python312Packages.triton*: repair package

, libedit

, libffi

, libpfm

, lit

, mpfr

, zlib

, ncurses

  isNative = stdenv.hostPlatform == stdenv.buildPlatform;

in stdenv.mkDerivation (finalAttrs: {

  pname = "triton-llvm";

  version = "17.0.0-c5dede880d17";

  version = "19.1.0-rc1"; # One of the tags at https://github.com/llvm/llvm-project/commit/10dc3a8e916d73291269e5e2b82dd22681489aa1

  outputs = [

    "out"

  src = fetchFromGitHub {

    owner = "llvm";

    repo = "llvm-project";

    rev = "c5dede880d175f7229c9b2923f4753e12702305d";

    hash = "sha256-v4r3+7XVFK+Dzxt/rErZNJ9REqFO3JmGN4X4vZ+77ew=";

    rev = "10dc3a8e916d73291269e5e2b82dd22681489aa1";

    hash = "sha256-9DPvcFmhzw6MipQeCQnr35LktW0uxtEL8axMMPXIfWw=";

  };

  nativeBuildInputs = [

    doxygen

    sphinx

    python3Packages.recommonmark

    python3Packages.myst-parser

  ];

  buildInputs = [

    rm test/tools/llvm-exegesis/AArch64/latency-by-opcode-name.s

  '';

  postInstall = lib.optionalString (!isNative) ''

  postInstall = ''

    cp ${lib.getExe lit} $out/bin/llvm-lit

  '' + (lib.optionalString (!isNative) ''

    cp -a NATIVE/bin/llvm-config $out/bin/llvm-config-native

  '';

  '');

  doCheck = buildTests;

From 10f3d49aa6084d1b9b9624017cce7df106b9fb7e Fri Jul 19 00:00:00 2024

From: derdennisop <just@patch.local>

Date: Fri, 19 jul 2024 00:00:00 +0100

Subject: [PATCH] ptxas: disable version key for non-cuda targets

---

 python/setup.py | 4 ++--

 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/python/setup.py b/python/setup.py

index 18764ec13..b3bb5b60a 100644

index d55972b4b..bd875a701 100644

--- a/python/setup.py

+++ b/python/setup.py

@@ -269,10 +269,6 @@ class CMakeBuild(build_ext):

         subprocess.check_call(["cmake", self.base_dir] + cmake_args, cwd=cmake_dir, env=env)

         subprocess.check_call(["cmake", "--build", "."] + build_args, cwd=cmake_dir)

@@ -437,54 +117,5 @@

 with open(nvidia_version_path, "r") as nvidia_version_file:

     NVIDIA_TOOLCHAIN_VERSION = nvidia_version_file.read().strip()

-download_and_copy(

-    name="ptxas",

-    src_path="bin/ptxas",

-    variable="TRITON_PTXAS_PATH",

-    version=NVIDIA_TOOLCHAIN_VERSION,

-    url_func=lambda arch, version:

-    f"https://anaconda.org/nvidia/cuda-nvcc/{version}/download/linux-{arch}/cuda-nvcc-{version}-0.tar.bz2",

-)

-download_and_copy(

-    name="cuobjdump",

-    src_path="bin/cuobjdump",

-    variable="TRITON_CUOBJDUMP_PATH",

-    version=NVIDIA_TOOLCHAIN_VERSION,

-    url_func=lambda arch, version:

-    f"https://anaconda.org/nvidia/cuda-cuobjdump/{version}/download/linux-{arch}/cuda-cuobjdump-{version}-0.tar.bz2",

-)

-download_and_copy(

-    name="nvdisasm",

-    src_path="bin/nvdisasm",

-    variable="TRITON_NVDISASM_PATH",

-    version=NVIDIA_TOOLCHAIN_VERSION,

-    url_func=lambda arch, version:

-    f"https://anaconda.org/nvidia/cuda-nvdisasm/{version}/download/linux-{arch}/cuda-nvdisasm-{version}-0.tar.bz2",

-)

-download_and_copy(

-    name="cudacrt",

-    src_path="include",

-    variable="TRITON_CUDACRT_PATH",

-    version=NVIDIA_TOOLCHAIN_VERSION,

-    url_func=lambda arch, version:

-    f"https://anaconda.org/nvidia/cuda-nvcc/{version}/download/linux-{arch}/cuda-nvcc-{version}-0.tar.bz2",

-)

-download_and_copy(

-    name="cudart",

-    src_path="include",

-    variable="TRITON_CUDART_PATH",

-    version=NVIDIA_TOOLCHAIN_VERSION,

-    url_func=lambda arch, version:

-    f"https://anaconda.org/nvidia/cuda-cudart-dev/{version}/download/linux-{arch}/cuda-cudart-dev-{version}-0.tar.bz2",

-)

-download_and_copy(

-    name="cupti",

-    src_path="include",

-    variable="TRITON_CUPTI_PATH",

-    version=NVIDIA_TOOLCHAIN_VERSION,

-    url_func=lambda arch, version:

-    f"https://anaconda.org/nvidia/cuda-cupti/{version}/download/linux-{arch}/cuda-cupti-{version}-0.tar.bz2",

-)

-

-download_and_copy_ptxas()

-

-

 setup(

     name="triton",

     version="2.1.0",

 backends = [*BackendInstaller.copy(["nvidia", "amd"]), *BackendInstaller.copy_externals()]

From 10f3d49aa6084d1b9b9624017cce7df106b9fb7e Mon Sep 17 00:00:00 2001

From: Yaroslav Bolyukin <iam@lach.pw>

Date: Tue, 6 Feb 2024 13:51:28 +0100

Subject: [PATCH] ptxas: disable version key for non-cuda targets

---

 python/triton/runtime/jit.py | 4 ++--

 1 file changed, 2 insertions(+), 2 deletions(-)

diff --git a/python/triton/runtime/jit.py b/python/triton/runtime/jit.py

index d55972b4b..bd875a701 100644

--- a/python/triton/runtime/jit.py

+++ b/python/triton/runtime/jit.py

@@ -117,8 +117,8 @@ def version_key():

         with open(lib.module_finder.find_spec(lib.name).origin, "rb") as f:

             contents += [hashlib.md5(f.read()).hexdigest()]

     # ptxas version

-    ptxas = path_to_ptxas()[0]

-    ptxas_version = hashlib.md5(subprocess.check_output([ptxas, "--version"])).hexdigest()

+    # ptxas = path_to_ptxas()[0]

+    ptxas_version = "noptxas"

     return '-'.join(TRITON_VERSION) + '-' + ptxas_version + '-' + '-'.join(contents)

-- 

2.43.0

  cudaPackages,

  buildPythonPackage,

  fetchurl,

  isPy38,

  isPy39,

  isPy310,

  isPy311,

  python,

  pythonOlder,

  autoPatchelfHook,

  filelock,

  lit,

    in

    fetchurl srcs;

  disabled = !(isPy38 || isPy39 || isPy310 || isPy311);

  disabled = pythonOlder "3.8";

  pythonRemoveDeps = [

    "cmake"

{

  lib,

  config,

  addDriverRunpath,

  buildPythonPackage,

  fetchFromGitHub,

  setuptools,

  cmake,

  ninja,

  pybind11,

  config,

  cudaPackages,

  fetchFromGitHub,

  filelock,

  gtest,

  zlib,

  ncurses,

  libxml2,

  lit,

  llvm,

  filelock,

  torchWithRocm,

  ncurses,

  ninja,

  pybind11,

  python,

  runCommand,

  cudaPackages,

  setuptools,

  torchWithRocm,

  zlib,

  cudaSupport ? config.cudaSupport,

}:

let

  ptxas = lib.getExe' cudaPackages.cuda_nvcc "ptxas"; # Make sure cudaPackages is the right version each update (See python/setup.py)

in

buildPythonPackage rec {

buildPythonPackage {

  pname = "triton";

  version = "3.0.0";

  pyproject = true;

  src = fetchFromGitHub {

    owner = "triton-lang";

    repo = pname;

    repo = "triton";

    # latest branch commit from https://github.com/triton-lang/triton/commits/release/3.0.x/

    rev = "91f24d87e50cb748b121a6c24e65a01187699c22";

    hash = "sha256-L5KqiR+TgSyKjEBlkE0yOU1pemMHFk2PhEmxLdbbxUU=";

  };

  patches =

    [

      # Upstream startded pinning CUDA version and falling back to downloading from Conda

      # in https://github.com/triton-lang/triton/pull/1574/files#diff-eb8b42d9346d0a5d371facf21a8bfa2d16fb49e213ae7c21f03863accebe0fcfR120-R123

      ./0000-dont-download-ptxas.patch

    ]

    ++ lib.optionals (!cudaSupport) [

      # triton wants to get ptxas version even if ptxas is not

      # used, resulting in ptxas not found error.

      ./0001-ptxas-disable-version-key-for-non-cuda-targets.patch

    ];

  # triton wants to download every dependency, even if we are not using cuda.

  patches = lib.optionals (!cudaSupport) [ ./0000-dont-download-ptxas.patch ];

  postPatch =

    let

      quote = x: ''"${x}"'';

      subs.ldFlags =

        let

          # Bash was getting weird without linting,

          # but basically upstream contains [cc, ..., "-lcuda", ...]

          # and we replace it with [..., "-lcuda", "-L/run/opengl-driver/lib", "-L$stubs", ...]

          old = [ "-lcuda" ];

          new = [

            "-lcuda"

            "-L${addDriverRunpath.driverLink}"

            "-L${cudaPackages.cuda_cudart}/lib/stubs/"

          ];

        in

        {

          oldStr = lib.concatMapStringsSep ", " quote old;

          newStr = lib.concatMapStringsSep ", " quote new;

        };

    in

    ''

      # Use our `cmakeFlags` instead and avoid downloading dependencies

      # remove any downloads

      substituteInPlace python/setup.py \

        --replace "= get_thirdparty_packages(triton_cache_path)" "= os.environ[\"cmakeFlags\"].split()"

      # Already defined in llvm, when built with -DLLVM_INSTALL_UTILS

      substituteInPlace bin/CMakeLists.txt \

        --replace "add_subdirectory(FileCheck)" ""

        --replace-fail "get_json_package_info(), get_pybind11_package_info()" ""\

        --replace-fail "get_pybind11_package_info(), get_llvm_package_info()" ""\

        --replace-fail 'packages += ["triton/profiler"]' ""\

        --replace-fail "curr_version != version" "False"

      # Don't fetch googletest

      substituteInPlace unittest/CMakeLists.txt \

        --replace "include (\''${CMAKE_CURRENT_SOURCE_DIR}/googletest.cmake)" ""\

        --replace "include(GoogleTest)" "find_package(GTest REQUIRED)"

      cat << \EOF >> python/triton/common/build.py

      def libcuda_dirs():

          return [ "${addDriverRunpath.driverLink}/lib" ]

      EOF

    ''

    + lib.optionalString cudaSupport ''

      # Use our linker flags

      substituteInPlace python/triton/common/build.py \

        --replace '${subs.ldFlags.oldStr}' '${subs.ldFlags.newStr}'

        --replace-fail "include (\''${CMAKE_CURRENT_SOURCE_DIR}/googletest.cmake)" ""\

        --replace-fail "include(GoogleTest)" "find_package(GTest REQUIRED)"

    '';

  nativeBuildInputs = [

  ];

  # Avoid GLIBCXX mismatch with other cuda-enabled python packages

  preConfigure =

    ''

  preConfigure = ''

    # Ensure that the build process uses the requested number of cores

    export MAX_JOBS="$NIX_BUILD_CORES"

    # The rest (including buildPhase) is relative to ./python/

    cd python

    ''

    + lib.optionalString cudaSupport ''

      export CC=${cudaPackages.backendStdenv.cc}/bin/cc;

      export CXX=${cudaPackages.backendStdenv.cc}/bin/c++;

      # Work around download_and_copy_ptxas()

      mkdir -p $PWD/triton/third_party/cuda/bin

      ln -s ${ptxas} $PWD/triton/third_party/cuda/bin

  '';

  env = {

    TRITON_BUILD_PROTON = "OFF";

  } // lib.optionalAttrs cudaSupport {

    CC = "${cudaPackages.backendStdenv.cc}/bin/cc";

    CXX = "${cudaPackages.backendStdenv.cc}/bin/c++";

    TRITON_PTXAS_PATH = lib.getExe' cudaPackages.cuda_nvcc "ptxas"; # Make sure cudaPackages is the right version each update (See python/setup.py)

    TRITON_CUOBJDUMP_PATH = cudaPackages.cuda_cuobjdump;

    TRITON_NVDISASM_PATH = cudaPackages.cuda_nvdisasm;

    TRITON_CUDACRT_PATH = cudaPackages.cuda_nvcc;

    TRITON_CUDART_PATH = cudaPackages.cuda_cudart;

    TRITON_CUPTI_PATH = cudaPackages.cuda_cupti;

  };

  # CMake is run by setup.py instead

  dontUseCmakeConfigure = true;

  # Setuptools (?) strips runpath and +x flags. Let's just restore the symlink

  postFixup = lib.optionalString cudaSupport ''

    rm -f $out/${python.sitePackages}/triton/third_party/cuda/bin/ptxas

    ln -s ${ptxas} $out/${python.sitePackages}/triton/third_party/cuda/bin/ptxas

  '';

  checkInputs = [ cmake ]; # ctest

  dontUseSetuptoolsCheck = true;