nixpkgs/pkgs/development/cuda-modules/nccl/default.nix
2024-02-16 19:04:55 +00:00

115 lines
3.2 KiB
Nix

# NOTE: Though NCCL is called within the cudaPackages package set, we avoid passing in
# the names of dependencies from that package set directly to avoid evaluation errors
# in the case redistributable packages are not available.
{
lib,
fetchFromGitHub,
python3,
which,
cudaPackages,
# passthru.updateScript
gitUpdater,
}:
let
inherit (cudaPackages)
autoAddOpenGLRunpathHook
backendStdenv
cuda_cccl
cuda_cudart
cuda_nvcc
cudaFlags
cudatoolkit
cudaVersion
;
in
backendStdenv.mkDerivation (
finalAttrs: {
pname = "nccl";
version = "2.20.3-1";
src = fetchFromGitHub {
owner = "NVIDIA";
repo = finalAttrs.pname;
rev = "v${finalAttrs.version}";
hash = "sha256-7gI1q6uN3saz/twwLjWl7XmMucYjvClDPDdbVpVM0vU=";
};
strictDeps = true;
outputs = [
"out"
"dev"
];
nativeBuildInputs =
[
which
autoAddOpenGLRunpathHook
python3
]
++ lib.optionals (lib.versionOlder cudaVersion "11.4") [cudatoolkit]
++ lib.optionals (lib.versionAtLeast cudaVersion "11.4") [cuda_nvcc];
buildInputs =
lib.optionals (lib.versionOlder cudaVersion "11.4") [cudatoolkit]
++ lib.optionals (lib.versionAtLeast cudaVersion "11.4") [
cuda_nvcc.dev # crt/host_config.h
cuda_cudart
]
# NOTE: CUDA versions in Nixpkgs only use a major and minor version. When we do comparisons
# against other version, like below, it's important that we use the same format. Otherwise,
# we'll get incorrect results.
# For example, lib.versionAtLeast "12.0" "12.0.0" == false.
++ lib.optionals (lib.versionAtLeast cudaVersion "12.0") [cuda_cccl];
env.NIX_CFLAGS_COMPILE = toString ["-Wno-unused-function"];
preConfigure = ''
patchShebangs ./src/device/generate.py
makeFlagsArray+=(
"NVCC_GENCODE=${lib.concatStringsSep " " cudaFlags.gencode}"
)
'';
makeFlags =
["PREFIX=$(out)"]
++ lib.optionals (lib.versionOlder cudaVersion "11.4") [
"CUDA_HOME=${cudatoolkit}"
"CUDA_LIB=${lib.getLib cudatoolkit}/lib"
"CUDA_INC=${lib.getDev cudatoolkit}/include"
]
++ lib.optionals (lib.versionAtLeast cudaVersion "11.4") [
"CUDA_HOME=${cuda_nvcc}"
"CUDA_LIB=${lib.getLib cuda_cudart}/lib"
"CUDA_INC=${lib.getDev cuda_cudart}/include"
];
enableParallelBuilding = true;
postFixup = ''
moveToOutput lib/libnccl_static.a $dev
'';
passthru.updateScript = gitUpdater {
inherit (finalAttrs) pname version;
rev-prefix = "v";
};
meta = with lib; {
description = "Multi-GPU and multi-node collective communication primitives for NVIDIA GPUs";
homepage = "https://developer.nvidia.com/nccl";
license = licenses.bsd3;
platforms = platforms.linux;
# NCCL is not supported on Jetson, because it does not use NVLink or PCI-e for inter-GPU communication.
# https://forums.developer.nvidia.com/t/can-jetson-orin-support-nccl/232845/9
badPlatforms = lib.optionals cudaFlags.isJetsonBuild [ "aarch64-linux" ];
maintainers =
with maintainers;
[
mdaiter
orivej
]
++ teams.cuda.members;
};
}
)