{ stdenv, fetchFromGitHub, which, cudatoolkit, addOpenGLRunpath }: stdenv.mkDerivation rec { name = "nccl-${version}-cuda-${cudatoolkit.majorVersion}"; version = "2.4.8-1"; src = fetchFromGitHub { owner = "NVIDIA"; repo = "nccl"; rev = "v${version}"; sha256 = "05m66y64rgsdyybvjybhy6clikwv438b1m484ikai78fb2b7mvyq"; }; outputs = [ "out" "dev" ]; nativeBuildInputs = [ which addOpenGLRunpath ]; buildInputs = [ cudatoolkit ]; preConfigure = '' patchShebangs src/collectives/device/gen_rules.sh ''; makeFlags = [ "CUDA_HOME=${cudatoolkit}" "PREFIX=$(out)" ]; postFixup = '' moveToOutput lib/libnccl_static.a $dev # Set RUNPATH so that libnvidia-ml in /run/opengl-driver(-32)/lib can be found. # See the explanation in addOpenGLRunpath. addOpenGLRunpath $out/lib/lib*.so ''; NIX_CFLAGS_COMPILE = [ "-Wno-unused-function" ]; enableParallelBuilding = true; meta = with stdenv.lib; { description = "Multi-GPU and multi-node collective communication primitives for NVIDIA GPUs"; homepage = "https://developer.nvidia.com/nccl"; license = licenses.bsd3; platforms = [ "x86_64-linux" ]; maintainers = with maintainers; [ mdaiter orivej ]; }; }