = with lib; {
    description = "High-throughput and memory-efficient inference and serving engine for LLMs";
    changelog = "https://github.com/vllm-project/vllm/releases/tag/v${version}";
    homepage = "https://github.com/vllm-project/vllm";
    license = licenses.asl20;
    maintainers = with maintainers; [
      happysalada
      lach
    ];
    # RuntimeError: Unknown runtime environment
    broken = true;
    # broken = !cudaSupport && !rocmSupport;
  };
}