Commit a96e5f0a authored by cfhammill's avatar cfhammill Committed by Someone Serge
Browse files

python311Packages.vllm: 0.3.2->0.5.2 

vllm remove comment, minor version bump

use refs/tags and format=pyproject

use @args and realphabetize inputs

move stdenv, mv cuda_home, add/clean cuda deps, restore comment

nixfmt rfc style

remove stale comment
parent 5377367b

pkgs/development/python-modules/outlines/default.nix

+10 −3
Original line number Diff line number Diff line
@@ -6,9 +6,12 @@ 
  setuptools-scm,

  interegular,

  cloudpickle,

  datasets,

  diskcache,

  joblib,

  jsonschema,

  pyairports,

  pycountry,

  pydantic,

  lark,

  nest-asyncio,
@@ -38,6 +41,7 @@ buildPythonPackage rec { 
  propagatedBuildInputs = [

    interegular

    cloudpickle

    datasets

    diskcache

    joblib

    jsonschema
@@ -48,16 +52,19 @@ buildPythonPackage rec { 
    scipy

    torch

    transformers

    pycountry

    pyairports

  ];



  pythonImportsCheck = [ "outlines" ];

  checkPhase = ''

    export HOME=$(mktemp -d)

    python3 -c 'import outlines'

  '';



  meta = with lib; {

    description = "Structured text generation";

    homepage = "https://github.com/outlines-dev/outlines";

    license = licenses.asl20;

    maintainers = with maintainers; [ lach ];

    # Missing dependencies since the last update

    broken = true;

  };

}

pkgs/development/python-modules/vllm/default.nix

+54 −37
Original line number Diff line number Diff line 
{

  lib,

  stdenv,

  python,

  buildPythonPackage,

  pythonRelaxDepsHook,

  fetchFromGitHub,

  which,

  ninja,

  cmake,

  packaging,

  setuptools,

  torch,
@@ -23,6 +26,13 @@ 
  pydantic,

  aioprometheus,

  pynvml,

  openai,

  pyzmq,

  tiktoken,

  torchvision,

  py-cpuinfo,

  lm-format-enforcer,

  prometheus-fastapi-instrumentator,

  cupy,

  writeShellScript,
@@ -34,47 +44,41 @@ 
  rocmSupport ? config.rocmSupport,

  rocmPackages ? { },

  gpuTargets ? [ ],

}:

}@args:



let

  stdenv_pkg = stdenv;

  cutlass = fetchFromGitHub {

    owner = "NVIDIA";

    repo = "cutlass";

    rev = "refs/tags/v3.5.0";

    sha256 = "sha256-D/s7eYsa5l/mfx73tE4mnFcTQdYqGmXa9d9TCryw4e4=";

  };

in



buildPythonPackage rec {

  pname = "vllm";

  version = "0.3.3";

  format = "pyproject";

  version = "0.5.3.post1";

  pyproject = true;



  src = fetchFromGitHub {

    owner = "vllm-project";

    repo = pname;

    rev = "v${version}";

    hash = "sha256-LU5pCPVv+Ws9dL8oWL1sJGzwQKI1IFk2A1I6TP9gXL4=";

    rev = "refs/tags/v${version}";

    hash = "sha256-++DK2Y2zz+1KrEcdQc5XFrSjc7fCwMD2DQ/RqY7PoFU=";

  };



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



  # Otherwise it tries to enumerate host supported ROCM gfx archs, and that is not possible due to sandboxing.

  PYTORCH_ROCM_ARCH = lib.optionalString rocmSupport (

    lib.strings.concatStringsSep ";" rocmPackages.clr.gpuTargets

  );



  # cupy-cuda12x is the same wheel as cupy, but built with cuda dependencies, we already have it set up

  # like that in nixpkgs. Version upgrade is due to upstream shenanigans

  # https://github.com/vllm-project/vllm/pull/2845/commits/34a0ad7f9bb7880c0daa2992d700df3e01e91363

  #

  # hipcc --version works badly on NixOS due to unresolved paths.

  # Unclear why pythonRelaxDeps doesn't work here, but on last attempt, it didn't.

  postPatch =

    ''

      substituteInPlace requirements.txt \

        --replace "xformers == 0.0.23.post1" "xformers"

      substituteInPlace requirements.txt \

        --replace "cupy-cuda12x == 12.1.0" "cupy"

      substituteInPlace requirements-build.txt \

        --replace "torch==2.1.2" "torch"

      substituteInPlace pyproject.toml \

        --replace "torch == 2.1.2" "torch"

      substituteInPlace requirements.txt \

        --replace "torch == 2.1.2" "torch"

      substituteInPlace setup.py \

        --replace 'cmake_args = [' 'cmake_args = ["-DFETCHCONTENT_SOURCE_DIR_CUTLASS=${cutlass}",'

    ''

    + lib.optionalString rocmSupport ''

      substituteInPlace setup.py \
@@ -82,17 +86,19 @@ buildPythonPackage rec { 
    '';



  preBuild =

    lib.optionalString cudaSupport ''

      export CUDA_HOME=${cudaPackages.cuda_nvcc}

    ''

    + lib.optionalString rocmSupport ''

    lib.optionalString rocmSupport ''

      export ROCM_HOME=${rocmPackages.clr}

      export PATH=$PATH:${rocmPackages.hipcc}

    ''

    + lib.optionalString cudaSupport ''

      export CUDA_HOME=${cudaPackages.cuda_nvcc}

    '';



  nativeBuildInputs = [

    cmake

    ninja

    packaging

    pythonRelaxDepsHook

    setuptools

    torch

    wheel
@@ -104,10 +110,12 @@ buildPythonPackage rec { 
      with cudaPackages;

      [

        cuda_cudart # cuda_runtime.h, -lcudart

        cuda_cccl # <thrust/*>

        cuda_cccl

        libcusparse # cusparse.h

        libcublas # cublas_v2.h

        libcusolver # cusolverDn.h

        cuda_nvcc

        cuda_nvtx

        libcublas

      ]

    ))

    ++ (lib.optionals rocmSupport (
@@ -123,29 +131,38 @@ buildPythonPackage rec { 


  propagatedBuildInputs =

    [

      psutil

      ray

      aioprometheus

      fastapi

      lm-format-enforcer

      numpy

      openai

      outlines

      pandas

      prometheus-fastapi-instrumentator

      psutil

      py-cpuinfo

      pyarrow

      pydantic

      pyzmq

      ray

      sentencepiece

      numpy

      tiktoken

      torch

      torchvision

      transformers

      outlines

      xformers

      fastapi

      uvicorn

      pydantic

      aioprometheus

      xformers

    ]

    ++ uvicorn.optional-dependencies.standard

    ++ aioprometheus.optional-dependencies.starlette

    ++ lib.optionals cudaSupport [

      pynvml

      cupy

      pynvml

    ];



  stdenv = if cudaSupport then cudaPackages.backendStdenv else stdenv_pkg;

  dontUseCmakeConfigure = true;



  pythonRelaxDeps = true;



  pythonImportsCheck = [ "vllm" ];

pkgs/top-level/python-packages.nix

+4 −0
Original line number Diff line number Diff line
@@ -7167,6 +7167,8 @@ self: super: with self; { 
    llvm = pkgs.llvm_14;

  };













  lm-format-enforcer = callPackage ../development/python-modules/lm-format-enforcer { };



























  lmcloud = callPackage ../development/python-modules/lmcloud { };





























  lmdb = callPackage ../development/python-modules/lmdb {










@@ -10904,6 +10906,8 @@ self: super: with self; {



























  pyairnow = callPackage ../development/python-modules/pyairnow { };



























  pyairports = callPackage ../development/python-modules/pyairports { };



























  pyairvisual = callPackage ../development/python-modules/pyairvisual { };





























  pyais = callPackage ../development/python-modules/pyais { };