pytorch

987f7b38 - Revert "Deletes CCACHE_DISABLE and SCCACHE_DISABLE from nccl.cmake (#84007)"

Revert "Deletes CCACHE_DISABLE and SCCACHE_DISABLE from nccl.cmake (#84007)" This reverts commit 37d3db7579386e928f70815afa7b5a21ffb2fefd. After 37d3db75 landed, we observed errors like the following when building on the AWS cluster with cuda11.6: ``` ... FAILED: nccl_external-prefix/src/nccl_external-stamp/nccl_external-build nccl/lib/libnccl_static.a /scratch/dberard/nccl-bisect/pytorch/build/nccl_external-prefix/src/nccl_external-stamp/nccl_external-build /scratch/dberard/nccl-bisect/pytorch/build/nccl/lib/libnccl_static.a cd /scratch/dberard/nccl-bisect/pytorch/third_party/nccl/nccl && make -j24 -l24 CXX=/data/shared/bin/c++ CUDA_HOME=/usr/local/cuda-11.6 NVCC=/data/shared/bin/cuda/nvcc "NVCC_GENCODE=-gencode=arch=compute_70,code=sm_70 -gencode=arch=compute_80,code=sm_80" BUILDDIR=/scratch/dberard/nccl-bisect/pytorch/build/nccl VERBOSE=0 && /usr/local/lib/python3.6/dist-packages/cmake/data/bin/cmake -E touch /scratch/dberard/nccl-bisect/pytorch/build/nccl_external-prefix/src/nccl_external-stamp/nccl_external-build make -C src build BUILDDIR=/scratch/dberard/nccl-bisect/pytorch/build/nccl make[1]: Entering directory '/scratch/dberard/nccl-bisect/pytorch/third_party/nccl/nccl/src' make[2]: Entering directory '/scratch/dberard/nccl-bisect/pytorch/third_party/nccl/nccl/src/collectives/device' /usr/local/cuda-11.6/bin/crt/link.stub:111:118: error: redefinition of ‘const unsigned char def_module_id_str_5b13a6ab_11_sendrecv_cu_2dca4361_39257 []’ static const unsigned char __DEFSTRNAME(id) [] __attribute__((aligned(1))) __attribute__((section ("__nv_module_id"))) = "def " __TO_STRING(id); \ ^ /tmp/tmpxft_00010ccb_00000000-3_devlink.reg.c:4:1: note: in expansion of macro ‘DEFINE_REGISTER_FUNC’ DEFINE_REGISTER_FUNC(_5b13a6ab_11_sendrecv_cu_2dca4361_39257) ^~~~~~~~~~~~~~~~~~~~ /usr/local/cuda-11.6/bin/crt/link.stub:79:30: note: ‘const unsigned char def_module_id_str_5b13a6ab_11_sendrecv_cu_2dca4361_39257 [44]’ previously defined here #define __DEFSTRNAME_CORE(X) def_module_id_str##X ^ /usr/local/cuda-11.6/bin/crt/link.stub:80:25: note: in expansion of macro ‘__DEFSTRNAME_CORE’ #define __DEFSTRNAME(X) __DEFSTRNAME_CORE(X) ^~~~~~~~~~~~~~~~~ /usr/local/cuda-11.6/bin/crt/link.stub:111:28: note: in expansion of macro ‘__DEFSTRNAME’ static const unsigned char __DEFSTRNAME(id) [] __attribute__((aligned(1))) __attribute__((section ("__nv_module_id"))) = "def " __TO_STRING(id); \ ^~~~~~~~~~~~ /tmp/tmpxft_00010ccb_00000000-3_devlink.reg.c:3:1: note: in expansion of macro ‘DEFINE_REGISTER_FUNC’ DEFINE_REGISTER_FUNC(_5b13a6ab_11_sendrecv_cu_2dca4361_39257) ^~~~~~~~~~~~~~~~~~~~ ... ``` After reverting the issue goes away ghstack-source-id: aa98c460416eb6ba9cba4977a7b955dc36f91435 Pull Request resolved: https://github.com/pytorch/pytorch/pull/84209