Commit fca9a6cb authored by Ari Lotter's avatar Ari Lotter
Browse files

cudaPackages.nccl: 2.26.2 -> 2.27.6 

nccl 2.27.6 changed its CXXSTD from cpp11 to 14, so the workaround is no
longer required.

2.27 adds a new "symmetric" set of APIs which require their generated
script to have its shebang patched.
parent e7515332

pkgs/development/cuda-modules/packages/nccl.nix

+3 −8
Original line number Diff line number Diff line
@@ -24,10 +24,10 @@ let 
    ;

  # versions 2.26+ with CUDA 11.x error with

  # fatal error: cuda/atomic: No such file or directory

  version = if cudaAtLeast "12.0" then "2.26.2-1" else "2.25.1-1";

  version = if cudaAtLeast "12.0" then "2.27.6-1" else "2.25.1-1";

  hash =

    if cudaAtLeast "12.0" then

      "sha256-iLEuru3gaNLcAdH4V8VIv3gjdTGjgb2/Mr5UKOh69N4="

      "sha256-/BiLSZaBbVIqOfd8nQlgUJub0YR3SR4B93x2vZpkeiU="

    else

      "sha256-3snh0xdL9I5BYqdbqdl+noizJoI38mZRVOJChgEE1I8=";

in
@@ -74,12 +74,7 @@ backendStdenv.mkDerivation (finalAttrs: { 


  postPatch = ''

    patchShebangs ./src/device/generate.py

  ''

  # CUDA 12.8 uses GCC 14 and we need to bump C++ standard to C++14

  # in order to work with new constexpr handling

  + lib.optionalString (cudaAtLeast "12.8") ''

    substituteInPlace ./makefiles/common.mk \

      --replace-fail "-std=c++11" "-std=c++14"

    patchShebangs ./src/device/symmetric/generate.py

  '';



  makeFlags = [