Unverified Commit 7c342c55 authored by Gaétan Lepage's avatar Gaétan Lepage Committed by GitHub
Browse files

cudaPackages.cuda_crt: patch math_functions.h signatures,... 

cudaPackages.cuda_crt: patch math_functions.h signatures, python3Packages.deep-ep: cleanup, fix build on CUDA>=13.0 (#512796)
parents 76f759a9 6e46f918

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

+30 −1
Original line number Diff line number Diff line 
{ backendStdenv, buildRedist }:

{

  lib,

  backendStdenv,

  buildRedist,

  glibc,

}:

buildRedist {

  redistName = "cuda";

  pname = "cuda_crt";



  outputs = [ "out" ];



  # Fix compatibility with glibc 2.42:

  # CUDA >= 13.0 fixed sinpi/cospi (using __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER), but

  # rsqrt/rsqrtf in math_functions.h still lack noexcept, conflicting with glibc 2.42's

  # declarations.

  postInstall = lib.optionalString (lib.versionAtLeast glibc.version "2.42") ''

    nixLog "Patching math_functions.h rsqrt signatures to match glibc's ones"

    substituteInPlace "''${!outputInclude:?}/include/crt/math_functions.h" \

      --replace-fail \

        "rsqrt(double x);" \

        "rsqrt(double x) noexcept (true);" \

      --replace-fail \

        "rsqrtf(float x);" \

        "rsqrtf(float x) noexcept (true);"



    nixLog "Patching math_functions.hpp rsqrt signatures to match glibc's ones"

    substituteInPlace "''${!outputInclude:?}/include/crt/math_functions.hpp" \

      --replace-fail \

        "__func__(double rsqrt(const double a))" \

        "__func__(double rsqrt(const double a) throw())" \

      --replace-fail \

        "__func__(float rsqrtf(const float a))" \

        "__func__(float rsqrtf(const float a) throw())"

  '';



  brokenAssertions = [

    # TODO(@connorbaker): Build fails on x86 when using pkgsLLVM.

    #  .../include/crt/host_defines.h:67:2:

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

+6 −0
Original line number Diff line number Diff line
@@ -168,6 +168,11 @@ buildRedist (finalAttrs: { 
      # The cospi|sinpi|rsqrt function signatures in include/common/math_functions.h do not match

      # glibc 2.42's.

      # Indeed, there they are declared with noexcept(true) which is not the case in cuda_nvcc.

      # - In CUDA < 13.0, sinpi/cospi/rsqrt all lack exception specifiers.

      # - In CUDA >= 13.0, NVIDIA fixed sinpi/cospi (using __NV_IEC_60559_FUNCS_EXCEPTION_SPECIFIER)

      #   but rsqrt/rsqrtf still lack noexcept, so we only patch those.

      #   As the CRT headers (including math_functions.h) moved to the cuda_crt package, the glibc

      #   2.42 compatibility patch is applied there instead.

      + lib.optionalString (cudaOlder "13.0" && lib.versionAtLeast glibc.version "2.42") ''

        nixLog "Patching math_functions.h signatures to match glibc's ones"

        substituteInPlace "''${!outputInclude:?}/include/crt/math_functions.h" \
@@ -213,6 +218,7 @@ buildRedist (finalAttrs: { 
            "__func__(float cospif(const float a))" \

            "__func__(float cospif(const float a) throw())"

      ''



      # Fix clang CUDA compilation: host_defines.h redefines __noinline__ as

      # __attribute__((noinline)), which conflicts with libstdc++ >=12 using

      # __attribute__((__noinline__)) — the macro expands to

pkgs/development/python-modules/deep-ep/default.nix

+17 −20
Original line number Diff line number Diff line
@@ -8,7 +8,6 @@ 
  setuptools,



  # env

  symlinkJoin,

  cudaPackages,



  # buildInputs
@@ -20,11 +19,6 @@ 
  cudaSupport ? config.cudaSupport,

}:

let

  inherit (lib)

    getBin

    getInclude

    ;



  minSupportedCudaCapability = "8.0"; # build fails with 7.5



  minCudaCapability = builtins.head (
@@ -66,31 +60,34 @@ buildPythonPackage.override { inherit (torch) stdenv; } (finalAttrs: { 


      DISABLE_SM90_FEATURES = if disableSm90Features then "1" else "0";



      CUDA_HOME = symlinkJoin {

        name = "cuda-redist";

        paths = with cudaPackages; [

          (getBin cuda_nvcc)



          (getInclude cuda_cccl) # <nv/target>

          (getInclude cuda_cudart) # cuda_runtime.h

          (getInclude libcublas) # cublas_v2.h

          (getInclude libcusolver) # cusolverDn.h

          (getInclude libcusparse) # cusparse.h

        ];

      };

      CUDA_HOME = (lib.getBin cudaPackages.cuda_nvcc).outPath;

    }



    # nvshmem must be disabled (unsetting NVSHMEM_DIR) when supporting <9.0 capabilities

    # https://github.com/deepseek-ai/DeepEP/blob/v1.2.1/setup.py#L65

    // lib.optionalAttrs (!disableSm90Features) {

      NVSHMEM_DIR = (getInclude cudaPackages.libnvshmem).outPath;

      NVSHMEM_DIR = (lib.getInclude cudaPackages.libnvshmem).outPath;

    }

  );



  buildInputs = [

    pybind11

    rdma-core

  ];

  ]

  ++ lib.optionals cudaSupport (

    with cudaPackages;

    [

      cuda_cccl # <nv/target>

      cuda_cudart # cuda_runtime.h

      libcublas # cublas_v2.h

      libcusolver # cusolverDn.h

      libcusparse # cusparse.h

    ]

    # On CUDA >=13.0, crt/host_config.h is shipped in cudaPackages.cuda_crt

    ++ lib.optionals cuda_crt.meta.available [

      cuda_crt # crt/host_config.h

    ]

  );



  pythonImportsCheck = [ "deep_ep" ];