Merge branch 'develop' into F403/fix

.gitignore

@@ -72,7 +72,9 @@ tools/nvcc_lazy
 
 # these files (directories) are generated before build system generation
 paddle/fluid/operators/generated_op.cc
+paddle/fluid/operators/generated_sparse_op.cc
 paddle/phi/ops/compat/generated_sig.cc
+paddle/phi/ops/compat/generated_sparse_sig.cc
 paddle/phi/api/yaml/parsed_apis/
 python/paddle/utils/code_gen/
 paddle/fluid/pybind/tmp_eager_op_function_impl.h

AUTHORS.md

@@ -83,6 +83,7 @@ This is an incomplete list of authors of [Paddle](https://github.com/PaddlePaddl
 | xushaoyong | Shao-Yong Xu |
 | Yancey1989 | Xu Yan |
 | zhaopu7 | Pu Zhao |
+| zhiqiu | Qiu-Liang Chen |
 | zhouxiao-coder | Xiao Zhou |
 | Zrachel | Rui-Qing Zhang |
 | jeng1220 | Bai-Cheng(Ryan) Jeng (NVIDIA) |

CMakeLists.txt

@@ -47,6 +47,7 @@ find_package(CUDA QUIET)
 find_package(MKL CONFIG QUIET)
 option(WITH_ONEMKL "Compile PaddlePaddle with oneMKL" OFF)
 option(WITH_GPU "Compile PaddlePaddle with NVIDIA GPU" ${CUDA_FOUND})
+option(WITH_MPI "Compile PaddlePaddle with MPI" OFF)
 option(WITH_TENSORRT "Compile PaddlePaddle with NVIDIA TensorRT" OFF)
 option(WITH_XPU "Compile PaddlePaddle with BAIDU KUNLUN XPU" OFF)
 option(WITH_XPU_KP "Compile PaddlePaddle with BAIDU XPU compiler " OFF)
@@ -336,7 +337,7 @@ endif()
 
 if(LINUX
    AND NOT WITH_CUSTOM_DEVICE
-   AND NOT ON_INFER)
+   AND WITH_PYTHON)
   set(WITH_CUSTOM_DEVICE
       ON
       CACHE BOOL "Enable Custom Device when compiling for Linux" FORCE)
@@ -485,9 +486,6 @@ if(WITH_DISTRIBUTE)
         ON
         CACHE STRING "Enable GLOO when compiling WITH_DISTRIBUTE=ON." FORCE)
   endif()
-  set(WITH_MPI
-      ON
-      CACHE STRING "Enable MPI when compiling WITH_DISTRIBUTE=ON." FORCE)
   if(WITH_ASCEND_CL AND NOT WITH_ARM_BRPC)
     # disable WITH_PSCORE for NPU before include third_party
     message(

cmake/external/xpu.cmake

@@ -10,7 +10,7 @@ set(XPU_RT_LIB_NAME "libxpurt.so")
 if(NOT DEFINED XPU_BASE_URL)
   set(XPU_BASE_URL_WITHOUT_DATE
       "https://baidu-kunlun-product.su.bcebos.com/KL-SDK/klsdk-dev")
-  set(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20220928")
+  set(XPU_BASE_URL "${XPU_BASE_URL_WITHOUT_DATE}/20221016")
 else()
   set(XPU_BASE_URL "${XPU_BASE_URL}")
 endif()
@@ -19,7 +19,7 @@ endif()
 if(NOT DEFINED XPU_XDNN_BASE_URL)
   set(XPU_XDNN_BASE_URL_WITHOUT_DATE
       "https://klx-sdk-release-public.su.bcebos.com/xdnn/dev")
-  set(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL_WITHOUT_DATE}/20220928")
+  set(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL_WITHOUT_DATE}/20221016")
 else()
   set(XPU_XDNN_BASE_URL "${XPU_XDNN_BASE_URL}")
 endif()

cmake/operators.cmake

@@ -511,13 +511,6 @@ function(op_library TARGET)
     # Append first implemented MKLDNN activation operator
     if(${MKLDNN_FILE} STREQUAL "activation_mkldnn_op")
       file(APPEND ${pybind_file} "USE_OP_DEVICE_KERNEL(softplus, MKLDNN);\n")
-    elseif(${MKLDNN_FILE} STREQUAL "conv_mkldnn_op")
-      file(APPEND ${pybind_file}
-           "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, FP32);\n")
-      file(APPEND ${pybind_file}
-           "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, S8);\n")
-      file(APPEND ${pybind_file}
-           "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(conv2d, MKLDNN, U8);\n")
     elseif(${MKLDNN_FILE} STREQUAL "fc_mkldnn_op")
       file(APPEND ${pybind_file}
            "USE_OP_DEVICE_KERNEL_WITH_CUSTOM_TYPE(fc, MKLDNN, FP32);\n")

paddle/fluid/distributed/CMakeLists.txt

@@ -42,7 +42,9 @@ set(DISTRIBUTE_COMPILE_FLAGS
 if(CMAKE_CXX_COMPILER_VERSION VERSION_GREATER 7.0)
   set(DISTRIBUTE_COMPILE_FLAGS "${DISTRIBUTE_COMPILE_FLAGS} -faligned-new")
 endif()
-
+if(LINUX)
+  add_subdirectory(rpc)
+endif()
 add_subdirectory(common)
 add_subdirectory(ps)
 add_subdirectory(test)

paddle/fluid/distributed/collective/reducer.cc

@@ -254,6 +254,10 @@ static void ConcatTensorsWithType(
       ConcatTensorsForAllReduce<DeviceContext, double>()(
           context, dense_tensors_, p_dense_contents);
       break;
+    case phi::DataType::BFLOAT16:
+      ConcatTensorsForAllReduce<DeviceContext, platform::bfloat16>()(
+          context, dense_tensors_, p_dense_contents);
+      break;
     default:
       PADDLE_THROW(platform::errors::Unimplemented(
           "Data type (%s) is not supported when it concats tensors for "
@@ -281,6 +285,10 @@ static void SplitTensorsWithType(const DeviceContext &context,
       SplitTensorsForAllReduce<DeviceContext, double>()(
           context, p_dense_contents, p_dense_tensors);
       break;
+    case phi::DataType::BFLOAT16:
+      SplitTensorsForAllReduce<DeviceContext, platform::bfloat16>()(
+          context, p_dense_contents, p_dense_tensors);
+      break;
     default:
       PADDLE_THROW(platform::errors::Unimplemented(
           "Data type (%s) is not supported when it splits tensors for "