Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into...

Merge branch 'develop' of https://github.com/PaddlePaddle/Paddle into feature/large_scale_kv_save_delta

cmake/external/xpu.cmake

@@ -4,7 +4,7 @@ endif()
 
 INCLUDE(ExternalProject)
 SET(XPU_PROJECT                 "extern_xpu")
-SET(XPU_URL    "https://kunlun1.su.bcebos.com/xpu.tar.gz" CACHE STRING "" FORCE)
+SET(XPU_URL    "https://baidu-kunlun-public.su.bcebos.com/paddle_depence/xpu.tar.gz" CACHE STRING "" FORCE)
 SET(XPU_SOURCE_DIR              "${THIRD_PARTY_PATH}/xpu")
 SET(XPU_DOWNLOAD_DIR            "${XPU_SOURCE_DIR}/src/${XPU_PROJECT}")
 SET(XPU_INSTALL_DIR             "${THIRD_PARTY_PATH}/install/xpu")

cmake/generic.cmake

@@ -446,6 +446,9 @@ function(nv_library TARGET_NAME)
         message(FATAL "Please specify source file or library in nv_library.")
       endif()
     endif(nv_library_SRCS)
+    if (WIN32)
+      set_target_properties(${TARGET_NAME} PROPERTIES VS_USER_PROPS ${WIN_PROPS})
+    endif(WIN32)
   endif()
 endfunction(nv_library)
 
@@ -461,6 +464,9 @@ function(nv_binary TARGET_NAME)
       add_dependencies(${TARGET_NAME} ${nv_binary_DEPS})
       common_link(${TARGET_NAME})
     endif()
+    if (WIN32)
+      set_target_properties(${TARGET_NAME} PROPERTIES VS_USER_PROPS ${WIN_PROPS})
+    endif(WIN32)
   endif()
 endfunction(nv_binary)
 
@@ -482,6 +488,9 @@ function(nv_test TARGET_NAME)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cpu_deterministic=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_init_allocated_mem=true)
     set_property(TEST ${TARGET_NAME} PROPERTY ENVIRONMENT FLAGS_cudnn_deterministic=true)
+    if (WIN32)
+      set_target_properties(${TARGET_NAME} PROPERTIES VS_USER_PROPS ${WIN_PROPS})
+    endif(WIN32)
   endif()
 endfunction(nv_test)
 
@@ -712,6 +721,7 @@ function(proto_library TARGET_NAME)
   set(proto_hdrs)
   paddle_protobuf_generate_cpp(proto_srcs proto_hdrs ${proto_library_SRCS})
   cc_library(${TARGET_NAME} SRCS ${proto_srcs} DEPS ${proto_library_DEPS} protobuf)
+  add_dependencies(extern_xxhash ${TARGET_NAME})
 endfunction()
 
 function(py_proto_compile TARGET_NAME)

cmake/inference_lib.cmake

@@ -19,9 +19,8 @@ set(PADDLE_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_install_dir" CACHE STRING
 set(PADDLE_INFERENCE_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_inference_install_dir" CACHE STRING
   "A path setting paddle inference shared and static libraries")
   
-# TODO(zhaolong)
-# At present, the size of static lib in Windows exceeds the system limit,
-# so the generation of static lib is temporarily turned off.
+# At present, the size of static lib in Windows is very large,
+# so we need to crop the library size.
 if(WIN32)
     #todo: remove the option 
     option(WITH_STATIC_LIB "Compile demo with static/shared library, default use dynamic."   OFF)
@@ -196,7 +195,11 @@ set(PADDLE_INFERENCE_C_INSTALL_DIR "${CMAKE_BINARY_DIR}/paddle_inference_c_insta
 copy_part_of_thrid_party(inference_lib_dist ${PADDLE_INFERENCE_C_INSTALL_DIR})
 
 set(src_dir "${PADDLE_SOURCE_DIR}/paddle/fluid")
-set(paddle_fluid_c_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/capi/libpaddle_fluid_c.*)
+if(WIN32)
+  set(paddle_fluid_c_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/capi/${CMAKE_BUILD_TYPE}/paddle_fluid_c.*)
+else(WIN32)
+  set(paddle_fluid_c_lib ${PADDLE_BINARY_DIR}/paddle/fluid/inference/capi/libpaddle_fluid_c.*)
+endif(WIN32)
 
 copy(inference_lib_dist
       SRCS  ${src_dir}/inference/capi/paddle_c_api.h  ${paddle_fluid_c_lib}

cmake/init.cmake

@@ -26,4 +26,7 @@ if(WITH_GPU)
     set(CMAKE_CUDA_FLAGS_MINSIZEREL "-O1 -DNDEBUG")
 endif()
 
+if(WIN32)
+    set(WIN_PROPS ${CMAKE_SOURCE_DIR}/cmake/paddle_win.props)
+endif()
 

cmake/operators.cmake

@@ -62,9 +62,9 @@ function(op_library TARGET)
             endif()
         endif()
         if(WITH_XPU)
-            string(REPLACE "_op" "_xpu_op" XPU_FILE "${TARGET}")
-            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/xpu/${XPU_FILE}.cc)
-                list(APPEND xpu_cc_srcs xpu/${XPU_FILE}.cc)
+            string(REPLACE "_op" "_op_xpu" XPU_FILE "${TARGET}")
+            if (EXISTS ${CMAKE_CURRENT_SOURCE_DIR}/${XPU_FILE}.cc)
+                list(APPEND xpu_cc_srcs ${XPU_FILE}.cc)
             endif()
         endif()
     else()
@@ -83,7 +83,7 @@ function(op_library TARGET)
                 list(APPEND mkldnn_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cu.cc$")
                 list(APPEND cu_cc_srcs ${src})
-            elseif(WITH_XPU AND ${src} MATCHES ".*_xpu_op.cc$")
+            elseif(WITH_XPU AND ${src} MATCHES ".*_op_xpu.cc$")
                 list(APPEND xpu_cc_srcs ${src})
             elseif(${src} MATCHES ".*\\.cc$")
                 list(APPEND cc_srcs ${src})

cmake/paddle_win.props

+<?xml version="1.0" encoding="utf-8"?>
+<Project xmlns="http://schemas.microsoft.com/developer/msbuild/2003">
+    <ItemDefinitionGroup>
+        <CudaCompile>
+            <!-- Project schema: Host properties -->
+            <UseHostDefines>true</UseHostDefines>
+            <Emulation>false</Emulation>
+            <HostDebugInfo Condition="'$(Configuration)' == 'Debug'">true</HostDebugInfo>
+            <HostDebugInfo Condition="'$(Configuration)' != 'Debug'">false</HostDebugInfo>
+            <FastMath>false</FastMath>
+            <Optimization>InheritFromHost</Optimization>
+            <Runtime>InheritFromHost</Runtime>
+            <RuntimeChecks>InheritFromHost</RuntimeChecks>
+            <TypeInfo>InheritFromHost</TypeInfo>
+            <Warning>InheritFromHost</Warning>
+
+            <BaseCommandLineTemplate>-ccbin "%(VCBinDir)" -x cu [GenerateRelocatableDeviceCode] [Include] [RequiredIncludes] [InterleaveSourceInPTX] [GPUDebugInfo] [GenerateLineInfo] [Keep] [KeepDir] [MaxRegCount] [PtxAsOptionV] [TargetMachinePlatform] [NvccCompilation] [CudaRuntime] [AdditionalOptions]</BaseCommandLineTemplate>
+            <BuildCommandLineTemplate>--use-local-env</BuildCommandLineTemplate>
+            <BuildDynamicCommandLineTemplate>[CodeGeneration]</BuildDynamicCommandLineTemplate>
+            <CleanCommandLineTemplate>-clean</CleanCommandLineTemplate>
+            <!-- <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] [ProgramDataBaseFileName] $(CudaForceSynchronousPdbWrites) /Zi [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate> -->
+            <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] [ProgramDataBaseFileName] $(CudaForceSynchronousPdbWrites) [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate>
+
+            <DriverApiCommandLineTemplate>%(BaseCommandLineTemplate) [CompileOut] "%(FullPath)"</DriverApiCommandLineTemplate>
+            <RuntimeApiCommandLineTemplate>%(BaseCommandLineTemplate) [HostDebugInfo] [Emulation] [FastMath] [Defines] %(HostCommandLineTemplate) [CompileOut] "%(FullPath)"</RuntimeApiCommandLineTemplate>
+
+            <CommandLineTemplate>
+# (Approximate command-line.  Settings inherited from host are not visible below.)
+# (Please see the output window after a build for the full command-line)
+
+# Driver API (NVCC Compilation Type is .cubin, .gpu, or .ptx)
+set CUDAFE_FLAGS=--sdk_dir "$(WindowsSdkDir)"
+"$(CudaToolkitNvccPath)" %(BuildCommandLineTemplate) %(DriverApiCommandLineTemplate)
+
+# Runtime API (NVCC Compilation Type is hybrid object or .c file)
+set CUDAFE_FLAGS=--sdk_dir "$(WindowsSdkDir)"
+"$(CudaToolkitNvccPath)" %(BuildCommandLineTemplate) %(RuntimeApiCommandLineTemplate)
+            </CommandLineTemplate>
+            <ExecutionDescription>Compiling CUDA source file %(Identity)...</ExecutionDescription>
+            <ExclusionDescription>Skipping CUDA source file %(Identity) (excluded from build).</ExclusionDescription>
+
+            <!-- Miscellaneous -->
+            <PropsCacheOutputFile>%(Filename)%(Extension).cache</PropsCacheOutputFile>
+            <PropsCacheOutputPath>$(IntDir)%(PropsCacheOutputFile)</PropsCacheOutputPath>
+
+            <CudaCompileCoreProject>$(MSBuildProjectFullPath)</CudaCompileCoreProject>
+        </CudaCompile>
+
+        <CudaLink>
+            <PerformDeviceLink>true</PerformDeviceLink>
+            <LinkOut>$(IntDir)$(TargetName).device-link.obj</LinkOut>
+
+            <AdditionalLibraryDirectories></AdditionalLibraryDirectories>
+            <UseHostLibraryDirectories>true</UseHostLibraryDirectories>
+            <AdditionalDependencies></AdditionalDependencies>
+            <UseHostLibraryDependencies>true</UseHostLibraryDependencies>
+
+            <GPUDebugInfo>InheritFromProject</GPUDebugInfo>
+            <Optimization>InheritFromProject</Optimization>
+
+            <!-- Implicitly inherited from the project via @(CudaCompile) -->
+            <CodeGeneration></CodeGeneration>
+            <RuntimeChecks></RuntimeChecks>
+            <Runtime></Runtime>
+            <TargetMachinePlatform></TargetMachinePlatform>
+            <TypeInfo></TypeInfo>
+            <Warning></Warning>
+
+            <Inputs></Inputs>
+
+            <!-- <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] /Zi [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate> -->
+            <HostCommandLineTemplate>-Xcompiler &quot;/EHsc [Warning] /nologo [Optimization] [RuntimeChecks] [Runtime] [TypeInfo]&quot;</HostCommandLineTemplate>
+            <LinkCommandLineTemplate>"$(CudaToolkitNvccPath)" -dlink [LinkOut] %(HostCommandLineTemplate) [AdditionalLibraryDirectories] [AdditionalDependencies] [AdditionalOptions] [CodeGeneration] [GPUDebugInfo] [TargetMachinePlatform] [Inputs]</LinkCommandLineTemplate>
+            <CommandLineTemplate>
+# (Approximate command-line.  Settings inherited from host are not visible below.)
+# (Please see the output window after a build for the full command-line)
+
+%(LinkCommandLineTemplate)
+            </CommandLineTemplate>
+        </CudaLink>
+
+        <Link>
+            <AdditionalLibraryDirectories>%(AdditionalLibraryDirectories);$(CudaToolkitLibDir)</AdditionalLibraryDirectories>
+        </Link>
+
+        <ClCompile>
+            <AdditionalIncludeDirectories>%(AdditionalIncludeDirectories);$(CudaToolkitIncludeDir)</AdditionalIncludeDirectories>
+        </ClCompile>
+    </ItemDefinitionGroup>
+</Project>
+

cmake/third_party.cmake

@@ -39,6 +39,7 @@ set(third_party_deps)
 #            REPOSITORY ${TARGET_REPOSITORY}
 #            TAG        ${TARGET_TAG}
 #            DIR        ${TARGET_SOURCE_DIR})
+
 FUNCTION(cache_third_party TARGET)
     SET(options "")
     SET(oneValueArgs URL REPOSITORY TAG DIR)
@@ -269,6 +270,10 @@ if(WITH_PSLIB)
     endif()
 endif(WITH_PSLIB)
 
+if(NOT WIN32 AND NOT APPLE)
+    include(external/gloo)
+    list(APPEND third_party_deps extern_gloo)
+endif()
 
 if(WITH_BOX_PS)
     include(external/box_ps)
@@ -276,10 +281,6 @@ if(WITH_BOX_PS)
 endif(WITH_BOX_PS)
 
 if(WITH_DISTRIBUTE)
-    if(WITH_GLOO)
-        include(external/gloo)
-        list(APPEND third_party_deps extern_gloo)
-    endif()
 
     if(WITH_GRPC)
         list(APPEND third_party_deps extern_grpc)

paddle/fluid/framework/attribute.cc

@@ -14,8 +14,6 @@ limitations under the License. */
 
 #include "paddle/fluid/framework/attribute.h"
 
-#include <vector>
-
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/block_desc.h

@@ -30,6 +30,8 @@ namespace paddle {
 namespace framework {
 
 class ProgramDesc;
+class OpDesc;
+class VarDesc;
 
 // Each Protobuf Message, we provide a XXXBind class. In that class, we optimize
 // read/write speed. Only when we want the protobuf message, the local changes

paddle/fluid/framework/c/c_api.cc

@@ -12,17 +12,7 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <string>
-#include <unordered_map>
-#include <unordered_set>
-#include <vector>
-
-#include "paddle/fluid/framework/block_desc.h"
 #include "paddle/fluid/framework/c/c_api.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/program_desc.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/init.h"
 
 extern "C" {
 

paddle/fluid/framework/c/c_api.h

@@ -24,6 +24,15 @@ limitations under the License. */
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class OpInfoMap;
+}  // namespace framework
+namespace platform {
+class DeviceContextPool;
+}  // namespace platform
+}  // namespace paddle
+
 #ifdef __cplusplus
 extern "C" {
 #endif

paddle/fluid/framework/channel.h

@@ -277,7 +277,7 @@ class ChannelObject {
     size_t finished = 0;
     while (finished < n && WaitForWrite(lock)) {
       size_t m =
-          std::min(n - finished, capacity_ + reading_count_ - data_.size());
+          (std::min)(n - finished, capacity_ + reading_count_ - data_.size());
       for (size_t i = 0; i < m; i++) {
         data_.push_back(std::move(p[finished++]));
       }

paddle/fluid/framework/copy_same_tensor_test.cc

@@ -12,8 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
-#include <cstring>
 #include <random>
+
 #include "gflags/gflags.h"
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/tensor.h"

paddle/fluid/framework/data_device_transform.h

@@ -21,6 +21,8 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class Tensor;
+
 void TransDataDevice(const Tensor& in, const platform::Place& dst_place,
                      Tensor* out);
 

paddle/fluid/framework/data_feed.h

@@ -41,6 +41,15 @@ limitations under the License. */
 #include "paddle/fluid/framework/variable.h"
 #include "paddle/fluid/string/string_helper.h"
 
+namespace paddle {
+namespace framework {
+class DataFeedDesc;
+class LoDTensor;
+class Scope;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
@@ -418,6 +427,7 @@ class MultiSlotType {
 
   std::string DebugString() {
     std::stringstream ss;
+
     ss << "\ntype: " << type_ << "\n";
     ss << "offset: ";
     ss << "[";

paddle/fluid/framework/data_feed_factory.cc

@@ -17,10 +17,10 @@ limitations under the License. */
 #include <string>
 #include <unordered_map>
 
-#include "paddle/fluid/framework/data_feed.h"
-
 namespace paddle {
 namespace framework {
+class DataFeed;
+
 typedef std::shared_ptr<DataFeed> (*Createdata_feedFunction)();
 typedef std::unordered_map<std::string, Createdata_feedFunction> data_feedMap;
 data_feedMap g_data_feed_map;

paddle/fluid/framework/data_feed_factory.h

@@ -16,10 +16,13 @@ limitations under the License. */
 
 #include <memory>
 #include <string>
+
 #include "paddle/fluid/framework/data_feed.h"
 
 namespace paddle {
 namespace framework {
+class DataFeed;
+
 class DataFeedFactory {
  public:
   static std::string DataFeedTypeList();

paddle/fluid/framework/data_layout_transform.cc

@@ -13,8 +13,8 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/data_layout_transform.h"
+
 #include <string>
-#include <vector>
 
 #include "paddle/fluid/operators/math/math_function.h"
 #ifdef PADDLE_WITH_MKLDNN

paddle/fluid/framework/data_layout_transform.h

@@ -17,10 +17,18 @@
 #include <map>
 #include <unordered_map>
 #include <vector>
+
 #include "paddle/fluid/framework/op_kernel_type.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/variable.h"
 
+namespace paddle {
+namespace framework {
+class OpKernelType;
+class Tensor;
+}  // namespace framework
+}  // namespace paddle
+
 #ifdef PADDLE_WITH_MKLDNN
 #include "paddle/fluid/platform/mkldnn_helper.h"
 #endif

paddle/fluid/framework/data_layout_transform_test.cc

@@ -15,7 +15,6 @@
 #include "paddle/fluid/framework/data_layout_transform.h"
 
 #include "gtest/gtest.h"
-#include "paddle/fluid/platform/device_context.h"
 
 TEST(DataTransform, DataLayoutFunction) {
   auto place = paddle::platform::CPUPlace();

paddle/fluid/framework/data_transform.cc

@@ -18,8 +18,13 @@ limitations under the License. */
 #include "paddle/fluid/framework/data_layout_transform.h"
 #include "paddle/fluid/framework/data_type_transform.h"
 
+namespace paddle {
+namespace framework {
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 #ifdef PADDLE_WITH_MKLDNN
-#include <algorithm>
 #include "paddle/fluid/platform/mkldnn_helper.h"
 #endif
 

paddle/fluid/framework/data_transform.h

@@ -30,6 +30,10 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class OpKernelType;
+class Tensor;
+class Variable;
+
 void TransformData(const OpKernelType &expected_kernel_type,
                    const OpKernelType &kernel_type_for_var,
                    const Tensor &input_tensor, Tensor *out);

paddle/fluid/framework/data_type.cc

@@ -13,7 +13,6 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/data_type.h"
-#include <stdint.h>
 #include <string>
 #include <unordered_map>
 

paddle/fluid/framework/data_type.h

@@ -15,12 +15,19 @@ limitations under the License. */
 #pragma once
 #include <string>
 #include <typeindex>
-#include "paddle/fluid/framework/framework.pb.h"
-#include "paddle/fluid/platform/enforce.h"
 
+#include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/platform/bfloat16.h"
+#include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/float16.h"
 
+namespace paddle {
+namespace platform {
+struct bfloat16;
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/data_type_test.cc

@@ -14,9 +14,17 @@
 #include "paddle/fluid/framework/data_type.h"
 
 #include <string>
+
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/tensor.h"
 
+namespace paddle {
+namespace platform {
+struct bfloat16;
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 TEST(DataType, float16) {
   using paddle::framework::Tensor;
   using paddle::platform::CPUPlace;

paddle/fluid/framework/data_type_transform.h

@@ -15,6 +15,7 @@ limitations under the License. */
 #pragma once
 
 #include <utility>
+
 #include "paddle/fluid/framework/op_kernel_type.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/framework/variable.h"
@@ -23,6 +24,9 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class OpKernelType;
+class Tensor;
+
 using KernelTypePair = std::pair<OpKernelType, OpKernelType>;
 
 void TransDataType(const OpKernelType& kernel_type_for_var,

paddle/fluid/framework/dataset_factory.cc

@@ -13,12 +13,9 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/dataset_factory.h"
-#include <memory>
 #include <string>
 #include <unordered_map>
 
-#include "paddle/fluid/framework/data_set.h"
-
 namespace paddle {
 namespace framework {
 typedef std::unique_ptr<Dataset> (*CreateDatasetFunction)();

paddle/fluid/framework/dataset_factory.h

@@ -16,6 +16,7 @@ limitations under the License. */
 
 #include <memory>
 #include <string>
+
 #include "paddle/fluid/framework/data_set.h"
 
 namespace paddle {

paddle/fluid/framework/ddim.h

@@ -18,6 +18,7 @@ limitations under the License. */
 #include <stdexcept>
 #include <string>
 #include <vector>
+
 #include "paddle/fluid/framework/dim.h"
 
 namespace paddle {

paddle/fluid/framework/ddim_test.cc

@@ -12,7 +12,6 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 #include <sstream>
-#include <vector>
 
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/ddim.h"

paddle/fluid/framework/details/all_reduce_op_handle.cc

@@ -76,7 +76,7 @@ void AllReduceOpHandle::AllReduceImpl(
                     platform::errors::InvalidArgument(
                         "The NoDummyInputSize should be equal "
                         "to the number of places, but got NoDummyInputSize is "
-                        "%d and the number of place is %d.",
+                        "%d and the number of places is %d.",
                         in_var_handles.size(), num_places));
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), out_var_handles.size(),
@@ -89,7 +89,7 @@ void AllReduceOpHandle::AllReduceImpl(
       platform::errors::InvalidArgument(
           "The number of local scopes should be equal "
           "to the number of places, but got the number of local scopes is "
-          "%d and the number of place is %d.",
+          "%d and the number of places is %d.",
           in_var_handles.size(), num_places));
 
   std::vector<const void *> lod_tensor_data;

paddle/fluid/framework/details/all_reduce_op_handle.h

@@ -20,6 +20,17 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/framework/details/nccl_op_handle.h"
 #include "paddle/fluid/platform/nccl_helper.h"

paddle/fluid/framework/details/broadcast_op_handle.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
+
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -31,10 +32,15 @@ void BroadcastOpHandle::RunImpl() {
   auto out_var_handles = DynamicCast<VarHandle>(outputs_);
 
   PADDLE_ENFORCE_EQ(in_var_handles.size(), 1UL,
-                    "The number of input should be one.");
-  PADDLE_ENFORCE_EQ(
-      out_var_handles.size(), places_.size(),
-      "The number of output should equal to the number of places.");
+                    platform::errors::PreconditionNotMet(
+                        "The number of inputs should be 1, but got %d.",
+                        in_var_handles.size()));
+  PADDLE_ENFORCE_EQ(out_var_handles.size(), places_.size(),
+                    platform::errors::PreconditionNotMet(
+                        "The number of outputs and the number of places should "
+                        "be equal, but got the number of outputs is %d and the "
+                        "number of places is %d.",
+                        out_var_handles.size(), places_.size()));
 
   VarHandle *in_var_handle = in_var_handles[0];
 
@@ -47,7 +53,9 @@ void BroadcastOpHandle::BroadcastOneVar(
     const std::vector<Scope *> &var_scopes) {
   auto *in_var =
       var_scopes.at(in_var_handle.scope_idx())->FindVar(in_var_handle.name());
-  PADDLE_ENFORCE_NOT_NULL(in_var);
+  PADDLE_ENFORCE_NOT_NULL(
+      in_var, platform::errors::NotFound("Variable %s is not found in scopes.",
+                                         in_var_handle.name()));
   Tensor &in_tensor = VariableVisitor::GetMutableTensor(in_var);
   if (UNLIKELY(!in_tensor.IsInitialized())) {
     VLOG(3) << "in var " << in_var_handle.name() << "not inited, return!";
@@ -103,7 +111,7 @@ void BroadcastOpHandle::BroadcastOneVar(
 
       broadcast_calls.emplace_back(
           [send_recv_buffer, numel, type, root_id, &nccl_ctx] {
-            PADDLE_ENFORCE(platform::dynload::ncclBcast(
+            PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclBcast(
                 send_recv_buffer, numel, static_cast<ncclDataType_t>(type),
                 root_id, nccl_ctx.comm_, nccl_ctx.stream()));
           });
@@ -131,7 +139,8 @@ void BroadcastOpHandle::BroadcastOneVar(
       nccl_ctxs_->DevCtx(p)->Wait();
     }
 #else
-    PADDLE_THROW("CUDA is not enabled.");
+    PADDLE_THROW(
+        platform::errors::PreconditionNotMet("Not compiled with NCLL."));
 #endif
   }
 }
@@ -154,10 +163,13 @@ void BroadcastOpHandle::InitOutputValue(
     auto t_out_p = out_var_handle->place();
     auto *out_var = var_scopes.at(out_var_handle->scope_idx())
                         ->FindVar(out_var_handle->name());
-    PADDLE_ENFORCE_NOT_NULL(out_var);
+    PADDLE_ENFORCE_NOT_NULL(out_var, platform::errors::NotFound(
+                                         "Variable %s is not found in scopes.",
+                                         out_var_handle->name()));
     if (is_gpu_place(in_tensor.place())) {
-      PADDLE_ENFORCE(platform::is_gpu_place(t_out_p),
-                     "Places of input and output must be all on GPU.");
+      PADDLE_ENFORCE_EQ(platform::is_gpu_place(t_out_p), true,
+                        platform::errors::PreconditionNotMet(
+                            "Places of input and output must be all on GPU."));
     } else {
       t_out_p = platform::CPUPlace();
     }

paddle/fluid/framework/details/broadcast_op_handle.h

@@ -24,6 +24,20 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace details {
+struct VarHandle;
+}  // namespace details
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+struct NCCLContextMap;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/broadcast_op_handle_test.h

@@ -21,13 +21,15 @@
 
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/details/broadcast_op_handle.h"
-
 #include "paddle/fluid/platform/device_context.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
 
+struct DummyVarHandle;
+struct VarHandle;
+
 namespace f = paddle::framework;
 namespace p = paddle::platform;
 
@@ -77,7 +79,8 @@ struct TestBroadcastOpHandle {
       }
       nccl_ctxs_.reset(new platform::NCCLContextMap(place_list_));
 #else
-      PADDLE_THROW("CUDA is not support.");
+      PADDLE_THROW(
+          platform::errors::PreconditionNotMet("Not compiled with NCLL."));
 #endif
     } else {
       int count = 8;
@@ -111,7 +114,8 @@ struct TestBroadcastOpHandle {
       op_handle_ = new BroadcastOpHandle(nodes_.back().get(), local_scopes_,
                                          place_list_, nccl_ctxs_.get());
 #else
-      PADDLE_THROW("CUDA is not support.");
+      PADDLE_THROW(
+          platform::errors::PreconditionNotMet("Not compiled with NCLL."));
 #endif
     } else {
 #if defined(PADDLE_WITH_NCCL)
@@ -169,7 +173,9 @@ struct TestBroadcastOpHandle {
                                    float val_scalar = 0.0) {
     auto var = param_scopes_[input_scope_idx]->FindVar(varname);
 
-    PADDLE_ENFORCE_NOT_NULL(var);
+    PADDLE_ENFORCE_NOT_NULL(
+        var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                        varname));
     auto lod_tensor = var->GetMutable<f::LoDTensor>();
     std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
     for (size_t k = 0; k < send_vector.size(); ++k) {
@@ -192,7 +198,9 @@ struct TestBroadcastOpHandle {
     }
 
     auto var = param_scopes_[input_scope_idx]->FindVar(varname);
-    PADDLE_ENFORCE_NOT_NULL(var);
+    PADDLE_ENFORCE_NOT_NULL(
+        var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                        varname));
     auto selected_rows = var->GetMutable<f::SelectedRows>();
     auto value = selected_rows->mutable_value();
     value->mutable_data<float>(kDims, place_list_[input_scope_idx]);
@@ -209,13 +217,24 @@ struct TestBroadcastOpHandle {
                          const std::vector<float>& send_vector,
                          const std::vector<int64_t>& rows, int height) {
     auto var = param_scopes_[input_scope_idx]->FindVar(varname);
-    PADDLE_ENFORCE_NOT_NULL(var);
+    PADDLE_ENFORCE_NOT_NULL(
+        var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                        varname));
     auto& selected_rows = var->Get<f::SelectedRows>();
     auto rt = selected_rows.value();
-    PADDLE_ENFORCE_EQ(selected_rows.height(), height, "height is not equal.");
+    PADDLE_ENFORCE_EQ(selected_rows.height(), height,
+                      platform::errors::InvalidArgument(
+                          "The height of SelectedRows is not equal to "
+                          "the expected, expect %d, but got %ld.",
+                          height, selected_rows.height()));
 
     for (size_t k = 0; k < selected_rows.rows().size(); ++k) {
-      PADDLE_ENFORCE_EQ(selected_rows.rows()[k], rows[k]);
+      PADDLE_ENFORCE_EQ(
+          selected_rows.rows()[k], rows[k],
+          platform::errors::InvalidArgument(
+              "The item at position %zu of rows of SelectedRows "
+              "is not equal to the expected, expect %ld, but got %ld.",
+              k, rows[k], selected_rows.rows()[k]));
     }
 
     p::CPUPlace cpu_place;
@@ -233,9 +252,15 @@ struct TestBroadcastOpHandle {
                       framework::Scope* scope) {
     p::CPUPlace cpu_place;
     auto var = scope->FindVar(varname);
-    PADDLE_ENFORCE_NOT_NULL(var);
+    PADDLE_ENFORCE_NOT_NULL(
+        var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                        varname));
     auto tensor = var->Get<f::LoDTensor>();
-    PADDLE_ENFORCE_EQ(tensor.lod(), lod, "lod is not equal.");
+    PADDLE_ENFORCE_EQ(tensor.lod(), lod,
+                      platform::errors::InvalidArgument(
+                          "The LoD of tensor is not equal to "
+                          "the expected, expect %s, but got %s.",
+                          lod, tensor.lod()));
     f::Tensor result_tensor;
     f::TensorCopySync(tensor, cpu_place, &result_tensor);
     float* ct = result_tensor.mutable_data<float>(cpu_place);

paddle/fluid/framework/details/build_strategy.cc

@@ -235,7 +235,8 @@ class ParallelExecutorPassBuilder : public ir::PassBuilder {
               AppendPass("reduce_mode_multi_devices_pass").get();
           break;
         default:
-          PADDLE_THROW("Unknown reduce strategy.");
+          PADDLE_THROW(
+              platform::errors::Unimplemented("Unknown reduce strategy."));
       }
     }
     multi_devices_pass->SetNotOwned<const BuildStrategy>("strategy",

paddle/fluid/framework/details/build_strategy.h

@@ -27,6 +27,18 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/enforce.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Graph;
+class PassBuilder;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/computation_op_handle.cc

@@ -19,6 +19,8 @@
 namespace paddle {
 namespace framework {
 namespace details {
+struct VarHandleBase;
+
 ComputationOpHandle::ComputationOpHandle(ir::Node *node, Scope *scope,
                                          platform::Place place,
                                          size_t scope_idx)

paddle/fluid/framework/details/computation_op_handle.h

@@ -24,9 +24,21 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class OperatorBase;
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
+struct VarHandleBase;
+
 class ComputationOpHandle : public OpHandleBase {
  public:
   ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,

paddle/fluid/framework/details/eager_deletion_op_handle.cc

@@ -12,11 +12,12 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
+
 #include <memory>
 #include <unordered_set>
 #include <utility>
 
-#include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
 #include "paddle/fluid/framework/ir/memory_optimize_pass/memory_optimization_var_info.h"
 #include "paddle/fluid/framework/lod_tensor_array.h"
 #include "paddle/fluid/framework/scope.h"
@@ -47,15 +48,19 @@ EagerDeletionOpHandle::EagerDeletionOpHandle(
     if (dynamic_cast<StreamGarbageCollector *>(gc_)) {
       platform::CUDADeviceGuard guard(
           BOOST_GET_CONST(platform::CUDAPlace, place).device);
-      PADDLE_ENFORCE(cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
-      PADDLE_ENFORCE_NOT_NULL(event_);
+      PADDLE_ENFORCE_CUDA_SUCCESS(
+          cudaEventCreateWithFlags(&event_, cudaEventDisableTiming));
+      PADDLE_ENFORCE_NOT_NULL(event_, platform::errors::InvalidArgument(
+                                          "The cuda envet created is NULL."));
     }
   }
 #endif
-  PADDLE_ENFORCE_NE(vars.empty(), true, platform::errors::InvalidArgument(
-                                            "Variable names are empty."));
+  PADDLE_ENFORCE_NE(vars.empty(), true,
+                    platform::errors::InvalidArgument(
+                        "The variables to be deleted are empty."));
   for (auto *var : var_infos_) {
-    PADDLE_ENFORCE_NOT_NULL(var);
+    PADDLE_ENFORCE_NOT_NULL(var, platform::errors::InvalidArgument(
+                                     "The memory optimization info is NULL."));
   }
 }
 
@@ -64,7 +69,7 @@ EagerDeletionOpHandle::~EagerDeletionOpHandle() {
   if (event_) {
     auto gpu_place = BOOST_GET_CONST(platform::CUDAPlace, dev_ctx_->GetPlace());
     platform::CUDADeviceGuard guard(gpu_place.device);
-    PADDLE_ENFORCE(cudaEventDestroy(event_));
+    PADDLE_ENFORCE_CUDA_SUCCESS(cudaEventDestroy(event_));
   }
 #endif
 }
@@ -78,12 +83,17 @@ void EagerDeletionOpHandle::InitCUDA() {
 }
 
 void EagerDeletionOpHandle::CallOnce() {
-  PADDLE_ENFORCE(vars_.empty(), "vars_ must be initialized here");
+  PADDLE_ENFORCE_EQ(
+      vars_.empty(), true,
+      platform::errors::InvalidArgument(
+          "The variables to be deleted should be initialized here."));
   Scope *exec_scope = local_exec_scopes_[0];
   for (auto *var_info : var_infos_) {
     auto *var = exec_scope->FindVar(var_info->Name());
-    PADDLE_ENFORCE_NOT_NULL(var, "Variable %s should not be nullptr",
-                            var_info->Name());
+    PADDLE_ENFORCE_NOT_NULL(
+        var, platform::errors::NotFound(
+                 "The variable(%s) to be inplaced is not found in scope.",
+                 var_info->Name()));
     vars_.emplace_back(var);
   }
 }
@@ -119,8 +129,9 @@ void EagerDeletionOpHandle::RunImpl() {
         garbages.emplace_back(t.MoveMemoryHolder());
       }
     } else {
-      PADDLE_THROW("Type %s of %s is not supported eager deletion",
-                   framework::ToTypeName(var->Type()), var_info->Name());
+      PADDLE_THROW(platform::errors::Unimplemented(
+          "The variable(%s) of type %s is not supported in eager deletion.",
+          framework::ToTypeName(var->Type()), var_info->Name()));
     }
   }
 
@@ -137,8 +148,9 @@ void EagerDeletionOpHandle::ClearGarbages(
     auto callback_stream =
         reinterpret_cast<StreamGarbageCollector *>(gc_)->stream();
     auto callback_func = [=]() {
-      PADDLE_ENFORCE(cudaEventRecord(event_, compute_stream));
-      PADDLE_ENFORCE(cudaStreamWaitEvent(callback_stream, event_, 0));
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaEventRecord(event_, compute_stream));
+      PADDLE_ENFORCE_CUDA_SUCCESS(
+          cudaStreamWaitEvent(callback_stream, event_, 0));
     };
     gc_->Add(std::move(*garbages), callback_func);
   } else {

paddle/fluid/framework/details/eager_deletion_op_handle.h

@@ -19,12 +19,23 @@
 #include <string>
 #include <unordered_set>
 #include <vector>
+
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/ir/memory_optimize_pass/reference_count_pass_helper.h"
 
+namespace paddle {
+namespace platform {
+class CUDADeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 class Scope;
+class GarbageCollector;
+namespace ir {
+class Node;
+}  // namespace ir
 
 namespace ir {
 class MemOptVarInfo;

paddle/fluid/framework/details/exception_holder_test.cc

@@ -13,8 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/exception_holder.h"
-#include <memory>
-#include <unordered_map>
+
 #include "gtest/gtest.h"
 #include "paddle/fluid/memory/allocation/allocator.h"
 

paddle/fluid/framework/details/fetch_async_op_handle.cc

@@ -15,9 +15,14 @@
 #include "paddle/fluid/framework/details/fetch_async_op_handle.h"
 #include <string>
 #include <utility>
-#include <vector>
 #include "paddle/fluid/platform/profiler.h"
 
+namespace paddle {
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/fetch_async_op_handle.h

@@ -22,6 +22,18 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class LoDTensor;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/fetch_barrier_op_handle.cc

@@ -19,6 +19,8 @@
 namespace paddle {
 namespace framework {
 namespace details {
+struct VarHandleBase;
+
 FetchBarrierOpHandle::FetchBarrierOpHandle(
     ir::Node *node, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places)

paddle/fluid/framework/details/fetch_barrier_op_handle.h

@@ -24,6 +24,15 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
@@ -32,6 +41,8 @@ namespace details {
 // all places if there are multiple places, must init with
 // multiple dev_ctxes_ !!!!
 
+struct VarHandleBase;
+
 struct FetchBarrierOpHandle : public OpHandleBase {
  public:
   FetchBarrierOpHandle(ir::Node *node, const std::vector<Scope *> &local_scopes,

paddle/fluid/framework/details/fetch_op_handle.h

@@ -22,6 +22,17 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/fused_all_reduce_op_handle.cc

@@ -12,8 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 #include "paddle/fluid/framework/details/fused_all_reduce_op_handle.h"
+
 #include <algorithm>
 #include <utility>
+
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
@@ -56,10 +58,20 @@ void FusedAllReduceOpHandle::RunImpl() {
   size_t place_num = places_.size();
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), place_num * num_of_all_reduce_,
-      "The NoDummyInputSize should be equal to the number of places.");
+      platform::errors::PreconditionNotMet(
+          "The number of input variable handles should be equal to the number "
+          "of places plus the number of all reduce handles, "
+          "but got the number of input variable handles is %d, the "
+          "number of places is %d, and the number of all reduce handles "
+          "is %d.",
+          in_var_handles.size(), place_num, num_of_all_reduce_));
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), out_var_handles.size(),
-      "The NoDummyInputSize and NoDummyOutputSize should be equal.");
+      platform::errors::PreconditionNotMet(
+          "The number of input variable handles should be equal to the number "
+          "of output variable handles, but got the number of input variable "
+          "handles is %d, and the number of  output variable handles is %d.",
+          in_var_handles.size(), out_var_handles.size()));
 
   // Note: some gradient op doesn't have CUDAKernel, so the gradients of
   // those op are in CPUPlace, in this case, the all reduce should not be fused.
@@ -106,7 +118,13 @@ void FusedAllReduceOpHandle::FusedAllReduceFunc(
       dtype = ele_dtype;
     }
 
-    PADDLE_ENFORCE_EQ(ele_dtype, dtype);
+    PADDLE_ENFORCE_EQ(
+        ele_dtype, dtype,
+        platform::errors::InvalidArgument(
+            "The DataType of grad tensors of fused_all_reduce_op_handle  "
+            "must be consistent. The current dtype is %s, but the "
+            "previous dtype is %s.",
+            DataTypeToString(ele_dtype), DataTypeToString(dtype)));
 
     // Check whether the address space is contiguous.
     std::sort(
@@ -130,16 +148,29 @@ void FusedAllReduceOpHandle::FusedAllReduceFunc(
           "input[%d] address: 0X%02x. The offset: %d",
           k - 1, g_tensor.at(k - 1).first, cur_address, g_tensor.at(k).first, k,
           next_address, k, infer_next_address, offset);
-      PADDLE_ENFORCE_EQ(infer_next_address, next_address,
-                        "The address is not consistent.");
+      PADDLE_ENFORCE_EQ(
+          infer_next_address, next_address,
+          platform::errors::InvalidArgument(
+              "The infered address of the next tensor should be equal to the "
+              "real address of the next tensor. But got infered address is %p "
+              "and real address is %p.",
+              infer_next_address, next_address));
     }
   }
 
   if (!FLAGS_skip_fused_all_reduce_check) {
     for (size_t scope_idx = 0; scope_idx < place_num; ++scope_idx) {
       for (size_t j = 1; j < num_of_all_reduce_; ++j) {
-        PADDLE_ENFORCE_EQ(grads_tensor.at(0).at(j).first,
-                          grads_tensor.at(scope_idx).at(j).first);
+        PADDLE_ENFORCE_EQ(
+            grads_tensor.at(0).at(j).first,
+            grads_tensor.at(scope_idx).at(j).first,
+            platform::errors::InvalidArgument(
+                "The variable name of grad tensors of "
+                "fused_all_reduce_op_handle  "
+                "must be consistent. The current name is %s, but the "
+                "previous name is %s.",
+                grads_tensor.at(0).at(j).first,
+                grads_tensor.at(scope_idx).at(j).first));
       }
     }
   }
@@ -167,7 +198,9 @@ bool FusedAllReduceOpHandle::InputIsInDifferentPlace(
     for (size_t j = 0; j < in_var_handles.size(); j += place_num) {
       auto var_name = in_var_handles[j]->name();
       auto var = local_scope->FindVar(var_name);
-      PADDLE_ENFORCE_NOT_NULL(var, "%s is not found in local scope.", var_name);
+      PADDLE_ENFORCE_NOT_NULL(
+          var, platform::errors::NotFound(
+                   "The variable '%s' is not found in local scope.", var_name));
       auto &lod_tensor = var->Get<LoDTensor>();
       if (!is_same_place(lod_tensor.place(), places_.at(scope_idx))) {
         return true;
@@ -185,14 +218,24 @@ void FusedAllReduceOpHandle::GetGradLoDTensor(
   size_t place_num = places_.size();
   for (size_t j = 0; j < in_var_handles.size(); j += place_num) {
     auto var_name = in_var_handles[j]->name();
-    PADDLE_ENFORCE_EQ(var_name, out_var_handles[j]->name());
+    PADDLE_ENFORCE_EQ(
+        var_name, out_var_handles[j]->name(),
+        platform::errors::InvalidArgument(
+            "The name of input variable should be equal "
+            "to the name of output variable. But got the name of input "
+            "variable is %s and the name of output variable is %s.",
+            var_name, out_var_handles[j]->name()));
     auto var = local_scope->FindVar(var_name);
-    PADDLE_ENFORCE_NOT_NULL(var, "%s is not found in local scope.", var_name);
+    PADDLE_ENFORCE_NOT_NULL(
+        var, platform::errors::NotFound(
+                 "The variable '%s' is not found in local scope.", var_name));
     auto &lod_tensor = var->Get<LoDTensor>();
 
     PADDLE_ENFORCE_EQ(
         platform::is_same_place(lod_tensor.place(), places_.at(scope_idx)),
-        true, "%s(%d) is not in the right place.", var_name, scope_idx);
+        true, platform::errors::InvalidArgument(
+                  "The variable '%s' at scope %d is not in the right place.",
+                  var_name, scope_idx));
     grad_tensor->emplace_back(std::make_pair(var_name, &lod_tensor));
   }
 }
@@ -204,16 +247,26 @@ void FusedAllReduceOpHandle::GetDTypeAndNumel(
   size_t size_of_dtype = 0;
   for (size_t i = 0; i < grad_tensor.size(); ++i) {
     // Get dtype
-    auto ele_type = grad_tensor.at(i).second->type();
+    auto ele_dtype = grad_tensor.at(i).second->type();
     if (i == 0) {
-      *dtype = ele_type;
-      size_of_dtype = framework::SizeOfType(ele_type);
+      *dtype = ele_dtype;
+      size_of_dtype = framework::SizeOfType(ele_dtype);
     }
-    PADDLE_ENFORCE_EQ(ele_type, *dtype);
+    PADDLE_ENFORCE_EQ(
+        ele_dtype, *dtype,
+        platform::errors::InvalidArgument(
+            "The DataType of grad tensors of fused_all_reduce_op_handle  "
+            "must be consistent. The current dtype is %s, but the "
+            "previous dtype is %s.",
+            DataTypeToString(ele_dtype), DataTypeToString(*dtype)));
 
     // Get element number
     int64_t len = grad_tensor.at(i).second->numel();
-    PADDLE_ENFORCE_GT(len, 0);
+    PADDLE_ENFORCE_GT(
+        len, 0, platform::errors::InvalidArgument(
+                    "The size of grad tensors of fused_all_reduce_op_handle  "
+                    "must be > 0, but got %d.",
+                    len));
     *numel +=
         platform::Alignment(len * size_of_dtype, places_[0]) / size_of_dtype;
   }

paddle/fluid/framework/details/fused_all_reduce_op_handle.h

@@ -17,10 +17,22 @@
 #include <string>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/framework/details/all_reduce_op_handle.h"
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
+
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/framework/details/nccl_op_handle.h"
 #include "paddle/fluid/platform/nccl_helper.h"

paddle/fluid/framework/details/fused_broadcast_op_handle.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
+
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -32,7 +33,15 @@ void FusedBroadcastOpHandle::RunImpl() {
   WaitInputVarGenerated();
 
   size_t place_num = places_.size();
-  PADDLE_ENFORCE_EQ(in_var_handles.size() * place_num, out_var_handles.size());
+  PADDLE_ENFORCE_EQ(
+      in_var_handles.size() * place_num, out_var_handles.size(),
+      platform::errors::PreconditionNotMet(
+          "The number of input variable handles plus the number "
+          "of places should be equal to the number of output variable handles, "
+          "but got the number of input variable handles is %d, the "
+          "number of places is %d, and the number of output variable handles "
+          "is %d.",
+          in_var_handles.size(), place_num, out_var_handles.size()));
 
   for (size_t i = 0; i < in_var_handles.size(); ++i) {
     BroadcastOneVar(

paddle/fluid/framework/details/fused_broadcast_op_handle.h

@@ -25,6 +25,17 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+struct NCCLContextMap;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/fused_broadcast_op_handle_test.cc

@@ -13,15 +13,26 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/fused_broadcast_op_handle.h"
+
 #include <memory>
 #include <unordered_map>
+
 #include "gtest/gtest.h"
 #include "paddle/fluid/framework/details/broadcast_op_handle_test.h"
+#include "paddle/fluid/framework/details/op_handle_base.h"
+
+namespace paddle {
+namespace framework {
+class Scope;
+}  // namespace framework
+}  // namespace paddle
 
 namespace paddle {
 namespace framework {
 namespace details {
 
+struct VarHandle;
+
 struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
   std::vector<std::string> out_varnames_;
   std::vector<std::unique_ptr<ir::Node>> nodes_;
@@ -49,7 +60,8 @@ struct TestFusedBroadcastOpHandle : TestBroadcastOpHandle {
       op_handle_ = new FusedBroadcastOpHandle(
           nodes_.back().get(), local_scopes_, place_list_, nccl_ctxs_.get());
 #else
-      PADDLE_THROW("CUDA is not supported.");
+      PADDLE_THROW(
+          platform::errors::PreconditionNotMet("Not compiled with CUDA."));
 #endif
     } else {
 #if defined(PADDLE_WITH_NCCL)

paddle/fluid/framework/details/gather_op_handle.cc

@@ -13,6 +13,7 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/gather_op_handle.h"
+
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
 
@@ -32,13 +33,20 @@ void GatherOpHandle::RunImpl() {
 
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), places_.size(),
-      "The number of output should equal to the number of places.");
+      platform::errors::InvalidArgument(
+          "The number of input variables should be equal "
+          "to the number of places, but got the number of input variables is "
+          "%d and the number of places is %d.",
+          in_var_handles.size(), places_.size()));
 
   VarHandle *out_var_handle;
   {
     auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
-    PADDLE_ENFORCE_EQ(out_var_handles.size(), 1,
-                      "The number of output should be one.");
+    PADDLE_ENFORCE_EQ(
+        out_var_handles.size(), 1,
+        platform::errors::InvalidArgument(
+            "The number of output variables should be 1, but got %d.",
+            out_var_handles.size()));
     out_var_handle = out_var_handles.front();
   }
 
@@ -47,10 +55,14 @@ void GatherOpHandle::RunImpl() {
   auto in_0_handle = in_var_handles[0];
   auto pre_in_var =
       var_scopes.at(in_0_handle->scope_idx())->FindVar(in_0_handle->name());
-  PADDLE_ENFORCE_NOT_NULL(pre_in_var);
+  PADDLE_ENFORCE_NOT_NULL(
+      pre_in_var,
+      platform::errors::NotFound("The variable '%s' is not found in the scope.",
+                                 in_0_handle->name()));
 
-  PADDLE_ENFORCE(pre_in_var->IsType<framework::SelectedRows>(),
-                 "Currently, gather_op only can gather SelectedRows.");
+  PADDLE_ENFORCE_EQ(pre_in_var->IsType<framework::SelectedRows>(), true,
+                    platform::errors::Unimplemented(
+                        "Currently, gather_op only supports SelectedRows."));
 
   // Wait input done, this Wait is asynchronous operation
   WaitInputVarGenerated();
@@ -63,7 +75,10 @@ void GatherOpHandle::RunImpl() {
   for (auto *in_handle : in_var_handles) {
     auto *in_var =
         var_scopes.at(in_handle->scope_idx())->FindVar(in_handle->name());
-    PADDLE_ENFORCE_NOT_NULL(in_var);
+    PADDLE_ENFORCE_NOT_NULL(
+        in_var,
+        platform::errors::NotFound(
+            "The variable '%s' is not found in the scope.", in_handle->name()));
     VariableVisitor::EnforceShapeAndDTypeEQ(*in_var, *pre_in_var);
 
     auto &in_sr_value = in_var->Get<framework::SelectedRows>();
@@ -76,15 +91,19 @@ void GatherOpHandle::RunImpl() {
   // NOTE: The Places of all input tensor must be all on CPU or all on GPU.
   platform::Place t_out_p = out_var_handle->place();
   if (platform::is_gpu_place(pre_in_value.place())) {
-    PADDLE_ENFORCE(platform::is_gpu_place(t_out_p),
-                   "Places of input and output must be all on GPU.");
+    PADDLE_ENFORCE_EQ(platform::is_gpu_place(t_out_p), true,
+                      platform::errors::PreconditionNotMet(
+                          "Places of input and output must be all on GPU."));
   } else {
     t_out_p = platform::CPUPlace();
   }
 
   auto out_var = var_scopes.at(out_var_handle->scope_idx())
                      ->FindVar(out_var_handle->name());
-  PADDLE_ENFORCE_NOT_NULL(out_var);
+  PADDLE_ENFORCE_NOT_NULL(
+      out_var,
+      platform::errors::NotFound("The variable '%s' is not found in the scope.",
+                                 out_var_handle->name()));
   auto out_value = out_var->GetMutable<framework::SelectedRows>();
   out_value->set_height(pre_in_value.height());
   out_value->set_rows(out_rows);

paddle/fluid/framework/details/gather_op_handle.h

@@ -24,6 +24,14 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/gather_op_handle_test.cc

@@ -13,15 +13,17 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/gather_op_handle.h"
+
 #include <memory>
 #include <unordered_map>
-#include "gtest/gtest.h"
 
-#include "paddle/fluid/platform/device_context.h"
+#include "gtest/gtest.h"
 
 namespace paddle {
 namespace framework {
 namespace details {
+struct DummyVarHandle;
+
 namespace f = paddle::framework;
 namespace p = paddle::platform;
 
@@ -60,7 +62,8 @@ struct TestGatherOpHandle {
         ctxs_.emplace_back(new p::CUDADeviceContext(p));
       }
 #else
-      PADDLE_THROW("CUDA is not support.");
+      PADDLE_THROW(
+          platform::errors::PreconditionNotMet("Not compiled with CUDA."));
 #endif
     } else {
       int count = 8;
@@ -141,7 +144,9 @@ struct TestGatherOpHandle {
     for (size_t input_scope_idx = 0; input_scope_idx < gpu_list_.size();
          ++input_scope_idx) {
       auto in_var = param_scopes_.at(input_scope_idx)->FindVar("input");
-      PADDLE_ENFORCE_NOT_NULL(in_var);
+      PADDLE_ENFORCE_NOT_NULL(
+          in_var, platform::errors::NotFound(
+                      "The variable '%s' is not found in the scope.", "input"));
       auto in_selected_rows = in_var->GetMutable<f::SelectedRows>();
       auto value = in_selected_rows->mutable_value();
       value->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
@@ -155,7 +160,9 @@ struct TestGatherOpHandle {
     }
 
     auto out_var = param_scopes_.at(output_scope_idx)->FindVar("out");
-    PADDLE_ENFORCE_NOT_NULL(out_var);
+    PADDLE_ENFORCE_NOT_NULL(
+        out_var, platform::errors::NotFound(
+                     "The variable '%s' is not found in the scope.", "out"));
     auto out_selected_rows = out_var->GetMutable<f::SelectedRows>();
 
     auto in_var = param_scopes_.at(output_scope_idx)->FindVar("input");
@@ -173,9 +180,19 @@ struct TestGatherOpHandle {
     auto& out_select_rows = out_var->Get<f::SelectedRows>();
     auto rt = out_select_rows.value();
 
-    PADDLE_ENFORCE_EQ(out_select_rows.height(), height, "height is not equal.");
+    PADDLE_ENFORCE_EQ(out_select_rows.height(), height,
+                      platform::errors::InvalidArgument(
+                          "The height of SelectedRows is not equal to "
+                          "the expected, expect %d, but got %d.",
+                          height, out_select_rows.height()));
+
     for (size_t k = 0; k < out_select_rows.rows().size(); ++k) {
-      PADDLE_ENFORCE_EQ(out_select_rows.rows()[k], rows[k % rows.size()]);
+      PADDLE_ENFORCE_EQ(
+          out_select_rows.rows()[k], rows[k % rows.size()],
+          platform::errors::InvalidArgument(
+              "The item at position %d of rows of SelectedRows is not equal to "
+              "the expected, expect %d, but got %d.",
+              k, rows[k % rows.size()], out_select_rows.rows()[k]));
     }
 
     f::Tensor result_tensor;
@@ -207,6 +224,7 @@ TEST(GatherTester, TestGPUGatherTestSelectedRows) {
   test_op.TestGatherSelectedRows(input_scope_idx);
 }
 #endif
+
 }  // namespace details
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/multi_devices_helper.h

@@ -20,16 +20,21 @@
 #include <unordered_set>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/details/var_handle.h"
-
+#include "paddle/fluid/framework/ir/graph.h"
+#include "paddle/fluid/framework/ir/pass.h"
 #include "paddle/fluid/framework/op_desc.h"
 #include "paddle/fluid/framework/op_proto_maker.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/platform/place.h"
 
-#include "paddle/fluid/framework/ir/graph.h"
-#include "paddle/fluid/framework/ir/pass.h"
+namespace paddle {
+namespace framework {
+class OpDesc;
+}  // namespace framework
+}  // namespace paddle
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/details/nan_inf_utils_detail.h

@@ -19,6 +19,12 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/place.h"
 
+namespace paddle {
+namespace framework {
+class Tensor;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/nccl_op_handle.h

@@ -46,14 +46,17 @@ class NCCLOpHandleBase : public OpHandleBase {
   }
   virtual ~NCCLOpHandleBase() {
     for (auto& ev : inter_events_) {
-      PADDLE_ENFORCE(cudaEventDestroy(ev.second));
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaEventDestroy(ev.second));
     }
     for (auto& ev : exter_events_) {
-      PADDLE_ENFORCE(cudaEventDestroy(ev.second));
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaEventDestroy(ev.second));
     }
   }
   void SetRunEnv(int run_order, bool use_hierarchical_allreduce) {
-    PADDLE_ENFORCE(run_order >= 0, "run_order must >= 0");
+    PADDLE_ENFORCE_GE(
+        run_order, 0,
+        platform::errors::InvalidArgument(
+            "The argument run_order must be >= 0, but got %d.", run_order));
     run_order_ = run_order;
     use_hierarchical_allreduce_ = use_hierarchical_allreduce;
 
@@ -74,8 +77,11 @@ class NCCLOpHandleBase : public OpHandleBase {
       return;
     }
 
-    PADDLE_ENFORCE(places_.size() == 1,
-                   "HierarchicalAllReduce run one proc with one card mode.");
+    PADDLE_ENFORCE_EQ(places_.size(), 1,
+                      platform::errors::InvalidArgument(
+                          "HierarchicalAllReduce can only run "
+                          "one proccess with one card mode, but got %d cards.",
+                          places_.size()));
 
     for (auto& p : places_) {
       auto ctxs = nccl_ctxs_->GetHierarchicalInterCtx(run_order);
@@ -88,11 +94,11 @@ class NCCLOpHandleBase : public OpHandleBase {
         continue;
       }
 
-      PADDLE_ENFORCE(cudaSetDevice(dev_id));
-      PADDLE_ENFORCE(cudaEventCreateWithFlags(&inter_events_[dev_id],
-                                              cudaEventDisableTiming));
-      PADDLE_ENFORCE(cudaEventCreateWithFlags(&exter_events_[dev_id],
-                                              cudaEventDisableTiming));
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaSetDevice(dev_id));
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaEventCreateWithFlags(
+          &inter_events_[dev_id], cudaEventDisableTiming));
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaEventCreateWithFlags(
+          &exter_events_[dev_id], cudaEventDisableTiming));
       VLOG(10) << "Create events on dev_id:" << dev_id
                << ", inter_event:" << &inter_events_[dev_id]
                << ", exter_event:" << &exter_events_[dev_id];
@@ -102,7 +108,10 @@ class NCCLOpHandleBase : public OpHandleBase {
   void FlatNCCLAllReduce(platform::Place place, const void* sendbuff,
                          void* recvbuff, size_t count, ncclDataType_t datatype,
                          ncclRedOp_t op) {
-    PADDLE_ENFORCE(run_order_ >= 0, "run_order must > 0");
+    PADDLE_ENFORCE_GE(
+        run_order_, 0,
+        platform::errors::InvalidArgument(
+            "The argument run_order_ must be >= 0, but got %d.", run_order_));
     auto flat_nccl_ctxs = nccl_ctxs_->GetFlatCtx(run_order_);
     int dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
     auto& nccl_ctx = flat_nccl_ctxs->at(dev_id);
@@ -113,14 +122,17 @@ class NCCLOpHandleBase : public OpHandleBase {
              << ", dev_id:" << dev_id << ", dtype:" << datatype
              << ", place:" << place;
 
-    PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclAllReduce(
         sendbuff, recvbuff, count, datatype, op, comm, stream));
   }
 
   void NCCLAllReduce(platform::Place place, const void* sendbuff,
                      void* recvbuff, size_t count, ncclDataType_t datatype,
                      ncclRedOp_t op) {
-    PADDLE_ENFORCE(run_order_ >= 0, "run_order must > 0");
+    PADDLE_ENFORCE_GE(
+        run_order_, 0,
+        platform::errors::InvalidArgument(
+            "The argument run_order_ must be >= 0, but got %d.", run_order_));
     if (!use_hierarchical_allreduce_) {
       FlatNCCLAllReduce(place, sendbuff, recvbuff, count, datatype, op);
       return;
@@ -132,7 +144,10 @@ class NCCLOpHandleBase : public OpHandleBase {
   void HierarchicalAllReduce(platform::Place place, const void* sendbuff,
                              void* recvbuff, size_t count,
                              ncclDataType_t datatype, ncclRedOp_t op) {
-    PADDLE_ENFORCE(run_order_ >= 0, "run_order must > 0");
+    PADDLE_ENFORCE_GE(
+        run_order_, 0,
+        platform::errors::InvalidArgument(
+            "The argument run_order_ must be >= 0, but got %d.", run_order_));
     InterReduce(place, sendbuff, recvbuff, count, datatype, op);
     // When a trainer is not in exter allreduce ring
     // they need not to call this.
@@ -157,14 +172,13 @@ class NCCLOpHandleBase : public OpHandleBase {
              << ", dtype:" << datatype << ", place:" << place
              << ", stream:" << stream;
 
-    PADDLE_ENFORCE(platform::dynload::ncclReduce(
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclReduce(
         sendbuff, recvbuff, count, datatype, ncclSum, 0, comm, stream));
 
     cudaEventRecord(inter_events_.at(dev_id), stream);
 
     if (FLAGS_sync_nccl_allreduce) {
-      PADDLE_ENFORCE(cudaStreamSynchronize(stream),
-                     "sync HierarchicalAllReduce inter stream error");
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaStreamSynchronize(stream));
     }
   }
 
@@ -172,7 +186,9 @@ class NCCLOpHandleBase : public OpHandleBase {
                       void* recvbuff, size_t count, ncclDataType_t datatype,
                       ncclRedOp_t op) {
     auto nccl_ctxs = nccl_ctxs_->GetHierarchicalExterCtx(run_order_);
-    PADDLE_ENFORCE(nccl_ctxs_, "can't get exter %d nccl_ctxs", run_order_);
+    PADDLE_ENFORCE_NOT_NULL(
+        nccl_ctxs_, platform::errors::NotFound(
+                        "Can't get exter %d nccl contexts.", run_order_));
     int dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
     auto& nccl_ctx = nccl_ctxs->at(dev_id);
     auto stream = nccl_ctx.stream();
@@ -185,14 +201,13 @@ class NCCLOpHandleBase : public OpHandleBase {
 
     cudaStreamWaitEvent(stream, inter_events_.at(dev_id), 0);
 
-    PADDLE_ENFORCE(platform::dynload::ncclAllReduce(
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclAllReduce(
         sendbuff, recvbuff, count, datatype, op, comm, stream));
 
     cudaEventRecord(exter_events_.at(dev_id), stream);
 
     if (FLAGS_sync_nccl_allreduce) {
-      PADDLE_ENFORCE(cudaStreamSynchronize(stream),
-                     "sync HierarchicalAllReduce exter stream error");
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaStreamSynchronize(stream));
     }
   }
 
@@ -210,8 +225,8 @@ class NCCLOpHandleBase : public OpHandleBase {
              << ", stream:" << stream;
 
     cudaStreamWaitEvent(stream, exter_events_.at(dev_id), 0);
-    PADDLE_ENFORCE(platform::dynload::ncclBcast(sendbuff, count, datatype, 0,
-                                                comm, stream));
+    PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclBcast(
+        sendbuff, count, datatype, 0, comm, stream));
   }
 
  protected:

paddle/fluid/framework/details/op_handle_base.cc

@@ -47,8 +47,8 @@ void OpHandleBase::InitCUDA() {
 #ifdef PADDLE_WITH_CUDA
   for (auto &p : dev_ctxes_) {
     int dev_id = BOOST_GET_CONST(platform::CUDAPlace, p.first).device;
-    PADDLE_ENFORCE(cudaSetDevice(dev_id));
-    PADDLE_ENFORCE(
+    PADDLE_ENFORCE_CUDA_SUCCESS(cudaSetDevice(dev_id));
+    PADDLE_ENFORCE_CUDA_SUCCESS(
         cudaEventCreateWithFlags(&events_[dev_id], cudaEventDisableTiming));
   }
   if (IsMultiDeviceTransfer() && dev_ctxes_.size() > 0) {
@@ -62,17 +62,22 @@ void OpHandleBase::InitCUDA() {
       }
     }
   } else {
-    PADDLE_ENFORCE_EQ(dev_ctxes_.size(), 1UL,
-                      "%s should have only one dev_ctx.", Name());
+    PADDLE_ENFORCE_EQ(
+        dev_ctxes_.size(), 1UL,
+        platform::errors::InvalidArgument(
+            "Operator %s should have only one dev_ctx, but got %d.", Name(),
+            dev_ctxes_.size()));
     auto &place = dev_ctxes_.begin()->first;
     int dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
     for (auto &out_var : outputs_) {
       auto *out_var_handle = dynamic_cast<VarHandle *>(out_var);
       if (out_var_handle) {
-        PADDLE_ENFORCE(platform::is_same_place(place, out_var_handle->place()),
-                       "The place of output(%s) is not consistent with the "
-                       "place of current op(%s).",
-                       out_var_handle->Name(), Name());
+        PADDLE_ENFORCE_EQ(
+            platform::is_same_place(place, out_var_handle->place()), true,
+            platform::errors::InvalidArgument(
+                "The place of output(%s) is not consistent with the "
+                "place of current op(%s).",
+                out_var_handle->Name(), Name()));
         out_var_handle->SetGenerateEvent(events_.at(dev_id));
       }
     }
@@ -86,7 +91,10 @@ void OpHandleBase::Run(bool use_cuda) {
     InitCUDA();
   }
 #else
-  PADDLE_ENFORCE(!use_cuda);
+  PADDLE_ENFORCE_EQ(use_cuda, false,
+                    platform::errors::InvalidArgument(
+                        "Argument use_cuda should be false when Paddle is not "
+                        "compiled with CUDA."));
 #endif
 
   // skip running current op, used with inplace_addto_op_pass
@@ -100,17 +108,20 @@ void OpHandleBase::Run(bool use_cuda) {
 
 void OpHandleBase::RecordWaitEventOnCtx(platform::DeviceContext *waited_ctx) {
 #ifdef PADDLE_WITH_CUDA
-  PADDLE_ENFORCE_NOT_NULL(waited_ctx);
+  PADDLE_ENFORCE_NOT_NULL(waited_ctx, platform::errors::InvalidArgument(
+                                          "Argument waited_ctx is NULL."));
   if (platform::is_cpu_place(waited_ctx->GetPlace()) || events_.empty()) {
     for (auto &dev_ctx : dev_ctxes_) {
-      PADDLE_ENFORCE_NOT_NULL(dev_ctx.second);
+      PADDLE_ENFORCE_NOT_NULL(
+          dev_ctx.second,
+          platform::errors::InvalidArgument("The device context is NULL."));
       dev_ctx.second->Wait();
     }
   } else {
     auto stream =
         static_cast<platform::CUDADeviceContext *>(waited_ctx)->stream();
     for (auto &ev : events_) {
-      PADDLE_ENFORCE(cudaStreamWaitEvent(stream, ev.second, 0));
+      PADDLE_ENFORCE_CUDA_SUCCESS(cudaStreamWaitEvent(stream, ev.second, 0));
     }
   }
 #else
@@ -145,10 +156,11 @@ void OpHandleBase::WaitInputVarGenerated() {
           auto stream =
               static_cast<platform::CUDADeviceContext *>(dev_ctxes_.at(place))
                   ->stream();
-          PADDLE_ENFORCE(
+          PADDLE_ENFORCE_CUDA_SUCCESS(
               cudaStreamWaitEvent(stream, in_var_handle->GetEvent(), 0));
 #else
-          PADDLE_THROW("Doesn't compile the GPU.");
+          PADDLE_THROW(
+              platform::errors::PreconditionNotMet("Not compiled with CUDA."));
 #endif
         }
         // There are nothing to do when the place is CPUPlace.
@@ -169,10 +181,11 @@ void OpHandleBase::WaitInputVarGenerated(const platform::Place &place) {
           auto stream = static_cast<platform::CUDADeviceContext *>(
                             dev_ctxes_.at(in_var_handle->place()))
                             ->stream();
-          PADDLE_ENFORCE(
+          PADDLE_ENFORCE_CUDA_SUCCESS(
               cudaStreamWaitEvent(stream, in_var_handle->GetEvent(), 0));
 #else
-          PADDLE_THROW("Doesn't compile the GPU.");
+          PADDLE_THROW(
+              platform::errors::PreconditionNotMet("Not compiled with CUDA."));
 #endif
         }
         // There are nothing to do when the place is CPUPlace.
@@ -242,7 +255,9 @@ void OpHandleBase::SetLocalExecScopes(
   auto scopes = GetLocalScopes();
   for (auto *scope : scopes) {
     auto iter = scope_map.find(scope);
-    PADDLE_ENFORCE(iter != scope_map.end(), "Local scope not found");
+    PADDLE_ENFORCE_NE(
+        iter, scope_map.end(),
+        platform::errors::NotFound("Local scope not found in scope map."));
     local_exec_scopes_.emplace_back(iter->second);
   }
 }

paddle/fluid/framework/details/op_handle_base.h

@@ -24,10 +24,22 @@
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/macros.h"
 
+namespace paddle {
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
 class Scope;
+namespace details {
+struct VarHandleBase;
+}  // namespace details
+namespace ir {
+class Node;
+}  // namespace ir
 
 namespace details {
 

paddle/fluid/framework/details/op_registry.h

@@ -21,6 +21,7 @@ limitations under the License. */
 #include <unordered_map>
 #include <unordered_set>
 #include <vector>
+
 #include "paddle/fluid/framework/grad_op_desc_maker.h"
 #include "paddle/fluid/framework/inplace_op_inference.h"
 #include "paddle/fluid/framework/no_need_buffer_vars_inference.h"
@@ -186,19 +187,20 @@ struct OpInfoFiller<T, kOpProtoAndCheckerMaker> {
   void operator()(const char* op_type, OpInfo* info) const {
     PADDLE_ENFORCE_EQ(info->proto_, nullptr,
                       platform::errors::AlreadyExists(
-                          "OpProto of %s has been registered", op_type));
+                          "OpProto of %s has been registered.", op_type));
     PADDLE_ENFORCE_EQ(info->checker_, nullptr,
                       platform::errors::AlreadyExists(
-                          "OpAttrChecker of %s has been registered", op_type));
+                          "OpAttrChecker of %s has been registered.", op_type));
     info->proto_ = new proto::OpProto;
     info->checker_ = new OpAttrChecker();
     T maker;
     maker(info->proto_, info->checker_);
     info->proto_->set_type(op_type);
-    PADDLE_ENFORCE(
-        info->proto_->IsInitialized(),
-        "Fail to initialize %s's OpProto, because %s is not initialized",
-        op_type, info->proto_->InitializationErrorString());
+    PADDLE_ENFORCE_EQ(
+        info->proto_->IsInitialized(), true,
+        platform::errors::PreconditionNotMet(
+            "Fail to initialize %s's OpProto, because %s is not initialized.",
+            op_type, info->proto_->InitializationErrorString()));
   }
 };
 

paddle/fluid/framework/details/reduce_and_gather.h

@@ -16,6 +16,7 @@
 #include <algorithm>
 #include <map>
 #include <vector>
+
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/selected_rows.h"
@@ -32,9 +33,13 @@ struct ReduceLoDTensor {
 
   template <typename T>
   void apply() const {
-    PADDLE_ENFORCE(!src_tensors_.empty());
+    PADDLE_ENFORCE_NE(src_tensors_.empty(), true,
+                      platform::errors::InvalidArgument(
+                          "The number of tensors to be reduced is 0."));
     auto &t0 = *src_tensors_[0];
-    PADDLE_ENFORCE_NE(t0.numel(), 0);
+    PADDLE_ENFORCE_NE(t0.numel(), 0,
+                      platform::errors::InvalidArgument(
+                          "The size of first tensor to be reduced is 0."));
 
     dst_tensor_.Resize(t0.dims());
     T *dst = dst_tensor_.mutable_data<T>(platform::CPUPlace());
@@ -45,8 +50,19 @@ struct ReduceLoDTensor {
         continue;
       }
 
-      PADDLE_ENFORCE_EQ(t.dims(), t0.dims());
-      PADDLE_ENFORCE_EQ(t.type(), t0.type());
+      PADDLE_ENFORCE_EQ(t.dims(), t0.dims(),
+                        platform::errors::InvalidArgument(
+                            "The shape of tensors to be reduced must be "
+                            "consistent. The shape of current tensor is %s, "
+                            "but the shape of the first tensor is %s.",
+                            t.dims(), t0.dims()));
+
+      PADDLE_ENFORCE_EQ(t.type(), t0.type(),
+                        platform::errors::InvalidArgument(
+                            "The type of tensors to be reduced must be "
+                            "consistent. The type of current tensor is %s, "
+                            "but the type of the first tensor is %s.",
+                            t.type(), t0.type()));
       std::transform(t.data<T>(), t.data<T>() + t.numel(), dst, dst,
                      [](T a, T b) -> T { return a + b; });
     }
@@ -88,7 +104,9 @@ struct GatherLocalSelectedRowsFunctor {
         in_places_(in_places),
         out_place_(out_place),
         dst_selected_rows_(dst_selected_rows) {
-    PADDLE_ENFORCE_EQ(src_selected_rows.empty(), false);
+    PADDLE_ENFORCE_NE(src_selected_rows.empty(), true,
+                      platform::errors::InvalidArgument(
+                          "The number of selected_rows to be gathered is 0."));
 
     std::vector<int64_t> out_rows;
 

paddle/fluid/framework/details/reduce_op_handle.cc

@@ -13,7 +13,9 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/reduce_op_handle.h"
+
 #include <memory>
+
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
 #include "paddle/fluid/framework/details/variable_visitor.h"
@@ -116,8 +118,15 @@ void ReduceOpHandle::GatherSelectedRows(
   merged_dev_ctx->Wait();
   scope->EraseVars(std::vector<std::string>{gathered_var_name});
 
-  PADDLE_ENFORCE(client->Gather(vars, &remote, *merged_dev_ctx, scope));
-  PADDLE_ENFORCE(remote.size() == vars.size());
+  PADDLE_ENFORCE_EQ(
+      client->Gather(vars, &remote, *merged_dev_ctx, scope), true,
+      platform::errors::PreconditionNotMet("Gather SelectedRows failed."));
+  PADDLE_ENFORCE_EQ(remote.size(), vars.size(),
+                    platform::errors::PreconditionNotMet(
+                        "The number of remotes should be equal to the number "
+                        "of variables to be gathered, but got the number of "
+                        "remotes is %d and the number of variables is %d.",
+                        remote.size(), vars.size()));
 
   // 4. merged local selected rows.
   std::vector<const SelectedRows *> all;
@@ -151,14 +160,19 @@ void ReduceOpHandle::RunImpl() {
 
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), places_.size(),
-      "The number of output should equal to the number of places.");
+      platform::errors::InvalidArgument(
+          "The number of inputs should equal to the number of places, but got "
+          "the number of inputs is %d and the number of places is %d.",
+          in_var_handles.size(), places_.size()));
 
   VarHandle *out_var_handle;
   {
     auto out_var_handles = DynamicCast<VarHandle>(outputs_);
 
     PADDLE_ENFORCE_EQ(out_var_handles.size(), 1UL,
-                      "The number of output should be one.");
+                      platform::errors::InvalidArgument(
+                          "The number of output should be one, but got %d.",
+                          out_var_handles.size()));
     out_var_handle = out_var_handles.front();
   }
 
@@ -168,7 +182,10 @@ void ReduceOpHandle::RunImpl() {
 
   auto pre_in_var =
       var_scopes.at(in_0_handle->scope_idx())->FindVar(in_0_handle->name());
-  PADDLE_ENFORCE_NOT_NULL(pre_in_var);
+
+  PADDLE_ENFORCE_NOT_NULL(pre_in_var, platform::errors::NotFound(
+                                          "Variable %s is not found in scope.",
+                                          in_0_handle->name()));
 
   // NOTE: The Places of all input tensor must be all on CPU or all on GPU.
   std::vector<platform::Place> in_places;  // used to get dev_ctx
@@ -176,21 +193,29 @@ void ReduceOpHandle::RunImpl() {
     in_places.emplace_back(in_handle->place());
     auto in_var =
         var_scopes.at(in_handle->scope_idx())->FindVar(in_handle->name());
-    PADDLE_ENFORCE_NOT_NULL(in_var);
+
+    PADDLE_ENFORCE_NOT_NULL(
+        in_var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                           in_handle->name()));
+
     VariableVisitor::EnforceShapeAndDTypeEQ(*pre_in_var, *in_var);
   }
 
   auto out_var = var_scopes.at(out_var_handle->scope_idx())
                      ->FindVar(out_var_handle->name());
-  PADDLE_ENFORCE_NOT_NULL(out_var);
+
+  PADDLE_ENFORCE_NOT_NULL(
+      out_var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                          out_var_handle->name()));
 
   // NOTE: The tensors' Place of input and output must be all on GPU or all on
   // CPU.
   auto in_p = VariableVisitor::GetMutableTensor(pre_in_var).place();
   platform::Place t_out_p;
   if (platform::is_gpu_place(in_p)) {
-    PADDLE_ENFORCE(platform::is_gpu_place(out_var_handle->place()),
-                   "Places of input and output must be all on GPU.");
+    PADDLE_ENFORCE_EQ(platform::is_gpu_place(out_var_handle->place()), true,
+                      platform::errors::PreconditionNotMet(
+                          "Places of input and output must be all on GPU."));
     t_out_p = out_var_handle->place();
   } else {
     t_out_p = platform::CPUPlace();
@@ -229,7 +254,10 @@ void ReduceOpHandle::RunImpl() {
             in_selected_rows, in_places, dev_ctxes_, out_var_handle, t_out_p,
             out_var->GetMutable<framework::SelectedRows>());
       } else {
-        PADDLE_THROW("only support double or float when gather SelectedRows");
+        PADDLE_THROW(platform::errors::Unimplemented(
+            "Only support double or float when gather SelectedRows, but got "
+            "%s.",
+            framework::DataTypeToString(in_selected_rows[0]->value().type())));
       }
 #endif
     });
@@ -292,7 +320,7 @@ void ReduceOpHandle::RunImpl() {
         size_t numel = static_cast<size_t>(lod_tensor.numel());
         all_reduce_calls.emplace_back(
             [buffer, recvbuffer, type, numel, root_id, &nccl_ctx] {
-              PADDLE_ENFORCE(platform::dynload::ncclReduce(
+              PADDLE_ENFORCE_CUDA_SUCCESS(platform::dynload::ncclReduce(
                   buffer, recvbuffer, numel, static_cast<ncclDataType_t>(type),
                   ncclSum, root_id, nccl_ctx.comm_, nccl_ctx.stream()));
             });
@@ -306,10 +334,13 @@ void ReduceOpHandle::RunImpl() {
         }
       });
 #else
-      PADDLE_THROW("CUDA is not enabled.");
+      PADDLE_THROW(
+          platform::errors::PreconditionNotMet("Not compiled with CUDA."));
 #endif
     } else {
-      PADDLE_THROW("Place should be CPUPlace or CUDAPlace.");
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "The place of tensor should be CPUPlace or CUDAPlace, but got %s.",
+          lod_tensors[0]->place()));
     }
   }
 }

paddle/fluid/framework/details/reduce_op_handle.h

@@ -24,6 +24,21 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/device_context.h"
+
+namespace paddle {
+namespace framework {
+class SelectedRows;
+namespace details {
+struct VarHandle;
+}  // namespace details
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+struct NCCLContextMap;
+}  // namespace platform
+}  // namespace paddle
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif

paddle/fluid/framework/details/reduce_op_handle_test.cc

@@ -13,7 +13,9 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/reduce_op_handle.h"
+
 #include <unordered_map>
+
 #include "gtest/gtest.h"
 #include "paddle/fluid/platform/device_context.h"
 
@@ -69,7 +71,8 @@ struct TestReduceOpHandle {
       }
       nccl_ctxs_.reset(new platform::NCCLContextMap(gpu_list_));
 #else
-      PADDLE_THROW("CUDA is not support.");
+      PADDLE_THROW(
+          platform::errors::PreconditionNotMet("Not compiled with NCLL."));
 #endif
     } else {
       int count = 8;
@@ -103,7 +106,8 @@ struct TestReduceOpHandle {
       op_handle_.reset(new ReduceOpHandle(nodes.back().get(), local_scopes_,
                                           gpu_list_, nccl_ctxs_.get()));
 #else
-      PADDLE_THROW("CUDA is not support.");
+      PADDLE_THROW(
+          platform::errors::PreconditionNotMet("Not compiled with NCLL."));
 #endif
     } else {
 #if defined(PADDLE_WITH_NCCL)
@@ -164,7 +168,10 @@ struct TestReduceOpHandle {
     for (size_t input_scope_idx = 0; input_scope_idx < gpu_list_.size();
          ++input_scope_idx) {
       auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
-      PADDLE_ENFORCE_NOT_NULL(in_var);
+
+      PADDLE_ENFORCE_NOT_NULL(
+          in_var, platform::errors::NotFound(
+                      "Variable %s is not found in scope.", "input"));
       auto in_selected_rows = in_var->GetMutable<f::SelectedRows>();
       auto value = in_selected_rows->mutable_value();
       value->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
@@ -178,7 +185,9 @@ struct TestReduceOpHandle {
     }
 
     auto out_var = param_scopes_[output_scope_idx]->FindVar("out");
-    PADDLE_ENFORCE_NOT_NULL(out_var);
+    PADDLE_ENFORCE_NOT_NULL(out_var,
+                            platform::errors::NotFound(
+                                "Variable %s is not found in scope.", "out"));
     auto out_selected_rows = out_var->GetMutable<f::SelectedRows>();
 
     auto in_var = param_scopes_[output_scope_idx]->FindVar("input");
@@ -196,9 +205,18 @@ struct TestReduceOpHandle {
     auto &out_select_rows = out_var->Get<f::SelectedRows>();
     auto rt = out_select_rows.value();
 
-    PADDLE_ENFORCE_EQ(out_select_rows.height(), height, "height is not equal.");
+    PADDLE_ENFORCE_EQ(out_select_rows.height(), height,
+                      platform::errors::InvalidArgument(
+                          "The height of SelectedRows is not equal to "
+                          "the expected, expect %d, but got %d.",
+                          height, out_select_rows.height()));
     for (size_t k = 0; k < out_select_rows.rows().size(); ++k) {
-      PADDLE_ENFORCE_EQ(out_select_rows.rows()[k], rows[k % rows.size()]);
+      PADDLE_ENFORCE_EQ(
+          out_select_rows.rows()[k], rows[k % rows.size()],
+          platform::errors::InvalidArgument(
+              "The item at position %d of rows of SelectedRows is not equal to "
+              "the expected, expect %d, but got %d.",
+              k, rows[k % rows.size()], out_select_rows.rows()[k]));
     }
 
     f::Tensor result_tensor;
@@ -208,7 +226,7 @@ struct TestReduceOpHandle {
     for (int64_t j = 0; j < f::product(result_tensor.dims()); ++j) {
       ASSERT_NEAR(ct[j], send_vector[j % send_vector.size()], 1e-5);
     }
-  }
+  }  // namespace details
 
   void TestReduceLodTensors(size_t output_scope_idx) {
     std::vector<float> send_vector(static_cast<size_t>(f::product(kDims)));
@@ -220,7 +238,9 @@ struct TestReduceOpHandle {
     for (size_t input_scope_idx = 0; input_scope_idx < gpu_list_.size();
          ++input_scope_idx) {
       auto in_var = param_scopes_[input_scope_idx]->FindVar("input");
-      PADDLE_ENFORCE_NOT_NULL(in_var);
+      PADDLE_ENFORCE_NOT_NULL(
+          in_var, platform::errors::NotFound(
+                      "Variable %s is not found in scope.", "input"));
       auto in_lod_tensor = in_var->GetMutable<f::LoDTensor>();
       in_lod_tensor->mutable_data<float>(kDims, gpu_list_[input_scope_idx]);
       in_lod_tensor->set_lod(lod);
@@ -230,7 +250,9 @@ struct TestReduceOpHandle {
     }
 
     auto out_var = param_scopes_[output_scope_idx]->FindVar("out");
-    PADDLE_ENFORCE_NOT_NULL(out_var);
+    PADDLE_ENFORCE_NOT_NULL(out_var,
+                            platform::errors::NotFound(
+                                "Variable %s is not found in scope.", "out"));
     auto out_lodtensor = out_var->GetMutable<f::LoDTensor>();
 
     auto in_var = param_scopes_[output_scope_idx]->FindVar("input");
@@ -254,7 +276,7 @@ struct TestReduceOpHandle {
       ASSERT_NEAR(ct[j], send_vector[j] * gpu_list_.size(), 1e-5);
     }
   }
-};
+};  // namespace details
 
 TEST(ReduceTester, TestCPUReduceTestSelectedRows) {
   TestReduceOpHandle test_op;

paddle/fluid/framework/details/rpc_op_handle.h

@@ -24,6 +24,16 @@
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/scope.h"
 
+namespace paddle {
+namespace framework {
+class OpDesc;
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/scale_loss_grad_op_handle.cc

@@ -13,9 +13,17 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/details/scale_loss_grad_op_handle.h"
+
 #include <string>
+
 #include "paddle/fluid/platform/profiler.h"
 
+namespace paddle {
+namespace framework {
+class Tensor;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/scale_loss_grad_op_handle.h

@@ -21,6 +21,18 @@
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/scope.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/scope_buffered_monitor.h

@@ -17,7 +17,9 @@
 #include <unordered_set>
 #include <utility>
 #include <vector>
+
 #include "paddle/fluid/framework/scope.h"
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/share_tensor_buffer_functor.cc

@@ -22,6 +22,17 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/enforce.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+class Tensor;
+class Variable;
+namespace ir {
+class MemOptVarInfo;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
@@ -100,13 +111,12 @@ void ShareTensorBufferFunctor::CallOnce() {
     auto *out_var = exec_scope_->FindVar(out_var_names_[i]);
     PADDLE_ENFORCE_NOT_NULL(
         in_var, platform::errors::NotFound(
-                    "The input variable(%s)to be inplaced should not be NULL.",
+                    "The variable(%s) to be inplaced is not found in scope.",
                     in_var_infos_[i]->Name()));
     PADDLE_ENFORCE_NOT_NULL(
-        out_var,
-        platform::errors::NotFound(
-            "The output variable(%s) to be inplaced should not be NULL.",
-            out_var_names_[i]));
+        out_var, platform::errors::NotFound(
+                     "The variable(%s) to be inplaced is not found in scope.",
+                     out_var_names_[i]));
     PADDLE_ENFORCE_NE(
         in_var, out_var,
         platform::errors::PreconditionNotMet(

paddle/fluid/framework/details/share_tensor_buffer_functor.h

@@ -25,6 +25,15 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/variable.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class MemOptVarInfo;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/share_tensor_buffer_op_handle.cc

@@ -23,10 +23,20 @@
 #include "paddle/fluid/framework/selected_rows.h"
 #include "paddle/fluid/platform/enforce.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class MemOptVarInfo;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
 
+class ComputationOpHandle;
+
 ComputationOpHandle *GetUniquePendingComputationOpHandle(
     ShareTensorBufferOpHandle *share_tensor_op) {
   ComputationOpHandle *result_op = nullptr;

paddle/fluid/framework/details/share_tensor_buffer_op_handle.h

@@ -22,10 +22,22 @@
 #include "paddle/fluid/framework/details/op_handle_base.h"
 #include "paddle/fluid/framework/details/share_tensor_buffer_functor.h"
 
+namespace paddle {
+namespace framework {
+class Scope;
+namespace ir {
+class MemOptVarInfo;
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {
 
+class ComputationOpHandle;
+
 class ShareTensorBufferOpHandle : public OpHandleBase {
  public:
   ShareTensorBufferOpHandle(

paddle/fluid/framework/details/sparse_all_reduce_op_handle.cc

@@ -12,8 +12,10 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 #include "paddle/fluid/framework/details/sparse_all_reduce_op_handle.h"
+
 #include <algorithm>
 #include <utility>
+
 #include "dgc/dgc.h"
 #include "paddle/fluid/framework/details/container_cast.h"
 #include "paddle/fluid/framework/details/reduce_and_gather.h"
@@ -38,18 +40,23 @@ SparseAllReduceOpHandle::SparseAllReduceOpHandle(
       is_encoded_(is_encoded),
       nranks_(nranks) {
   // TODO(gongwb) :polish them!
-  PADDLE_ENFORCE_EQ(is_encoded, true);
+  PADDLE_ENFORCE_EQ(is_encoded, true, platform::errors::InvalidArgument(
+                                          "The argument is_encoded is false."));
   VLOG(1) << "Use dgc allreduce mode"
           << ", nranks:" << nranks_;
 
-  PADDLE_ENFORCE_GT(local_scopes_.size(), 0);
+  PADDLE_ENFORCE_GT(local_scopes_.size(), 0,
+                    platform::errors::PreconditionNotMet(
+                        "The number of local scope should be > 0, but got %zu.",
+                        local_scopes_.size()));
   auto nranks_name = g_dgc_nranks;
   for (size_t i = 0; i < local_scopes_.size(); ++i) {
     auto *local_scope = local_scopes_[i];
     auto nranks_var = local_scope->FindVar(nranks_name);
-    if (nranks_var == nullptr) {
-      PADDLE_THROW("not find nranks_var:%s", nranks_name);
-    }
+
+    PADDLE_ENFORCE_NOT_NULL(
+        nranks_var, platform::errors::NotFound(
+                        "Variable %s is not found in scope.", nranks_name));
 
     float *dgc_nranks = nranks_var->GetMutable<LoDTensor>()->data<float>();
     *dgc_nranks = nranks;
@@ -64,10 +71,18 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
   auto out_var_handles = DynamicCast<VarHandle>(this->Outputs());
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), places_.size(),
-      "The NoDummyInputSize should be equal to the number of places.");
+      platform::errors::PreconditionNotMet(
+          "The number of input variables should be equal to the number of "
+          "places, but got the number of input variables is %zu and the the "
+          "number of places is %zu.",
+          in_var_handles.size(), places_.size()));
   PADDLE_ENFORCE_EQ(
       in_var_handles.size(), out_var_handles.size(),
-      "The NoDummyInputSize and NoDummyOutputSize should be equal.");
+      platform::errors::PreconditionNotMet(
+          "The number of input variables should be equal to the number of "
+          "output variables, but got the number of input variables is %zu and "
+          "the the number of output variables is %zu.",
+          in_var_handles.size(), out_var_handles.size()));
 
   std::vector<const LoDTensor *> ins;
   std::vector<LoDTensor *> gathers;
@@ -80,14 +95,17 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
 
     auto encode_var_name = original_name + g_dgc_encoded;
     auto *in_var = local_scope->FindVar(encode_var_name);
-    PADDLE_ENFORCE_NOT_NULL(in_var, "%s should not be null", encode_var_name);
+    PADDLE_ENFORCE_NOT_NULL(
+        in_var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                           encode_var_name));
     auto &in = in_var->Get<LoDTensor>();
     ins.emplace_back(&in);
 
     auto gather_var_name = original_name + g_dgc_gather;
     auto *gather_var = local_scope->FindVar(gather_var_name);
-    PADDLE_ENFORCE_NOT_NULL(gather_var, "%s should not be null",
-                            gather_var_name);
+    PADDLE_ENFORCE_NOT_NULL(
+        gather_var, platform::errors::NotFound(
+                        "Variable %s is not found in scope.", gather_var));
     auto *gather = gather_var->GetMutable<LoDTensor>();
     gathers.emplace_back(gather);
 
@@ -100,14 +118,26 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
     }
   }
 
-  PADDLE_ENFORCE(platform::is_gpu_place(ins[0]->place()));
-  PADDLE_ENFORCE(platform::is_gpu_place(outs[0]->place()));
-  PADDLE_ENFORCE(nccl_ctxs_, "nccl_ctxs should not be nullptr.");
+  PADDLE_ENFORCE_EQ(
+      platform::is_gpu_place(ins[0]->place()), true,
+      platform::errors::InvalidArgument(
+          "The place of input variable should be CUDAPlace, but got %s.",
+          ins[0]->place()));
+  PADDLE_ENFORCE_EQ(
+      platform::is_gpu_place(outs[0]->place()), true,
+      platform::errors::InvalidArgument(
+          "The place of input variable should be CUDAPlace, but got %s.",
+          outs[0]->place()));
+  PADDLE_ENFORCE_NOT_NULL(nccl_ctxs_, platform::errors::PreconditionNotMet(
+                                          "The nccl contexts are NULL."));
 
   int dtype = -1;
   size_t in_numel = 0;
   size_t out_numel = 0;
-  PADDLE_ENFORCE(nranks_ > 1);
+  PADDLE_ENFORCE_GT(
+      nranks_, 1,
+      platform::errors::PreconditionNotMet(
+          "The number of ranks should be > 1, but got %d.", nranks_));
   std::vector<std::function<void()>> all_gather_calls;
   std::vector<std::function<void()>> sparse_reduce_calls;
 
@@ -123,8 +153,16 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
 
     dtype = (dtype == -1) ? platform::ToNCCLDataType(in.type()) : dtype;
     in_numel = (in_numel == 0) ? static_cast<size_t>(in.numel()) : in_numel;
-    PADDLE_ENFORCE(in_numel % 2 == 0);
-    PADDLE_ENFORCE(in_numel / 2 == static_cast<size_t>(k));
+    PADDLE_ENFORCE_EQ(in_numel % 2, 0,
+                      platform::errors::InvalidArgument(
+                          "The number of elements of input variable should be "
+                          "even, but got %zu.",
+                          in_numel));
+    PADDLE_ENFORCE_EQ(in_numel / 2, static_cast<size_t>(k),
+                      platform::errors::InvalidArgument(
+                          "The number of elements of input variable should be "
+                          "even, but got %zu.",
+                          in_numel));
     out_numel = (out_numel == 0) ? static_cast<size_t>(out.numel()) : out_numel;
 
     int dev_id = BOOST_GET_CONST(platform::CUDAPlace, place).device;
@@ -154,7 +192,8 @@ void SparseAllReduceOpHandle::RunImplEncoded() {
       PADDLE_ENFORCE_EQ(paddle::communication::dgc::sparseReduce(
                             gather_buff, k, out_tensor_buf,
                             static_cast<int>(out_numel), nranks_, stream),
-                        true);
+                        true, platform::errors::Unavailable(
+                                  "Calling sparseReduce() failed."));
     });
   }
 
@@ -187,11 +226,16 @@ void SparseAllReduceOpHandle::SparseAllReduceFunc(
 int SparseAllReduceOpHandle::GetKValue(const std::string &grad_name) {
   auto original_name = paddle::framework::GradOriginalVarName(grad_name);
   auto var_name = original_name + g_dgc_k;
-  PADDLE_ENFORCE(local_scopes_.size() > 0);
+  PADDLE_ENFORCE_GT(local_scopes_.size(), 0,
+                    platform::errors::PreconditionNotMet(
+                        "The number of local scope should be > 0, but got %zu.",
+                        local_scopes_.size()));
 
   auto *scope = local_exec_scopes_[0];
   auto var = scope->FindVar(var_name);
-  PADDLE_ENFORCE_NOT_NULL(var);
+  PADDLE_ENFORCE_NOT_NULL(
+      var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                      var_name));
   auto tensor = var->Get<LoDTensor>().data<float>();
   return *tensor;
 }
@@ -202,15 +246,22 @@ bool SparseAllReduceOpHandle::IsEncoded() {
   }
   auto counter_name = g_dgc_counter_name;
   auto step_name = g_dgc_rampup_begin_step;
-  PADDLE_ENFORCE(local_scopes_.size() > 0);
+
+  PADDLE_ENFORCE_GT(local_scopes_.size(), 0,
+                    platform::errors::PreconditionNotMet(
+                        "The number of local scope should be > 0, but got %zu.",
+                        local_scopes_.size()));
 
   auto *local_scope = local_exec_scopes_[0];
   auto count_var = local_scope->FindVar(counter_name);
   auto step_var = local_scope->FindVar(step_name);
-  if (count_var == nullptr || step_var == nullptr) {
-    PADDLE_THROW("not find count_var:%s or step_var:%s", counter_name,
-                 step_var);
-  }
+
+  PADDLE_ENFORCE_NOT_NULL(
+      count_var, platform::errors::NotFound(
+                     "Variable %s is not found in scope.", counter_name));
+  PADDLE_ENFORCE_NOT_NULL(
+      step_var, platform::errors::NotFound("Variable %s is not found in scope.",
+                                           step_var));
 
   float count = *count_var->Get<LoDTensor>().data<float>();
   float step = *step_var->Get<LoDTensor>().data<float>();

paddle/fluid/framework/details/sparse_all_reduce_op_handle.h

@@ -23,6 +23,17 @@
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/platform/nccl_helper.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+namespace platform {
+class NCCLCommunicator;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/var_handle.h

@@ -24,6 +24,14 @@
 #include "paddle/fluid/platform/macros.h"
 #include "paddle/fluid/platform/place.h"
 
+namespace paddle {
+namespace framework {
+namespace ir {
+class Node;
+}  // namespace ir
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/variable_visitor.cc

@@ -15,6 +15,14 @@
 #include "paddle/fluid/framework/details/variable_visitor.h"
 
 #include "paddle/fluid/framework/selected_rows.h"
+
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/details/variable_visitor.h

@@ -17,6 +17,13 @@
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/variable.h"
 
+namespace paddle {
+namespace framework {
+class Tensor;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 namespace details {

paddle/fluid/framework/device_worker.cc

@@ -13,11 +13,13 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/device_worker.h"
-#include "xxhash.h"  // NOLINT
 
 namespace paddle {
 namespace framework {
 
+class LoDTensor;
+class Scope;
+
 void DeviceWorker::SetRootScope(Scope* root_scope) { root_scope_ = root_scope; }
 
 void DeviceWorker::SetDataFeed(DataFeed* data_feed) {

paddle/fluid/framework/device_worker.h

@@ -39,6 +39,18 @@ limitations under the License. */
 #include "paddle/fluid/platform/port.h"
 #include "paddle/fluid/platform/timer.h"
 
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class ProgramDesc;
+class Scope;
+class Tensor;
+}  // namespace framework
+namespace platform {
+class DeviceContext;
+}  // namespace platform
+}  // namespace paddle
+
 #if defined(PADDLE_WITH_NCCL)
 #include "paddle/fluid/platform/nccl_helper.h"
 #endif
@@ -62,7 +74,9 @@ class PullDenseWorker {
   virtual void Initialize(const TrainerDesc& param);
 #ifdef PADDLE_WITH_CUDA
   void AddStream(const cudaStream_t stream) { copy_streams_.push_back(stream); }
+#endif
 
+#if (defined PADDLE_WITH_CUDA) || (defined PADDLE_WITH_XPU)
   void AddPlace(const paddle::platform::Place place) {
     places_.push_back(place);
   }
@@ -123,9 +137,9 @@ class PullDenseWorker {
 
 #ifdef PADDLE_WITH_CUDA
   std::vector<cudaStream_t> copy_streams_;
+#endif
   std::vector<paddle::platform::Place> places_;
   std::vector<Scope*> thread_scopes_;
-#endif
 };
 
 // should incorporate different type of device
@@ -149,6 +163,7 @@ class DeviceWorker {
   virtual void SetDataFeed(DataFeed* data_feed);
   virtual void SetWorkerNum(int num) {}
   virtual void CacheProgram(const ProgramDesc& main_program) {}
+  virtual void GetXpuOpIndex() {}
   virtual void SetNeedDumpField(bool need_dump_field) {
     need_dump_field_ = need_dump_field;
   }

paddle/fluid/framework/device_worker_factory.cc

@@ -20,6 +20,8 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class DeviceWorker;
+
 typedef std::shared_ptr<DeviceWorker> (*Createdevice_workerFunction)();
 typedef std::unordered_map<std::string, Createdevice_workerFunction>
     device_workerMap;

paddle/fluid/framework/device_worker_factory.h

@@ -16,11 +16,14 @@ limitations under the License. */
 
 #include <memory>
 #include <string>
+
 #include "paddle/fluid/framework/device_worker.h"
 
 namespace paddle {
 namespace framework {
 
+class DeviceWorker;
+
 class DeviceWorkerFactory {
  public:
   static std::string DeviceWorkerTypeList();

paddle/fluid/framework/device_worker_test.cc

@@ -13,9 +13,10 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/device_worker.h"
+
 #include <gtest/gtest.h>
+
 #include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/trainer.h"
 
 namespace paddle {
 namespace framework {

paddle/fluid/framework/distributed_strategy.proto

@@ -97,6 +97,7 @@ message AsyncConfig {
   optional int32 thread_pool_size = 6 [ default = 1 ];
   optional int32 send_wait_times = 7 [ default = 1 ];
   optional bool runtime_split_send_recv = 8 [ default = false ];
+  optional bool launch_barrier = 9 [ default = true ];
 }
 
 message PipelineConfig { optional int32 micro_batch = 1 [ default = 1 ]; }
@@ -127,6 +128,7 @@ message DistributedStrategy {
   optional int32 conv_workspace_size_limit = 22 [ default = 4000 ];
   optional bool cudnn_batchnorm_spatial_persistent = 23 [ default = true ];
   optional bool adaptive_localsgd = 24 [ default = false ];
+  optional bool fp16_allreduce = 25 [ default = false ];
 
   optional RecomputeConfig recompute_configs = 101;
   optional AMPConfig amp_configs = 102;

paddle/fluid/framework/dlpack_tensor.cc

@@ -11,10 +11,17 @@
 // WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 // See the License for the specific language governing permissions and
 // limitations under the License.
+#include "paddle/fluid/framework/dlpack_tensor.h"
 #include <unordered_map>
-
 #include "paddle/fluid/framework/data_type.h"
-#include "paddle/fluid/framework/dlpack_tensor.h"
+
+namespace paddle {
+namespace platform {
+struct bfloat16;
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/dlpack_tensor.h

@@ -15,11 +15,14 @@
 #pragma once
 
 #include <dlpack/dlpack.h>
+
 #include "paddle/fluid/framework/tensor.h"
 
 namespace paddle {
 namespace framework {
 
+class Tensor;
+
 class DLPackTensor {
  public:
   using LaneType = decltype(::DLTensor::dtype.lanes);  // uint16_t

paddle/fluid/framework/dlpack_tensor_test.cc

@@ -17,6 +17,12 @@
 #include <gtest/gtest.h>
 #include <vector>
 
+namespace paddle {
+namespace platform {
+struct float16;
+}  // namespace platform
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 

paddle/fluid/framework/downpour_worker.cc

@@ -13,10 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/device_worker_factory.h"
-#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include "paddle/fluid/platform/cpu_helper.h"
-#include "paddle/fluid/string/string_helper.h"
+
+namespace paddle {
+namespace framework {
+class LoDTensor;
+class Variable;
+}  // namespace framework
+}  // namespace paddle
 
 #if defined _WIN32 || defined __APPLE__
 #else

paddle/fluid/framework/downpour_worker_opt.cc

@@ -12,18 +12,16 @@ WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 See the License for the specific language governing permissions and
 limitations under the License. */
 
-#include <set>
-#include <unordered_map>
-#include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/device_worker_factory.h"
-#include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include "paddle/fluid/platform/cpu_helper.h"
-#include "paddle/fluid/platform/lodtensor_printer.h"
 
 namespace paddle {
 namespace framework {
 
+class OpDesc;
+class OperatorBase;
+class ProgramDesc;
+
 bool HasDependentOutput(const OpDesc& op_desc,
                         const std::unordered_set<std::string>& dependent_vars) {
   for (auto& var : op_desc.Outputs()) {

paddle/fluid/framework/eigen.h

@@ -14,6 +14,8 @@ limitations under the License. */
 
 #pragma once
 
+#include <stdint.h>
+
 #include "paddle/fluid/framework/tensor.h"
 #include "unsupported/Eigen/CXX11/Tensor"
 

paddle/fluid/framework/executor.h

@@ -19,6 +19,7 @@ limitations under the License. */
 #include <string>
 #include <unordered_map>
 #include <vector>
+
 #include "paddle/fluid/framework/data_set.h"
 #include "paddle/fluid/framework/executor_gc_helper.h"
 #include "paddle/fluid/framework/garbage_collector.h"
@@ -32,6 +33,11 @@ limitations under the License. */
 namespace paddle {
 namespace framework {
 
+class Dataset;
+class ProgramDesc;
+class Scope;
+class TrainerBase;
+
 struct ExecutorPrepareContext {
   ExecutorPrepareContext(const framework::ProgramDesc& prog, size_t block_id);
 

paddle/fluid/framework/executor_gc_helper.cc

@@ -13,16 +13,19 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/executor_gc_helper.h"
+
 #include <deque>
 #include <string>
-#include <unordered_map>
 #include <unordered_set>
 #include <utility>
-#include <vector>
+
 #include "glog/logging.h"
-#include "paddle/fluid/framework/lod_tensor.h"
-#include "paddle/fluid/framework/lod_tensor_array.h"
-#include "paddle/fluid/framework/selected_rows.h"
+#include "paddle/fluid/framework/block_desc.h"
+#include "paddle/fluid/framework/framework.pb.h"
+#include "paddle/fluid/framework/no_need_buffer_vars_inference.h"
+#include "paddle/fluid/framework/op_info.h"
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/framework/var_desc.h"
 #include "paddle/fluid/platform/enforce.h"
 
 namespace paddle {

paddle/fluid/framework/executor_gc_helper.h

@@ -18,6 +18,7 @@
 #include <string>
 #include <unordered_map>
 #include <vector>
+
 #include "paddle/fluid/framework/garbage_collector.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/scope.h"
@@ -26,6 +27,10 @@ namespace paddle {
 namespace framework {
 
 // Result map: op -> variable names that can be deleted after op runs
+class GarbageCollector;
+class OperatorBase;
+class Scope;
+
 std::unordered_map<const OperatorBase *, std::vector<std::string>>
 GetUnusedVars(const BlockDesc &block,
               const std::vector<std::unique_ptr<OperatorBase>> &ops,

paddle/fluid/framework/feed_fetch_method.cc

@@ -13,16 +13,17 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/feed_fetch_method.h"
+
 #include <string>
-#include <vector>
+
 #include "glog/logging.h"
-#include "paddle/fluid/framework/var_type.h"
-#include "paddle/fluid/framework/variable.h"
-#include "paddle/fluid/platform/place.h"
 
 namespace paddle {
 namespace framework {
 
+class LoDTensor;
+class Variable;
+
 void SetFeedVariable(Scope* scope, const LoDTensor& input,
                      const std::string& var_name, size_t index) {
   // If var_name Variable is not found in GlobalScope, a new variable will

paddle/fluid/framework/feed_fetch_method.h

@@ -15,12 +15,16 @@ limitations under the License. */
 #pragma once
 
 #include <string>
+
 #include "paddle/fluid/framework/feed_fetch_type.h"
 #include "paddle/fluid/framework/scope.h"
 
 namespace paddle {
 namespace framework {
 
+class LoDTensor;
+class Scope;
+
 void SetFeedVariable(Scope* scope, const LoDTensor& input,
                      const std::string& var_name, size_t index);
 

paddle/fluid/framework/fleet/fleet_wrapper.cc

@@ -29,12 +29,6 @@ limitations under the License. */
 #include "paddle/fluid/framework/fleet/fleet_wrapper.h"
 #include <algorithm>
 #include <utility>
-#include "paddle/fluid/framework/channel.h"
-#include "paddle/fluid/framework/data_feed.h"
-#include "paddle/fluid/framework/io/fs.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/timer.h"
 
 namespace paddle {
 namespace framework {
@@ -751,7 +745,57 @@ void FleetWrapper::PushDenseVarsAsync(
     push_sparse_status->push_back(std::move(status));
   }
 }
+#endif
+
+#ifdef PADDLE_WITH_XPU
+void FleetWrapper::PushDenseVarsAsync(
+    const Scope& scope, const uint64_t table_id,
+    const std::vector<std::string>& var_names,
+    std::vector<::std::future<int32_t>>* push_sparse_status,
+    float scale_datanorm, int batch_size,
+    const paddle::platform::Place& place) {
+#ifdef PADDLE_WITH_PSLIB
+  std::vector<paddle::ps::Region> regions;
+  for (auto& t : var_names) {
+    Variable* var = scope.FindVar(t);
+    LoDTensor* tensor = var->GetMutable<LoDTensor>();
+    int count = tensor->numel();
+    float* g_data = tensor->data<float>();
+
+    Variable* pin_var = scope.FindVar(t + "pin");
+    LoDTensor* pin_tensor = pin_var->GetMutable<LoDTensor>();
+    float* pin_g =
+        pin_tensor->mutable_data<float>(tensor->dims(), platform::CPUPlace());
+    memory::Copy(platform::CPUPlace(), pin_g,
+                 BOOST_GET_CONST(platform::XPUPlace, place), g_data,
+                 sizeof(float) * count);
+
+    float* g = pin_g;
+    if (scale_datanorm >= 0) {
+      if (t.find(".batch_size@GRAD") != std::string::npos ||
+          t.find(".batch_sum@GRAD") != std::string::npos) {
+        Eigen::Map<Eigen::MatrixXf> mat(g, 1, count);
+        float scale = 1.0 / batch_size;
+        mat *= scale;
+      } else if (t.find(".batch_square_sum@GRAD") != std::string::npos) {
+        VLOG(3) << "epsilon: " << scale_datanorm;
+        for (int i = 0; i < count; ++i) {
+          g[i] = (g[i] - batch_size * scale_datanorm) / batch_size +
+                 batch_size * scale_datanorm;
+        }
+      }
+    }
+    paddle::ps::Region reg(g, count);
+    regions.emplace_back(std::move(reg));
+  }
 
+  auto status = pslib_ptr_->_worker_ptr->push_dense(regions.data(),
+                                                    regions.size(), table_id);
+  if (push_sparse_status) {
+    push_sparse_status->push_back(std::move(status));
+  }
+#endif
+}
 #endif
 void FleetWrapper::PushDenseVarsAsync(
     const Scope& scope, const uint64_t table_id,

paddle/fluid/framework/fleet/fleet_wrapper.h

@@ -35,6 +35,12 @@ limitations under the License. */
 #include "paddle/fluid/framework/variable_helper.h"
 #include "paddle/fluid/platform/macros.h"  // for DISABLE_COPY_AND_ASSIGN
 
+namespace paddle {
+namespace framework {
+class Scope;
+}  // namespace framework
+}  // namespace paddle
+
 namespace paddle {
 namespace framework {
 
@@ -154,6 +160,14 @@ class FleetWrapper {
       float scale_datanorm, int batch_size,
       const paddle::platform::Place& place, cudaStream_t stream,
       cudaEvent_t event);
+#endif
+#ifdef PADDLE_WITH_XPU
+  void PushDenseVarsAsync(
+      const Scope& scope, const uint64_t table_id,
+      const std::vector<std::string>& var_names,
+      std::vector<::std::future<int32_t>>* push_sparse_status,
+      float scale_datanorm, int batch_size,
+      const paddle::platform::Place& place);
 #endif
   void PushDenseVarsAsync(
       const Scope& scope, const uint64_t table_id,

paddle/fluid/framework/fleet/gloo_wrapper.cc

@@ -10,10 +10,7 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/fleet/gloo_wrapper.h"
-#include <thread>  // NOLINT
-#include <vector>
 #include "paddle/fluid/framework/io/fs.h"
-#include "paddle/fluid/platform/errors.h"
 #include "paddle/fluid/string/string_helper.h"
 
 namespace gloo {

paddle/fluid/framework/fleet/heter_wrapper.cc

@@ -27,15 +27,6 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #include "paddle/fluid/framework/fleet/heter_wrapper.h"
-#include <algorithm>
-#include <utility>
-#include "paddle/fluid/framework/channel.h"
-#include "paddle/fluid/framework/data_feed.h"
-#include "paddle/fluid/framework/device_worker.h"
-#include "paddle/fluid/framework/io/fs.h"
-#include "paddle/fluid/framework/op_registry.h"
-#include "paddle/fluid/framework/scope.h"
-#include "paddle/fluid/platform/timer.h"
 #ifdef PADDLE_WITH_PSLIB
 
 namespace paddle {
@@ -122,29 +113,66 @@ void HeterWrapper::SerializeToReq(const std::string& varname, Scope* scope,
   if (platform::is_cpu_place(tensor->place())) {
     memcpy(data_ptr, tensor->data<void>(),
            tensor->numel() * SizeOfType(tensor->type()));
-  }
+  } else {
 #ifdef PADDLE_WITH_CUDA
-  else {
     memory::Copy(platform::CPUPlace(), data_ptr,
                  BOOST_GET_CONST(platform::CUDAPlace, tensor->place()),
                  tensor->data<void>(),
                  tensor->numel() * SizeOfType(tensor->type()), nullptr);
-  }
 #endif
+#ifdef PADDLE_WITH_XPU
+    memory::Copy(platform::CPUPlace(), data_ptr,
+                 BOOST_GET_CONST(platform::XPUPlace, tensor->place()),
+                 tensor->data<void>(),
+                 tensor->numel() * SizeOfType(tensor->type()));
+#endif
+  }
 }
 
-// void HeterWrapper::DeSerializeToTensor(Scope* scope,
-// const HeterRequest* request) {
 #ifdef PADDLE_WITH_CUDA
 void HeterWrapper::DeSerializeToTensor(Scope* scope,
                                        const VariableMessage& req_var,
                                        platform::Place place,
                                        cudaStream_t stream) {
+  // const VariableMessage& req_var = request->vars();
+  auto* var = scope->FindVar(req_var.varname());
+  auto* tensor = var->GetMutable<LoDTensor>();
+
+  std::vector<int> vec_dim;
+  for (auto& x : req_var.dims()) {
+    vec_dim.push_back(x);
+  }
+  tensor->Resize(make_ddim(vec_dim));
+
+  LoD lod;
+  for (int i = 0; i < req_var.lod_level(); ++i) {
+    framework::Vector<size_t> v;
+    for (int j = 0; j < req_var.lod(i).lod_data_size(); ++j) {
+      v.push_back(req_var.lod(i).lod_data(j));
+    }
+    lod.push_back(v);
+  }
+  tensor->set_lod(lod);
+
+  void* tensor_data =
+      tensor->mutable_data(place, ToVarType(req_var.data_type()));
+
+#ifdef PADDLE_WITH_CUDA
+  memory::Copy(BOOST_GET_CONST(platform::CUDAPlace, place), tensor_data,
+               platform::CPUPlace(), req_var.data().data(),
+               tensor->numel() * SizeOfType(tensor->type()), stream);
 #else
+  memcpy(tensor_data, req_var.data().data(),
+         tensor->numel() * SizeOfType(tensor->type()));
+#endif
+}
+#endif
+
+// void HeterWrapper::DeSerializeToTensor(Scope* scope,
+// const HeterRequest* request) {
 void HeterWrapper::DeSerializeToTensor(Scope* scope,
                                        const VariableMessage& req_var,
                                        platform::Place place) {
-#endif
   // const VariableMessage& req_var = request->vars();
   auto* var = scope->FindVar(req_var.varname());
   auto* tensor = var->GetMutable<LoDTensor>();
@@ -168,10 +196,10 @@ void HeterWrapper::DeSerializeToTensor(Scope* scope,
   void* tensor_data =
       tensor->mutable_data(place, ToVarType(req_var.data_type()));
 
-#ifdef PADDLE_WITH_CUDA
-  memory::Copy(BOOST_GET_CONST(platform::CUDAPlace, place), tensor_data,
+#ifdef PADDLE_WITH_XPU
+  memory::Copy(BOOST_GET_CONST(platform::XPUPlace, place), tensor_data,
                platform::CPUPlace(), req_var.data().data(),
-               tensor->numel() * SizeOfType(tensor->type()), stream);
+               tensor->numel() * SizeOfType(tensor->type()));
 #else
   memcpy(tensor_data, req_var.data().data(),
          tensor->numel() * SizeOfType(tensor->type()));
@@ -192,7 +220,8 @@ framework::proto::VarType::Type HeterWrapper::ToVarType(
     case VariableMessage::BOOL:
       return framework::proto::VarType::BOOL;  // NOLINT
     default:
-      VLOG(0) << "Not support type " << type;
+      PADDLE_THROW(platform::errors::InvalidArgument(
+          "ToVarType:Unsupported type %d", type));
   }
 }
 
@@ -239,7 +268,7 @@ void HeterWrapper::CallRemoteXpu(std::shared_ptr<HeterTask> task,
   request.set_cur_batch(task->cur_batch_);
 
   OnHeterRpcDone* done = new OnHeterRpcDone([this, task, worker](void* done) {
-    auto* closure = (OnHeterRpcDone*)done;
+    auto* closure = reinterpret_cast<OnHeterRpcDone*>(done);
     if (closure->cntl.Failed()) {
       VLOG(0) << "call xpu fail: " << closure->cntl.ErrorText();
     } else {

paddle/fluid/framework/fleet/nccl_wrapper.cc

@@ -13,9 +13,6 @@
 // limitations under the License.
 
 #include "paddle/fluid/framework/fleet/nccl_wrapper.h"
-#include <utility>
-#include "paddle/fluid/framework/data_feed.h"
-#include "paddle/fluid/framework/scope.h"
 
 namespace paddle {
 namespace framework {