mirror of
https://github.com/pytorch/pytorch.git
synced 2025-10-20 12:54:11 +08:00
[RELAND] Always build USE_DISTRIBUTED (#160449) and Make distributed modules importable even when backend not built (#159889) (#162594)
Summary: Original: D81957844 and D81957923 Also, https://github.com/pytorch/pytorch/pull/162142 is patched in as well #buildall Test Plan: sandcastle and oss ci Rollback Plan: Reviewed By: H-Huang Pull Request resolved: https://github.com/pytorch/pytorch/pull/162594 Approved by: https://github.com/H-Huang, https://github.com/dcci
This commit is contained in:
Edward Yang
committed by
PyTorch MergeBot
PyTorch MergeBot
parent
4027e97791
commit
09cb34c1dc
@ -540,11 +540,9 @@ if(NOT INTERN_BUILD_MOBILE AND NOT BUILD_LITE_INTERPRETER)
|
||||
${TORCH_SRC_DIR}/csrc/utils/byte_order.cpp
|
||||
)
|
||||
|
||||
if(USE_DISTRIBUTED)
|
||||
append_filelist("libtorch_distributed_base_sources" TORCH_SRCS)
|
||||
if(NOT WIN32)
|
||||
append_filelist("libtorch_distributed_extra_sources" TORCH_SRCS)
|
||||
endif()
|
||||
append_filelist("libtorch_distributed_base_sources" TORCH_SRCS)
|
||||
if(NOT WIN32)
|
||||
append_filelist("libtorch_distributed_extra_sources" TORCH_SRCS)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
@ -575,32 +573,30 @@ if(USE_CUDA)
|
||||
list(APPEND Caffe2_GPU_SRCS
|
||||
${TORCH_SRC_DIR}/csrc/cuda/nccl.cpp)
|
||||
endif()
|
||||
if(USE_DISTRIBUTED)
|
||||
append_filelist("libtorch_cuda_distributed_base_sources" Caffe2_GPU_SRCS)
|
||||
if(NOT WIN32)
|
||||
append_filelist("libtorch_cuda_distributed_extra_sources" Caffe2_GPU_SRCS)
|
||||
set_source_files_properties(
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/ProcessGroupNCCL.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/cuda/utils.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/intra_node_comm.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CudaDMAConnectivity.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CUDASymmetricMemory.cu
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CUDASymmetricMemoryOps.cu
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CUDASymmetricMemoryUtils.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/NCCLSymmetricMemory.cu
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/cuda_mem_pool.cpp
|
||||
PROPERTIES COMPILE_FLAGS "-DPYTORCH_C10_DRIVER_API_SUPPORTED=1"
|
||||
)
|
||||
endif()
|
||||
append_filelist("libtorch_cuda_distributed_base_sources" Caffe2_GPU_SRCS)
|
||||
if(NOT WIN32)
|
||||
append_filelist("libtorch_cuda_distributed_extra_sources" Caffe2_GPU_SRCS)
|
||||
set_source_files_properties(
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/ProcessGroupNCCL.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/cuda/utils.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/intra_node_comm.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CudaDMAConnectivity.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CUDASymmetricMemory.cu
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CUDASymmetricMemoryOps.cu
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/CUDASymmetricMemoryUtils.cpp
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/NCCLSymmetricMemory.cu
|
||||
${TORCH_SRC_DIR}/csrc/distributed/c10d/symm_mem/cuda_mem_pool.cpp
|
||||
PROPERTIES COMPILE_FLAGS "-DPYTORCH_C10_DRIVER_API_SUPPORTED=1"
|
||||
)
|
||||
endif()
|
||||
|
||||
set(ASYNC_MM_FILE "${TORCH_SRC_DIR}/csrc/distributed/c10d/cuda/AsyncMM.cu")
|
||||
# Disable the warning to make cutlass warp-specialized cooperative kernel build for gcc-9
|
||||
if(CMAKE_COMPILER_IS_GNUCXX)
|
||||
set_source_files_properties(${ASYNC_MM_FILE} PROPERTIES COMPILE_FLAGS "-Wno-unused-but-set-variable")
|
||||
endif()
|
||||
if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.0 AND CUDA_NVCC_FLAGS MATCHES ".*compute_90.*")
|
||||
set_source_files_properties(${ASYNC_MM_FILE} PROPERTIES COMPILE_FLAGS "-gencode arch=compute_90a,code=sm_90a")
|
||||
endif()
|
||||
set(ASYNC_MM_FILE "${TORCH_SRC_DIR}/csrc/distributed/c10d/cuda/AsyncMM.cu")
|
||||
# Disable the warning to make cutlass warp-specialized cooperative kernel build for gcc-9
|
||||
if(CMAKE_COMPILER_IS_GNUCXX)
|
||||
set_source_files_properties(${ASYNC_MM_FILE} PROPERTIES COMPILE_FLAGS "-Wno-unused-but-set-variable")
|
||||
endif()
|
||||
if(CMAKE_CUDA_COMPILER_VERSION VERSION_GREATER_EQUAL 12.0 AND CUDA_NVCC_FLAGS MATCHES ".*compute_90.*")
|
||||
set_source_files_properties(${ASYNC_MM_FILE} PROPERTIES COMPILE_FLAGS "-gencode arch=compute_90a,code=sm_90a")
|
||||
endif()
|
||||
set_source_files_properties(
|
||||
${TORCH_ROOT}/aten/src/ATen/cuda/detail/LazyNVRTC.cpp
|
||||
@ -633,11 +629,9 @@ if(USE_ROCM)
|
||||
list(APPEND Caffe2_HIP_SRCS
|
||||
${TORCH_SRC_DIR}/csrc/cuda/nccl.cpp)
|
||||
endif()
|
||||
if(USE_DISTRIBUTED)
|
||||
append_filelist("libtorch_cuda_distributed_base_sources" Caffe2_HIP_SRCS)
|
||||
if(NOT WIN32)
|
||||
append_filelist("libtorch_cuda_distributed_extra_sources" Caffe2_HIP_SRCS)
|
||||
endif()
|
||||
append_filelist("libtorch_cuda_distributed_base_sources" Caffe2_HIP_SRCS)
|
||||
if(NOT WIN32)
|
||||
append_filelist("libtorch_cuda_distributed_extra_sources" Caffe2_HIP_SRCS)
|
||||
endif()
|
||||
# caffe2_nvrtc's stubs to driver APIs are useful for HIP.
|
||||
# See NOTE [ ATen NVRTC Stub and HIP ]
|
||||
@ -1358,12 +1352,10 @@ if(BUILD_TEST)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/jit ${CMAKE_BINARY_DIR}/test_jit)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/nativert ${CMAKE_BINARY_DIR}/test_nativert)
|
||||
add_subdirectory(${TORCH_ROOT}/test/inductor ${CMAKE_BINARY_DIR}/test_inductor)
|
||||
if(USE_DISTRIBUTED)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/c10d ${CMAKE_BINARY_DIR}/test_cpp_c10d)
|
||||
if(NOT WIN32)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/dist_autograd ${CMAKE_BINARY_DIR}/dist_autograd)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/rpc ${CMAKE_BINARY_DIR}/test_cpp_rpc)
|
||||
endif()
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/c10d ${CMAKE_BINARY_DIR}/test_cpp_c10d)
|
||||
if(NOT WIN32)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/dist_autograd ${CMAKE_BINARY_DIR}/dist_autograd)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/rpc ${CMAKE_BINARY_DIR}/test_cpp_rpc)
|
||||
endif()
|
||||
if(NOT NO_API)
|
||||
add_subdirectory(${TORCH_ROOT}/test/cpp/api ${CMAKE_BINARY_DIR}/test_api)
|
||||
@ -1468,46 +1460,40 @@ if(BUILD_LITE_INTERPRETER)
|
||||
endif()
|
||||
endif()
|
||||
|
||||
|
||||
# Pass USE_DISTRIBUTED to torch_cpu, as some codes in jit/pickler.cpp and
|
||||
# jit/unpickler.cpp need to be compiled only when USE_DISTRIBUTED is set
|
||||
if(USE_DISTRIBUTED)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_DISTRIBUTED)
|
||||
if(USE_GLOO AND USE_C10D_GLOO)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_C10D_GLOO)
|
||||
if(USE_GLOO AND USE_C10D_GLOO)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_C10D_GLOO)
|
||||
endif()
|
||||
if(USE_UCC AND USE_C10D_UCC)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_C10D_UCC)
|
||||
if(USE_CUDA)
|
||||
target_compile_definitions(torch_cuda PUBLIC USE_C10D_UCC)
|
||||
endif()
|
||||
if(USE_UCC AND USE_C10D_UCC)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_C10D_UCC)
|
||||
if(USE_CUDA)
|
||||
target_compile_definitions(torch_cuda PUBLIC USE_C10D_UCC)
|
||||
endif()
|
||||
endif()
|
||||
if(USE_NCCL AND USE_C10D_NCCL)
|
||||
if(USE_ROCM)
|
||||
target_compile_definitions(torch_hip PUBLIC USE_C10D_NCCL)
|
||||
else()
|
||||
target_compile_definitions(torch_cuda PUBLIC USE_C10D_NCCL)
|
||||
endif()
|
||||
if(USE_NCCL AND USE_C10D_NCCL)
|
||||
if(USE_ROCM)
|
||||
target_compile_definitions(torch_hip PUBLIC USE_C10D_NCCL)
|
||||
else()
|
||||
target_compile_definitions(torch_cuda PUBLIC USE_C10D_NCCL)
|
||||
endif()
|
||||
endif()
|
||||
if(USE_MPI AND USE_C10D_MPI)
|
||||
if(CMAKE_CXX_COMPILER_ID MATCHES "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
|
||||
set_source_files_properties(
|
||||
"${TORCH_SRC_DIR}/csrc/distributed/c10d/ProcessGroupMPI.cpp"
|
||||
PROPERTIES COMPILE_FLAGS -Wno-deprecated-declarations)
|
||||
endif()
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_C10D_MPI)
|
||||
endif()
|
||||
# Pass USE_RPC in order to reduce use of
|
||||
# #if defined(USE_DISTRIBUTED) && !defined(_WIN32)
|
||||
# need to be removed when RPC is supported
|
||||
if(NOT WIN32)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_RPC)
|
||||
endif()
|
||||
# Pass USE_TENSORPIPE to torch_cpu as some parts of rpc/utils.cpp
|
||||
# can only be compiled with USE_TENSORPIPE is set.
|
||||
if(USE_TENSORPIPE)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_TENSORPIPE)
|
||||
endif()
|
||||
if(USE_MPI AND USE_C10D_MPI)
|
||||
if(CMAKE_CXX_COMPILER_ID MATCHES "Clang" OR CMAKE_CXX_COMPILER_ID STREQUAL "GNU")
|
||||
set_source_files_properties(
|
||||
"${TORCH_SRC_DIR}/csrc/distributed/c10d/ProcessGroupMPI.cpp"
|
||||
PROPERTIES COMPILE_FLAGS -Wno-deprecated-declarations)
|
||||
endif()
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_C10D_MPI)
|
||||
endif()
|
||||
# Pass USE_RPC in order to reduce use of
|
||||
# #if defined(USE_DISTRIBUTED) && !defined(_WIN32)
|
||||
# need to be removed when RPC is supported
|
||||
if(NOT WIN32)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_RPC)
|
||||
endif()
|
||||
# Pass USE_TENSORPIPE to torch_cpu as some parts of rpc/utils.cpp
|
||||
# can only be compiled with USE_TENSORPIPE is set.
|
||||
if(USE_TENSORPIPE)
|
||||
target_compile_definitions(torch_cpu PUBLIC USE_TENSORPIPE)
|
||||
endif()
|
||||
|
||||
if(NOT INTERN_BUILD_MOBILE)
|
||||
|
||||
Reference in New Issue
Block a user