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

cmake/inference_lib.cmake

@@ -70,6 +70,12 @@ copy(glog_lib
   DSTS ${dst_dir} ${dst_dir}/lib
 )
 
+set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/boost/")
+copy(boost_lib
+  SRCS ${BOOST_INCLUDE_DIR}/boost
+  DSTS ${dst_dir}
+)
+
 if(NOT PROTOBUF_FOUND)
     set(dst_dir "${CMAKE_INSTALL_PREFIX}/third_party/install/protobuf")
     copy(protobuf_lib

doc/fluid/design/dist_train/async_update.md

@@ -4,34 +4,37 @@
 
 For the typical synchronous distributed training, some significant steps are as follows:
 
-1. A Trainer will compute the gradients and SEND them to the Parameter Server(PServer) nodes.
-1. After the PServer node received gradients came from all the Trainers, It will aggregate the
+1. A trainer process will compute the gradients and **send** them to the parameter server (PS) nodes.
+1. After the PS node received gradients came from all the Trainers, It will aggregate the
 gradient variables for the same parameter into one gradient variable and then apply the aggregated
 gradient to the respective parameter, finally using an optimize algorithms(SGD, Monument...)
 to update the parameters.
-1. The Trainer would wait for the PServers finished the optimize stage, and GET the parameters from PServer,
+1. The Trainer would wait for the PS finished the optimize stage, and GET the parameters from PS,
 so all the Trainers would get the same parameters.
 
-In the synchronously distributed training, there should be a `Barrier` to synchronise the
-parameters after the optimizing stage. The performance of a distributed training job would
-depend on the slowest node if there were hundreds or thousands of training nodes in a
-Job, the performance of synchronously distributed training might be very poor because of
-the slow node. So this design doc would introduce an approach to implement
-*asynchronously* distributed training in PaddlePaddle Fluid.
+In Synchronous Distributed Training, there is a **barrier** on each PS to wait until all trainers processes
+have completed running current mini-batch. After that, all trainers can continue to run the next
+mini-batch. So, we can find that the overall performance of Synchronous Distributed Training depends 
+on the slowest node.
+
+In Asynchronous Distributed Training, we don't need to wait for a global mini-bach, the optimizer on
+the PS will run immediately when the gradient is uploaded to the PS from one trainer. This mode would
+train such models that achieve scaling, better throughput. In this design doc, we will introduce how to 
+implement the Asynchronous Distributed Training base on PaddlePaddle Fluid.
 
 ## Design
 
 <img src="./src/async_update.png" width="600"/>
 
-As the figure above, we describe a global view of asynchronously update process and use
+As the figure above, we describe a global view of the asynchronous update process and use
 the parameter `w1` as an example to introduce the steps:
 1. For each gradient variables, they may distribute on different GPU card and aggregate
 them while they are all calculated.
-1. Split the gradient variable into multiple blocks according to the number of PServer
+1. Split the gradient variable into multiple blocks according to the number of PS
 instances and then send them.
-1. PServer would run an `Optimize Block` using a specified optimize algorithm to update
+1. PS would run an `Optimize Block` using a specified optimize algorithm to update
 the specified parameter.
-1. The trainer will fetch latest parameter from PServer before running forward Op which depends
+1. The trainer will fetch the latest parameter from PS before running forward Op which depends
 on the specified parameter.
 1. Broadcast the received variable into multiple GPU cards and continue to run the next
 mini-batch.
@@ -40,8 +43,8 @@ mini-batch.
 
 - For the multiple devices distributed training, we need to aggregate the gradient
 variables which placed on different devices firstly and then schedule a `SendVars` Operator to
-send the gradient variables to the multiple PServer instances.
-- Schedule `FetchVars` operator to fetch the latest parameter from PServer before running
+send the gradient variables to the multiple PS instances.
+- Schedule `FetchVars` operator to fetch the latest parameter from PS before running
 the forward ops.
 - There could be a large number of gradient variables to be sent, so we need to use another
 thread pool(IO Threadpool) whose a number of the schedulable threads is larger than the

doc/v2/build_and_install/build_from_source_cn.rst

@@ -19,8 +19,9 @@
 ----------------
 
 PaddlePaddle需要使用Docker环境完成编译，这样可以免去单独安装编译依赖的步骤，可选的不同编译环境Docker镜像
-可以在 `这里 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 找到。或者
-参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。
+可以在 `这里 <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ 找到，您也可以
+在 `这里 <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`_ 找到 paddle_manylinux_devel
+镜像的编译以及使用方法。或者参考下述可选步骤，从源码中构建用于编译PaddlePaddle的Docker镜像。
 
 如果您选择不使用Docker镜像，则需要在本机安装下面章节列出的 `编译依赖`_ 之后才能开始编译的步骤。
 

doc/v2/build_and_install/build_from_source_en.rst

@@ -22,6 +22,8 @@ How To Build
 You need to use Docker to build PaddlePaddle
 to avoid installing dependencies by yourself. We have several pre-built
 Docker images `here <https://hub.docker.com/r/paddlepaddle/paddle_manylinux_devel/tags/>`_ ,
+you can also find how to build and use paddle_manylinux_devel Docker image from
+`here <https://github.com/PaddlePaddle/Paddle/tree/develop/tools/manylinux1/>`_
 Or you can build your own image from source as the optional step below:
 
 .. code-block:: bash

paddle/fluid/framework/CMakeLists.txt

@@ -5,11 +5,11 @@ proto_library(framework_proto SRCS framework.proto)
 cc_library(ddim SRCS ddim.cc DEPS eigen3 boost)
 cc_test(ddim_test SRCS ddim_test.cc DEPS ddim)
 nv_test(dim_test SRCS dim_test.cu DEPS ddim)
-
+cc_library(data_type SRCS data_type.cc DEPS framework_proto ddim device_context)
 if(WITH_GPU)
-  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS ddim place memory device_context framework_proto)
+  nv_library(tensor SRCS tensor.cc tensor_util.cu DEPS place memory data_type)
 else()
-  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS ddim place memory device_context framework_proto)
+  cc_library(tensor SRCS tensor.cc tensor_util.cc DEPS place memory data_type)
 endif()
 
 cc_test(tensor_test SRCS tensor_test.cc DEPS tensor)

paddle/fluid/framework/data_type.cc

+//   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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/data_type.h"
+#include <stdint.h>
+#include <string>
+#include <unordered_map>
+
+namespace paddle {
+namespace framework {
+
+struct DataTypeMap {
+  std::unordered_map<std::type_index, proto::VarType::Type> cpp_to_proto_;
+  std::unordered_map<int, std::type_index> proto_to_cpp_;
+  std::unordered_map<int, std::string> proto_to_str_;
+  std::unordered_map<std::type_index, size_t> cpp_to_size_;
+};
+
+static DataTypeMap* InitDataTypeMap();
+static DataTypeMap& gDataTypeMap() {
+  static DataTypeMap* g_data_type_map_ = InitDataTypeMap();
+  return *g_data_type_map_;
+}
+
+template <typename T>
+static inline void RegisterType(DataTypeMap* map,
+                                proto::VarType::Type proto_type,
+                                const std::string& name) {
+  map->proto_to_cpp_.emplace(static_cast<int>(proto_type), typeid(T));
+  map->cpp_to_proto_.emplace(typeid(T), proto_type);
+  map->proto_to_str_.emplace(static_cast<int>(proto_type), name);
+  map->cpp_to_size_.emplace(typeid(T), sizeof(T));
+}
+
+static DataTypeMap* InitDataTypeMap() {
+  auto retv = new DataTypeMap();
+
+#define RegType(cc_type, proto_type) \
+  RegisterType<cc_type>(retv, proto_type, #cc_type)
+
+  // NOTE: Add your customize type here.
+  RegType(platform::float16, proto::VarType::FP16);
+  RegType(float, proto::VarType::FP32);
+  RegType(double, proto::VarType::FP64);
+  RegType(int, proto::VarType::INT32);
+  RegType(int64_t, proto::VarType::INT64);
+  RegType(bool, proto::VarType::BOOL);
+  RegType(size_t, proto::VarType::SIZE_T);
+  RegType(int16_t, proto::VarType::INT16);
+
+#undef RegType
+  return retv;
+}
+
+proto::VarType::Type ToDataType(std::type_index type) {
+  auto it = gDataTypeMap().cpp_to_proto_.find(type);
+  if (it != gDataTypeMap().cpp_to_proto_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support %s as tensor type", type.name());
+}
+
+std::type_index ToTypeIndex(proto::VarType::Type type) {
+  auto it = gDataTypeMap().proto_to_cpp_.find(static_cast<int>(type));
+  if (it != gDataTypeMap().proto_to_cpp_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type",
+               static_cast<int>(type));
+}
+
+std::string DataTypeToString(const proto::VarType::Type type) {
+  auto it = gDataTypeMap().proto_to_str_.find(static_cast<int>(type));
+  if (it != gDataTypeMap().proto_to_str_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support proto::VarType::Type(%d) as tensor type",
+               static_cast<int>(type));
+}
+
+size_t SizeOfType(std::type_index type) {
+  auto it = gDataTypeMap().cpp_to_size_.find(type);
+  if (it != gDataTypeMap().cpp_to_size_.end()) {
+    return it->second;
+  }
+  PADDLE_THROW("Not support %s as tensor type", type.name());
+}
+
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/data_type.h

@@ -17,51 +17,14 @@ limitations under the License. */
 #include <typeindex>
 #include "paddle/fluid/framework/framework.pb.h"
 #include "paddle/fluid/platform/enforce.h"
+
 #include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace framework {
 
-inline proto::VarType::Type ToDataType(std::type_index type) {
-  if (typeid(platform::float16).hash_code() == type.hash_code()) {
-    return proto::VarType::FP16;
-  } else if (typeid(const float).hash_code() == type.hash_code()) {
-    // CPPLint complains Using C-style cast.  Use static_cast<float>() instead
-    // One fix to this is to replace float with const float because
-    // typeid(T) == typeid(const T)
-    // http://en.cppreference.com/w/cpp/language/typeid
-    return proto::VarType::FP32;
-  } else if (typeid(const double).hash_code() == type.hash_code()) {
-    return proto::VarType::FP64;
-  } else if (typeid(const int).hash_code() == type.hash_code()) {
-    return proto::VarType::INT32;
-  } else if (typeid(const int64_t).hash_code() == type.hash_code()) {
-    return proto::VarType::INT64;
-  } else if (typeid(const bool).hash_code() == type.hash_code()) {
-    return proto::VarType::BOOL;
-  } else {
-    PADDLE_THROW("Not supported");
-  }
-}
-
-inline std::type_index ToTypeIndex(proto::VarType::Type type) {
-  switch (type) {
-    case proto::VarType::FP16:
-      return typeid(platform::float16);
-    case proto::VarType::FP32:
-      return typeid(float);
-    case proto::VarType::FP64:
-      return typeid(double);
-    case proto::VarType::INT32:
-      return typeid(int);
-    case proto::VarType::INT64:
-      return typeid(int64_t);
-    case proto::VarType::BOOL:
-      return typeid(bool);
-    default:
-      PADDLE_THROW("Not support type %d", type);
-  }
-}
+extern proto::VarType::Type ToDataType(std::type_index type);
+extern std::type_index ToTypeIndex(proto::VarType::Type type);
 
 template <typename Visitor>
 inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
@@ -89,32 +52,12 @@ inline void VisitDataType(proto::VarType::Type type, Visitor visitor) {
   }
 }
 
-inline std::string DataTypeToString(const proto::VarType::Type type) {
-  switch (type) {
-    case proto::VarType::FP16:
-      return "float16";
-    case proto::VarType::FP32:
-      return "float32";
-    case proto::VarType::FP64:
-      return "float64";
-    case proto::VarType::INT16:
-      return "int16";
-    case proto::VarType::INT32:
-      return "int32";
-    case proto::VarType::INT64:
-      return "int64";
-    case proto::VarType::BOOL:
-      return "bool";
-    default:
-      PADDLE_THROW("Not support type %d", type);
-  }
-}
-
+extern std::string DataTypeToString(const proto::VarType::Type type);
+extern size_t SizeOfType(std::type_index type);
 inline std::ostream& operator<<(std::ostream& out,
                                 const proto::VarType::Type& type) {
   out << DataTypeToString(type);
   return out;
 }
-
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/details/build_strategy.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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.
+
+#pragma once
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+struct BuildStrategy {
+  enum class ReduceStrategy { kAllReduce = 0, kReduce = 1 };
+
+  enum class GradientScaleStrategy {
+    kCoeffNumDevice = 0,
+    kOne = 1,
+    kCustomized = 2,
+  };
+
+  ReduceStrategy reduce_{ReduceStrategy::kAllReduce};
+  GradientScaleStrategy gradient_scale_{GradientScaleStrategy::kCoeffNumDevice};
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle

paddle/fluid/framework/details/execution_strategy.h

+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// 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.
+
+#pragma once
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+struct ExecutionStrategy {
+  size_t num_threads_{0};
+  bool use_event_{true};
+  bool allow_op_delay_{false};
+};
+
+}  //  namespace details
+}  //  namespace framework
+}  //  namespace paddle

paddle/fluid/framework/details/multi_devices_graph_builder.cc

@@ -37,31 +37,26 @@ MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
     const std::vector<Scope *> &local_scopes,
-    platform::NCCLContextMap *nccl_ctxs, bool use_default_grad_scale,
-    bool balance_parameter_opt_between_cards)
+    platform::NCCLContextMap *nccl_ctxs, const BuildStrategy &strategy)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes),
       nccl_ctxs_(nccl_ctxs),
-      balance_parameter_opt_between_cards_(
-          balance_parameter_opt_between_cards) {
+      strategy_(strategy) {
 #else
 MultiDevSSAGraphBuilder::MultiDevSSAGraphBuilder(
     const std::vector<platform::Place> &places,
     const std::string &loss_var_name,
     const std::unordered_set<std::string> &params,
-    const std::vector<Scope *> &local_scopes, bool use_default_grad_scale,
-    bool balance_parameter_opt_between_cards)
+    const std::vector<Scope *> &local_scopes, const BuildStrategy &strategy)
     : loss_var_name_(loss_var_name),
       places_(places),
       local_scopes_(local_scopes),
-      balance_parameter_opt_between_cards_(
-          balance_parameter_opt_between_cards) {
+      strategy_(strategy) {
 #endif
   for (auto &p : params) {
     grad_names_.insert(GradVarName(p));
   }
-  use_default_grad_scale_ = use_default_grad_scale;
 }
 
 void MultiDevSSAGraphBuilder::CreateOpHandleIOs(SSAGraph *result,
@@ -146,7 +141,8 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
       CreateComputationalOps(&result, *op, 1);
     } else if (IsScaleLossOp(*op)) {
       // user can customize loss@grad if not use_default_grad_scale_
-      if (use_default_grad_scale_) {
+      if (strategy_.gradient_scale_ !=
+          BuildStrategy::GradientScaleStrategy::kCustomized) {
         CreateScaleLossGradOp(&result);
       }
       is_forwarding = false;
@@ -165,19 +161,22 @@ std::unique_ptr<SSAGraph> MultiDevSSAGraphBuilder::Build(
         // broadcast, and each gradient is only broadcast once.
         for (auto &og : op->OutputArgumentNames()) {
           if (IsParameterGradientOnce(og, &og_has_been_broadcast)) {
-            if (balance_parameter_opt_between_cards_) {
-              CreateReduceOp(&result, og, cur_device_id);
-              var_name_on_devices[cur_device_id].emplace(og);
-              bcast_var_name_set[cur_device_id].emplace(
-                  og.substr(0, og.size() - strlen(kGradVarSuffix)));
-              cur_device_id = (cur_device_id + 1) % places_.size();
-            } else {
-              if (IsSparseGradient(var_types, og)) {
-                CreateReduceOp(&result, og, 0);
-                CreateBroadcastOp(&result, og, 0);
-              } else {
-                InsertNCCLAllReduceOp(&result, og);
-              }
+            switch (strategy_.reduce_) {
+              case BuildStrategy::ReduceStrategy::kReduce:
+                CreateReduceOp(&result, og, cur_device_id);
+                var_name_on_devices[cur_device_id].emplace(og);
+                bcast_var_name_set[cur_device_id].emplace(
+                    og.substr(0, og.size() - strlen(kGradVarSuffix)));
+                cur_device_id = (cur_device_id + 1) % places_.size();
+                break;
+              case BuildStrategy::ReduceStrategy::kAllReduce:
+                if (IsSparseGradient(var_types, og)) {
+                  CreateReduceOp(&result, og, 0);
+                  CreateBroadcastOp(&result, og, 0);
+                } else {
+                  InsertNCCLAllReduceOp(&result, og);
+                }
+                break;
             }
           }
         }
@@ -303,7 +302,7 @@ bool MultiDevSSAGraphBuilder::IsParameterGradientOnce(
 int MultiDevSSAGraphBuilder::GetOpDeviceID(
     const std::vector<std::unordered_set<std::string>> &var_name_on_devices,
     const OpDesc &op) const {
-  if (!balance_parameter_opt_between_cards_) {
+  if (strategy_.reduce_ != BuildStrategy::ReduceStrategy::kReduce) {
     return -1;
   }
 

paddle/fluid/framework/details/multi_devices_graph_builder.h

@@ -17,6 +17,7 @@
 #include <utility>
 #include <vector>
 
+#include "paddle/fluid/framework/details/build_strategy.h"
 #include "paddle/fluid/framework/details/ssa_graph_builder.h"
 
 namespace paddle {
@@ -36,15 +37,13 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
                           platform::NCCLContextMap *nccl_ctxs,
-                          bool use_default_grad_scale,
-                          bool balance_parameter_opt_between_cards);
+                          const BuildStrategy &strategy);
 #else
   MultiDevSSAGraphBuilder(const std::vector<platform::Place> &places,
                           const std::string &loss_var_name,
                           const std::unordered_set<std::string> &params,
                           const std::vector<Scope *> &local_scopes,
-                          bool use_default_grad_scale,
-                          bool balance_parameter_opt_between_cards);
+                          const BuildStrategy &strategy);
 #endif
 
   std::unique_ptr<SSAGraph> Build(const ProgramDesc &program) const override;
@@ -62,8 +61,6 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
 #ifdef PADDLE_WITH_CUDA
   platform::NCCLContextMap *nccl_ctxs_;
 #endif
-  bool balance_parameter_opt_between_cards_;
-  bool use_default_grad_scale_;
 
   bool IsScaleLossOp(const OpDesc &op) const;
 
@@ -105,6 +102,9 @@ class MultiDevSSAGraphBuilder : public SSAGraphBuilder {
   bool IsSparseGradient(
       const std::unordered_map<std::string, proto::VarType::Type> &var_types,
       const std::string &og) const;
+
+ private:
+  BuildStrategy strategy_;
 };
 }  // namespace details
 }  // namespace framework

paddle/fluid/framework/details/threaded_ssa_graph_executor.cc

@@ -18,18 +18,17 @@ namespace paddle {
 namespace framework {
 namespace details {
 ThreadedSSAGraphExecutor::ThreadedSSAGraphExecutor(
-    size_t num_threads, bool use_event,
-    const std::vector<Scope *> &local_scopes,
+    const ExecutionStrategy &strategy, const std::vector<Scope *> &local_scopes,
     const std::vector<platform::Place> &places,
-    std::unique_ptr<SSAGraph> &&graph, bool allow_op_delay)
+    std::unique_ptr<SSAGraph> &&graph)
     : SSAGraphExecutor(std::move(graph)),
-      pool_(num_threads >= 2 ? new ::ThreadPool(num_threads) : nullptr),
+      pool_(strategy.num_threads_ >= 2 ? new ::ThreadPool(strategy.num_threads_)
+                                       : nullptr),
       local_scopes_(local_scopes),
       places_(places),
       fetch_ctxs_(places),
-      use_event_(use_event),
       running_ops_(0),
-      allow_op_delay_(allow_op_delay) {}
+      strategy_(strategy) {}
 
 FeedFetchList ThreadedSSAGraphExecutor::Run(
     const std::vector<std::string> &fetch_tensors) {
@@ -86,7 +85,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
     //
     // NOTE: DelayedOps have a lower priority. It will be scheduled after all
     // ready_ops have been performed.
-    if (ready_ops.empty() && allow_op_delay_ && running_ops_ == 0) {
+    if (ready_ops.empty() && strategy_.allow_op_delay_ && running_ops_ == 0) {
       run_all_ops(delayed_ops);
     } else {
       run_all_ops(ready_ops);
@@ -113,7 +112,7 @@ FeedFetchList ThreadedSSAGraphExecutor::Run(
         auto &deps = pending_ops[op];
         --deps;
         if (deps == 0) {
-          if (op->IsMultiDeviceTransfer() && allow_op_delay_) {
+          if (op->IsMultiDeviceTransfer() && strategy_.allow_op_delay_) {
             delayed_ops.insert(op);
           } else {
             ready_ops.insert(op);
@@ -191,7 +190,7 @@ void ThreadedSSAGraphExecutor::RunOp(
   auto op_run = [ready_var_q, op, this] {
     try {
       VLOG(10) << op << " " << op->Name() << " : " << op->DebugString();
-      op->Run(use_event_);
+      op->Run(strategy_.use_event_);
       VLOG(10) << op << " " << op->Name() << " Done ";
       running_ops_--;
       ready_var_q->Extend(op->Outputs());

paddle/fluid/framework/details/threaded_ssa_graph_executor.h

@@ -23,6 +23,7 @@
 #include <functional>
 #include "ThreadPool.h"  // ThreadPool in thrird party
 #include "paddle/fluid/framework/blocking_queue.h"
+#include "paddle/fluid/framework/details/execution_strategy.h"
 #include "paddle/fluid/framework/details/fetch_op_handle.h"
 #include "paddle/fluid/framework/details/ssa_graph_executor.h"
 
@@ -34,11 +35,10 @@ namespace details {
 
 class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
  public:
-  ThreadedSSAGraphExecutor(size_t num_threads, bool use_event,
+  ThreadedSSAGraphExecutor(const ExecutionStrategy &strategy,
                            const std::vector<Scope *> &local_scopes,
                            const std::vector<platform::Place> &places,
-                           std::unique_ptr<SSAGraph> &&graph,
-                           bool allow_op_delay);
+                           std::unique_ptr<SSAGraph> &&graph);
 
   // Run a SSAGraph by a thread pool
   // Use topological sort algorithm
@@ -55,10 +55,8 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
   std::vector<Scope *> local_scopes_;
   std::vector<platform::Place> places_;
   platform::DeviceContextPool fetch_ctxs_;
-  const bool use_event_;
   std::unique_ptr<platform::EnforceNotMet> exception_;
   std::atomic<int> running_ops_;
-  bool allow_op_delay_;
 
   void InsertPendingOp(std::unordered_map<OpHandleBase *, size_t> *pending_ops,
                        OpHandleBase *op_instance) const;
@@ -74,6 +72,9 @@ class ThreadedSSAGraphExecutor : public SSAGraphExecutor {
       std::unordered_map<OpHandleBase *, size_t> *pending_ops,
       std::unordered_set<VarHandleBase *> *pending_vars,
       BlockingQueue<VarHandleBase *> *ready_vars, FeedFetchList *fetch_data);
+
+ private:
+  ExecutionStrategy strategy_;
 };
 
 }  // namespace details

paddle/fluid/framework/framework.proto

@@ -101,6 +101,8 @@ message VarType {
     FP16 = 4;
     FP32 = 5;
     FP64 = 6;
+    // Tensor<size_t> is used in C++.
+    SIZE_T = 19;
 
     // Other types that may need additional descriptions
     LOD_TENSOR = 7;

paddle/fluid/framework/op_kernel_type_test.cc

@@ -27,7 +27,7 @@ TEST(OpKernelType, ToString) {
                               LibraryType::kCUDNN);
 
   ASSERT_EQ(paddle::framework::KernelTypeToString(op_kernel_type),
-            "data_type[float32]:data_layout[NCHW]:place[CPUPlace]:library_type["
+            "data_type[float]:data_layout[NCHW]:place[CPUPlace]:library_type["
             "CUDNN]");
 }
 

paddle/fluid/framework/parallel_executor.cc

@@ -52,13 +52,12 @@ std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
 }
 
 ParallelExecutor::ParallelExecutor(
-    size_t num_threads, bool use_event,
     const std::vector<platform::Place> &places,
     const std::unordered_set<std::string> &params,
     const std::unordered_set<std::string> &bcast_vars,
     const ProgramDesc &main_program, const std::string &loss_var_name,
-    Scope *scope, const std::vector<Scope *> &local_scopes, bool allow_op_delay,
-    bool use_default_grad_scale, bool balance_parameter_opt_between_cards,
+    Scope *scope, const std::vector<Scope *> &local_scopes,
+    const ExecutionStrategy &exec_strategy, const BuildStrategy &build_strategy,
     size_t num_trainers, size_t trainer_id)
     : member_(new ParallelExecutorPrivate(places)) {
   member_->global_scope_ = scope;
@@ -100,18 +99,16 @@ ParallelExecutor::ParallelExecutor(
 #ifdef PADDLE_WITH_CUDA
   details::MultiDevSSAGraphBuilder builder(
       member_->places_, loss_var_name, params, member_->local_scopes_,
-      member_->nccl_ctxs_.get(), use_default_grad_scale,
-      balance_parameter_opt_between_cards);
+      member_->nccl_ctxs_.get(), build_strategy);
 #else
-  details::MultiDevSSAGraphBuilder builder(
-      member_->places_, loss_var_name, params, member_->local_scopes_,
-      use_default_grad_scale, balance_parameter_opt_between_cards);
+  details::MultiDevSSAGraphBuilder builder(member_->places_, loss_var_name,
+                                           params, member_->local_scopes_,
+                                           build_strategy);
 #endif
   auto graph = builder.Build(main_program);
 
   member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
-      num_threads, use_event, member_->local_scopes_, places, std::move(graph),
-      allow_op_delay));
+      exec_strategy, member_->local_scopes_, places, std::move(graph)));
 
   // Step 3. Create vars in each scope;
   for (auto *var : main_program.Block(0).AllVars()) {

paddle/fluid/framework/parallel_executor.h

@@ -14,57 +14,60 @@ limitations under the License. */
 
 #pragma once
 
+#include <paddle/fluid/framework/details/build_strategy.h>
 #include <string>
 #include <unordered_set>
 #include <vector>
+#include "paddle/fluid/framework/details/execution_strategy.h"
 #include "paddle/fluid/framework/executor.h"
 #include "paddle/fluid/framework/op_info.h"
 #include "paddle/fluid/framework/program_desc.h"
 #include "paddle/fluid/framework/scope.h"
 #include "paddle/fluid/framework/tensor.h"
 #include "paddle/fluid/platform/device_context.h"
-
 namespace paddle {
 namespace framework {
 
 class ParallelExecutorPrivate;
 
+using details::BuildStrategy;
+using details::ExecutionStrategy;
+
 class ParallelExecutor {
   DISABLE_COPY_AND_ASSIGN(ParallelExecutor);
 
  public:
-  explicit ParallelExecutor(size_t num_threads, bool use_event,
-                            const std::vector<platform::Place>& places,
-                            const std::unordered_set<std::string>& params,
-                            const std::unordered_set<std::string>& bcast_vars,
-                            const ProgramDesc& main_program,
-                            const std::string& loss_var_name, Scope* scope,
-                            const std::vector<Scope*>& local_scopes,
-                            bool allow_op_delay, bool use_default_grad_scale,
-                            bool balance_parameter_opt_between_cards,
+  explicit ParallelExecutor(const std::vector<platform::Place> &places,
+                            const std::unordered_set<std::string> &params,
+                            const std::unordered_set<std::string> &bcast_vars,
+                            const ProgramDesc &main_program,
+                            const std::string &loss_var_name, Scope *scope,
+                            const std::vector<Scope *> &local_scopes,
+                            const ExecutionStrategy &exec_strategy,
+                            const BuildStrategy &build_strategy,
                             size_t num_trainers = 1, size_t trainer_id = 0);
 
   ~ParallelExecutor();
 
-  std::vector<Scope*>& GetLocalScopes();
+  std::vector<Scope *> &GetLocalScopes();
 
   /**
    * Feed tensors to local scopes. The size of tensors should be equal to the
    * size of local scopes.
    */
   void FeedTensorsIntoLocalScopes(
-      const std::vector<std::unordered_map<std::string, LoDTensor>>& tensors);
+      const std::vector<std::unordered_map<std::string, LoDTensor>> &tensors);
 
   void FeedAndSplitTensorIntoLocalScopes(
-      const std::unordered_map<std::string, LoDTensor>& tensors);
+      const std::unordered_map<std::string, LoDTensor> &tensors);
 
-  void Run(const std::vector<std::string>& fetch_tensors,
-           const std::string& fetched_var_name);
+  void Run(const std::vector<std::string> &fetch_tensors,
+           const std::string &fetched_var_name);
 
-  void BCastParamsToGPUs(const std::unordered_set<std::string>& vars) const;
+  void BCastParamsToGPUs(const std::unordered_set<std::string> &vars) const;
 
  private:
-  ParallelExecutorPrivate* member_;
+  ParallelExecutorPrivate *member_;
 };
 
 }  // namespace framework

paddle/fluid/framework/tensor_impl.h

@@ -13,54 +13,14 @@ See the License for the specific language governing permissions and
 limitations under the License. */
 
 #pragma once
+#include "paddle/fluid/framework/data_type.h"
 #include "paddle/fluid/memory/memcpy.h"
 #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/platform/float16.h"
 
 namespace paddle {
 namespace framework {
-
-template <typename... T>
-struct SizeOfTypeFunctor;
-
-template <typename T>
-struct SizeOfTypeFunctor<T> {
-  size_t operator()(std::type_index type) const {
-    if (typeid(T).hash_code() == type.hash_code()) {
-      return sizeof(T);
-    } else {
-      return 0UL;
-    }
-  }
-};
-
-template <>
-struct SizeOfTypeFunctor<> {
-  size_t operator()(std::type_index type) const { return 0UL; }
-};
-
-template <typename HEAD, typename... TAIL>
-struct SizeOfTypeFunctor<HEAD, TAIL...> {
-  size_t operator()(std::type_index type) const {
-    SizeOfTypeFunctor<HEAD> head;
-    size_t head_size = head(type);
-    if (head_size != 0) {
-      return head_size;
-    }
-    SizeOfTypeFunctor<TAIL...> tail;
-    return tail(type);
-  }
-};
-
-static inline size_t SizeOfType(std::type_index type) {
-  SizeOfTypeFunctor<int, float, double, int16_t, int64_t, bool, size_t,
-                    platform::float16>
-      functor;
-  size_t size = functor(type);
-  PADDLE_ENFORCE(size != 0UL, "Cannot get size of type %s", type.name());
-  return size;
-}
-
+extern size_t SizeOfType(std::type_index type);
 inline void Tensor::check_memory_size() const {
   PADDLE_ENFORCE_NOT_NULL(
       holder_, "Tensor holds no memory. Call Tensor::mutable_data first.");

paddle/fluid/inference/analysis/device.h

@@ -11,6 +11,7 @@ distributed under the License is distributed on an "AS IS" BASIS,
 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. */
+#pragma once
 
 namespace paddle {
 namespace inference {

paddle/fluid/inference/analysis/dot.h

@@ -21,6 +21,7 @@
 
 #include <glog/logging.h>
 #include <sstream>
+#include <string>
 #include <unordered_map>
 #include <vector>
 

paddle/fluid/inference/analysis/node.h

@@ -19,6 +19,7 @@ limitations under the License. */
  */
 #pragma once
 
+#include <limits>
 #include <memory>
 #include <string>
 #include <unordered_map>

paddle/fluid/inference/engine.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include "paddle/fluid/framework/framework.pb.h"
 
 namespace paddle {
@@ -58,8 +59,8 @@ class EngineBase {
 
 struct Buffer {
   void* buffer{nullptr};               // buffer should be allocated only once.
-  int max_size;                        // buffer allocated space.
-  int size;                            // data size.
+  size_t max_size;                     // buffer allocated space.
+  size_t size;                         // data size.
   DeviceType device{DeviceType::UNK};  // tells which device this buffer is on.
 };
 

paddle/fluid/inference/tensorrt/CMakeLists.txt

 nv_library(tensorrt_engine SRCS engine.cc DEPS framework_proto)
 nv_test(test_tensorrt SRCS test_tensorrt.cc DEPS dynload_cuda device_context dynamic_loader)
 nv_test(test_tensorrt_engine SRCS test_engine.cc DEPS dynload_cuda tensorrt_engine)
-
 add_subdirectory(convert)

paddle/fluid/inference/tensorrt/convert/CMakeLists.txt

-nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc op_converter.h DEPS ${FLUID_CORE_MODULES})
-nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc
+nv_test(test_op_converter SRCS test_op_converter.cc mul_op.cc conv2d_op.cc DEPS ${FLUID_CORE_MODULES})
+nv_test(test_trt_activation_op SRCS test_activation_op.cc activation_op.cc io_converter.cc
   DEPS ${FLUID_CORE_MODULES} activation_op tensorrt_engine)
 nv_test(test_io_converter SRCS test_io_converter.cc io_converter.cc DEPS dynload_cuda dynamic_loader lod_tensor)

paddle/fluid/inference/tensorrt/convert/io_converter.cc

@@ -23,26 +23,42 @@ namespace tensorrt {
 using platform::is_gpu_place;
 using platform::is_cpu_place;
 
-class DefaultInputConverter : public EngineInputConverter {
+class DefaultIOConverter : public EngineIOConverter {
  public:
-  DefaultInputConverter() {}
+  DefaultIOConverter() {}
   // NOTE out is GPU memory.
   virtual void operator()(const LoDTensor& in, void* out,
                           size_t max_size) override {
     PADDLE_ENFORCE(out != nullptr);
-    PADDLE_ENFORCE_LE(in.memory_size(), max_size);
+    PADDLE_ENFORCE(stream_ != nullptr);
     const auto& place = in.place();
+    size_t size = in.memory_size();
+    PADDLE_ENFORCE_LE(size, max_size);
     if (is_cpu_place(place)) {
-      PADDLE_ENFORCE(stream_ != nullptr);
-      PADDLE_ENFORCE_EQ(0,
-                        cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
-                                        cudaMemcpyHostToDevice, *stream_));
-
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
+                                           cudaMemcpyHostToDevice, *stream_));
     } else if (is_gpu_place(place)) {
-      PADDLE_ENFORCE_EQ(0,
-                        cudaMemcpyAsync(out, in.data<float>(), in.memory_size(),
-                                        cudaMemcpyHostToHost, *stream_));
-
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out, in.data<float>(), size,
+                                           cudaMemcpyDeviceToDevice, *stream_));
+    } else {
+      PADDLE_THROW("Unknown device for converter");
+    }
+    cudaStreamSynchronize(*stream_);
+  }
+  // NOTE in is GPU memory.
+  virtual void operator()(const void* in, LoDTensor* out,
+                          size_t max_size) override {
+    PADDLE_ENFORCE(in != nullptr);
+    PADDLE_ENFORCE(stream_ != nullptr);
+    const auto& place = out->place();
+    size_t size = out->memory_size();
+    PADDLE_ENFORCE_LE(size, max_size);
+    if (is_cpu_place(place)) {
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
+                                           cudaMemcpyDeviceToHost, *stream_));
+    } else if (is_gpu_place(place)) {
+      PADDLE_ENFORCE_EQ(0, cudaMemcpyAsync(out->data<float>(), in, size,
+                                           cudaMemcpyDeviceToDevice, *stream_));
     } else {
       PADDLE_THROW("Unknown device for converter");
     }
@@ -50,7 +66,8 @@ class DefaultInputConverter : public EngineInputConverter {
   }
 };
 
-REGISTER_TENSORRT_INPUT_CONVERTER(default, DefaultInputConverter);
+// fluid LodTensor <-> tensorrt ITensor
+REGISTER_TENSORRT_IO_CONVERTER(default, DefaultIOConverter);
 
 }  // namespace tensorrt
 }  // namespace inference

paddle/fluid/inference/tensorrt/convert/io_converter.h

@@ -14,6 +14,7 @@ limitations under the License. */
 
 #pragma once
 
+#include <string>
 #include <unordered_map>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/utils/singleton.h"
@@ -25,43 +26,57 @@ namespace tensorrt {
 using framework::LoDTensor;
 
 /*
- * Convert Input from Fluid to an Engine.
- * TensorRT's ITensor follows row major, NCHW. Fluid is also row major, so in
- * most cases just need to copy the data.
+ * Convert Input from Fluid to TensorRT Engine.
+ * Convert Output from TensorRT Engine to Fluid.
+ *
+ * Note that TensorRT's ITensor follows row major, NCHW. Fluid is also row
+ * major,
+ * so in the default case just need to copy the data.
  */
-class EngineInputConverter {
+class EngineIOConverter {
  public:
-  EngineInputConverter() {}
+  EngineIOConverter() {}
 
   virtual void operator()(const LoDTensor& in, void* out, size_t max_size) {}
+  virtual void operator()(const void* in, LoDTensor* out, size_t max_size) {}
 
   void SetStream(cudaStream_t* stream) { stream_ = stream; }
 
-  static void Run(const std::string& in_op_type, const LoDTensor& in, void* out,
-                  size_t max_size, cudaStream_t* stream) {
+  static void ConvertInput(const std::string& op_type, const LoDTensor& in,
+                           void* out, size_t max_size, cudaStream_t* stream) {
     PADDLE_ENFORCE(stream != nullptr);
-    auto* converter = Registry<EngineInputConverter>::Lookup(
-        in_op_type, "default" /* default_type */);
+    auto* converter = Registry<EngineIOConverter>::Lookup(
+        op_type, "default" /* default_type */);
     PADDLE_ENFORCE_NOT_NULL(converter);
     converter->SetStream(stream);
     (*converter)(in, out, max_size);
   }
 
-  virtual ~EngineInputConverter() {}
+  static void ConvertOutput(const std::string& op_type, const void* in,
+                            LoDTensor* out, size_t max_size,
+                            cudaStream_t* stream) {
+    PADDLE_ENFORCE(stream != nullptr);
+    auto* converter = Registry<EngineIOConverter>::Lookup(
+        op_type, "default" /* default_type */);
+    PADDLE_ENFORCE_NOT_NULL(converter);
+    converter->SetStream(stream);
+    (*converter)(in, out, max_size);
+  }
+
+  virtual ~EngineIOConverter() {}
 
  protected:
   cudaStream_t* stream_{nullptr};
 };
 
+#define REGISTER_TENSORRT_IO_CONVERTER(op_type__, Converter__)        \
+  struct trt_io_##op_type__##_converter {                             \
+    trt_io_##op_type__##_converter() {                                \
+      Registry<EngineIOConverter>::Register<Converter__>(#op_type__); \
+    }                                                                 \
+  };                                                                  \
+  trt_io_##op_type__##_converter trt_io_##op_type__##_converter__;
+
 }  // namespace tensorrt
 }  // namespace inference
 }  // namespace paddle
-
-#define REGISTER_TENSORRT_INPUT_CONVERTER(in_op_type__, Converter__) \
-  struct trt_input_##in_op_type__##_converter {                      \
-    trt_input_##in_op_type__##_converter() {                         \
-      ::paddle::inference::Registry<EngineInputConverter>::Register< \
-          Converter__>(#in_op_type__);                               \
-    }                                                                \
-  };                                                                 \
-  trt_input_##in_op_type__##_converter trt_input_##in_op_type__##_converter__;

paddle/fluid/inference/tensorrt/convert/test_activation_op.cc

@@ -16,6 +16,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/program_desc.h"
+#include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
 #include "paddle/fluid/inference/tensorrt/convert/op_converter.h"
 #include "paddle/fluid/platform/device_context.h"
 #include "paddle/fluid/platform/place.h"
@@ -26,7 +27,7 @@ namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-void Compare(float input, float expect) {
+void Compare(const std::string op_type, float input, float expect) {
   framework::Scope scope;
   platform::CUDAPlace place;
   platform::CUDADeviceContext ctx(place);
@@ -35,6 +36,7 @@ void Compare(float input, float expect) {
   auto x_var = scope.Var("X");
   auto x_tensor = x_var->GetMutable<framework::LoDTensor>();
   x_tensor->Resize({1, 1});
+  x_tensor->mutable_data<float>(place);
   std::vector<float> init;
   init.push_back(input);
   framework::TensorFromVector(init, ctx, x_tensor);
@@ -45,14 +47,15 @@ void Compare(float input, float expect) {
   out_tensor->mutable_data<float>(place);
 
   framework::OpDesc op_desc;
-  op_desc.SetType("relu");
+  op_desc.SetType(op_type);
   op_desc.SetInput("X", {"X"});
   op_desc.SetOutput("Out", {"Out"});
 
-  auto relu_op = framework::OpRegistry::CreateOp(*op_desc.Proto());
+  auto op = framework::OpRegistry::CreateOp(*op_desc.Proto());
 
   // run fluid op
-  relu_op->Run(scope, place);
+  op->Run(scope, place);
+  // get fluid output
   std::vector<float> out1;
   framework::TensorToVector(*out_tensor, ctx, &out1);
 
@@ -63,21 +66,28 @@ void Compare(float input, float expect) {
   engine->InitNetwork();
   engine->DeclareInput("X", nvinfer1::DataType::kFLOAT,
                        nvinfer1::DimsCHW{1, 1, 1});
-
+  // convert op
   OpConverter op_converter;
   op_converter.ConvertOp(*op_desc.Proto(), engine);
 
   engine->DeclareOutput("Out");
   engine->FreezeNetwork();
-  engine->SetInputFromCPU("X", &input, 1 * sizeof(float));
 
-  // run tensorrt op
+  // convert LoDTensor to ITensor
+  size_t size = x_tensor->memory_size();
+  EngineIOConverter::ConvertInput(op_type, *x_tensor,
+                                  engine->buffer("X").buffer, size, &stream);
+  // run tensorrt Outp
   engine->Execute(1);
-
-  float out2;
-  engine->GetOutputInCPU("Out", &out2, 1 * sizeof(float));
-
-  ASSERT_EQ(out1[0], out2);
+  // convert ITensor to LoDTensor
+  EngineIOConverter::ConvertOutput(op_type, engine->buffer("Out").buffer,
+                                   out_tensor, size, &stream);
+  // get tensorrt output
+  std::vector<float> out2;
+  framework::TensorToVector(*out_tensor, ctx, &out2);
+
+  // compare
+  ASSERT_EQ(out1[0], out2[0]);
   ASSERT_EQ(out1[0], expect);
 
   delete engine;
@@ -85,8 +95,8 @@ void Compare(float input, float expect) {
 }
 
 TEST(OpConverter, ConvertRelu) {
-  Compare(1, 1);   // relu(1) = 1
-  Compare(-5, 0);  // relu(-5) = 0
+  Compare("relu", 1, 1);   // relu(1) = 1
+  Compare("relu", -5, 0);  // relu(-5) = 0
 }
 
 }  // namespace tensorrt

paddle/fluid/inference/tensorrt/convert/test_io_converter.cc

@@ -12,40 +12,63 @@ 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 <gtest/gtest.h>
 #include "paddle/fluid/framework/lod_tensor.h"
 #include "paddle/fluid/inference/tensorrt/convert/io_converter.h"
 
-#include <gtest/gtest.h>
-
 namespace paddle {
 namespace inference {
 namespace tensorrt {
 
-class EngineInputConverterTester : public ::testing::Test {
- public:
-  void SetUp() override { tensor.Resize({10, 10}); }
+void IOConverterTester(const platform::DeviceContext& ctx) {
+  cudaStream_t stream;
+  ASSERT_EQ(0, cudaStreamCreate(&stream));
 
-  framework::LoDTensor tensor;
-};
+  // init fluid in_tensor
+  framework::LoDTensor in_tensor;
+  in_tensor.Resize({10, 10});
+  auto place = ctx.GetPlace();
+  in_tensor.mutable_data<float>(place);
+  std::vector<float> init;
+  for (int64_t i = 0; i < 10 * 10; ++i) {
+    init.push_back(i);
+  }
+  framework::TensorFromVector(init, ctx, &in_tensor);
 
-TEST_F(EngineInputConverterTester, DefaultCPU) {
+  // init tensorrt buffer
   void* buffer;
-  tensor.mutable_data<float>(platform::CPUPlace());
-  ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
+  size_t size = in_tensor.memory_size();
+  ASSERT_EQ(cudaMalloc(&buffer, size), 0);
 
-  cudaStream_t stream;
-  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
-                            &stream);
+  // convert fluid in_tensor to tensorrt buffer
+  EngineIOConverter::ConvertInput("test", in_tensor, buffer, size, &stream);
+
+  // convert tensorrt buffer to fluid out_tensor
+  framework::LoDTensor out_tensor;
+  out_tensor.Resize({10, 10});
+  out_tensor.mutable_data<float>(place);
+  EngineIOConverter::ConvertOutput("test", buffer, &out_tensor, size, &stream);
+
+  // compare in_tensor and out_tensor
+  std::vector<float> result;
+  framework::TensorToVector(out_tensor, ctx, &result);
+  EXPECT_EQ(init.size(), result.size());
+  for (size_t i = 0; i < init.size(); i++) {
+    EXPECT_EQ(init[i], result[i]);
+  }
+  cudaStreamDestroy(stream);
 }
 
-TEST_F(EngineInputConverterTester, DefaultGPU) {
-  void* buffer;
-  tensor.mutable_data<float>(platform::CUDAPlace());
-  ASSERT_EQ(cudaMalloc(&buffer, tensor.memory_size()), 0);
+TEST(EngineIOConverterTester, DefaultCPU) {
+  platform::CPUPlace place;
+  platform::CPUDeviceContext ctx(place);
+  IOConverterTester(ctx);
+}
 
-  cudaStream_t stream;
-  EngineInputConverter::Run("test", tensor, buffer, tensor.memory_size(),
-                            &stream);
+TEST(EngineIOConverterTester, DefaultGPU) {
+  platform::CUDAPlace place;
+  platform::CUDADeviceContext ctx(place);
+  IOConverterTester(ctx);
 }
 
 }  // namespace tensorrt

paddle/fluid/operators/detail/grpc_server.cc

@@ -306,7 +306,7 @@ void AsyncGRPCServer::TryToRegisterNewPrefetchOne() {
   }
   RequestPrefetch* prefetch =
       new RequestPrefetch(&service_, cq_prefetch_.get(), sync_mode_, scope_,
-                          dev_ctx_, executor_, program_, prefetch_ctx_);
+                          dev_ctx_, executor_, program_, prefetch_ctx_.get());
 
   VLOG(4) << "Create RequestPrefetch status:" << prefetch->Status();
 }

paddle/fluid/operators/detail/grpc_server.h

@@ -64,8 +64,9 @@ class AsyncGRPCServer final {
 
   void SetExecutor(framework::Executor *executor) { executor_ = executor; }
 
-  void SetPrefetchPreparedCtx(framework::ExecutorPrepareContext *prepared) {
-    prefetch_ctx_ = prepared;
+  void SetPrefetchPreparedCtx(
+      std::unique_ptr<framework::ExecutorPrepareContext> prepared) {
+    prefetch_ctx_.reset(prepared.release());
   }
 
   int GetSelectedPort() const { return selected_port_; }
@@ -116,7 +117,7 @@ class AsyncGRPCServer final {
   std::unique_ptr<std::thread> t_get_;
   std::unique_ptr<std::thread> t_prefetch_;
 
-  framework::ExecutorPrepareContext *prefetch_ctx_;
+  std::unique_ptr<framework::ExecutorPrepareContext> prefetch_ctx_;
   framework::ProgramDesc *program_;
   framework::Executor *executor_;
   int selected_port_;

paddle/fluid/operators/detail/grpc_server_test.cc

@@ -100,7 +100,7 @@ void StartServer(const std::string& endpoint) {
   InitTensorsOnServer(&scope, &place, 10);
 
   rpc_service_->SetProgram(&program);
-  rpc_service_->SetPrefetchPreparedCtx(prepared.get());
+  rpc_service_->SetPrefetchPreparedCtx(std::move(prepared));
   rpc_service_->SetDevCtx(&ctx);
   rpc_service_->SetScope(&scope);
   rpc_service_->SetExecutor(&exe);

paddle/fluid/operators/listen_and_serv_op.cc

@@ -322,8 +322,7 @@ void ListenAndServOp::RunImpl(const framework::Scope &scope,
   // prepare for prefetch
   VLOG(3) << "prefetch block id is " << prefetch_block->ID();
   auto prefetch_prepared = executor.Prepare(*program, prefetch_block->ID());
-  rpc_service_->SetPrefetchPreparedCtx(prefetch_prepared.get());
-  prefetch_prepared.release();
+  rpc_service_->SetPrefetchPreparedCtx(std::move(prefetch_prepared));
 
   // start the server listening after all member initialized.
   server_thread_.reset(new std::thread(RunServer, rpc_service_));

paddle/fluid/operators/load_combine_op.cc

@@ -12,7 +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 <fstream>
-
+#include "paddle/fluid/framework/data_type_transform.h"
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/platform/device_context.h"
 
@@ -31,6 +31,7 @@ class LoadCombineOp : public framework::OperatorBase {
   void RunImpl(const framework::Scope &scope,
                const platform::Place &place) const override {
     auto filename = Attr<std::string>("file_path");
+    auto load_as_fp16 = Attr<bool>("load_as_fp16");
 
     std::ifstream fin(filename);
     PADDLE_ENFORCE(static_cast<bool>(fin),
@@ -59,17 +60,25 @@ class LoadCombineOp : public framework::OperatorBase {
       // Get data from fin to tensor
       DeserializeFromStream(fin, tensor, dev_ctx);
 
-      if (platform::is_gpu_place(place)) {
-        // copy CPU to GPU
-        framework::LoDTensor cpu_tensor;
-        cpu_tensor.ShareDataWith(*tensor);
-        cpu_tensor.set_lod(tensor->lod());
-
-        // reset tensor
+      auto in_dtype = framework::ToDataType(tensor->type());
+      auto out_dtype =
+          load_as_fp16 ? framework::proto::VarType::FP16 : in_dtype;
+
+      if (in_dtype != out_dtype) {
+        // convert to float16 tensor
+        auto in_kernel_type = framework::OpKernelType(in_dtype, place);
+        auto out_kernel_type = framework::OpKernelType(out_dtype, place);
+        framework::LoDTensor fp16_tensor;
+        // copy LoD info to the new tensor
+        fp16_tensor.set_lod(tensor->lod());
+        framework::TransDataType(in_kernel_type, out_kernel_type, *tensor,
+                                 &fp16_tensor);
+
+        // reset output tensor
         out_var->Clear();
         tensor = out_var->GetMutable<framework::LoDTensor>();
-        tensor->set_lod(cpu_tensor.lod());
-        TensorCopy(cpu_tensor, place, dev_ctx, tensor);
+        tensor->set_lod(fp16_tensor.lod());
+        tensor->ShareDataWith(fp16_tensor);
       }
     }
   }
@@ -82,6 +91,13 @@ class LoadCombineOpProtoMaker : public framework::OpProtoAndCheckerMaker {
         "Out",
         "(vector) The output LoDTensors that will be read from the input file.")
         .AsDuplicable();
+    AddAttr<bool>(
+        "load_as_fp16",
+        "(boolean, default false)"
+        "If true, the tensor will be first loaded and then "
+        "converted to float16 data type. Otherwise, the tensor will be "
+        "directly loaded without data type conversion.")
+        .SetDefault(false);
     AddAttr<std::string>("file_path",
                          "(string) "
                          "LoDTensors will be loaded from \"file_path\".")

paddle/fluid/operators/save_load_combine_op_test.cc

@@ -139,8 +139,9 @@ TEST(SaveLoadCombineOp, CPU) {
   CheckValues<int, int>(expect4, actual4, expect_lod4, actual_lod4, numel4);
 }
 
-// FP16 version of SaveLoadCombineOp Test
-TEST(SaveLoadCombineFP16Op, CPU) {
+// FP16 version of SaveLoadCombineOp Test, only altering the saving aspect
+// to save as FP16.
+TEST(SaveCombineFP16Op, CPU) {
   paddle::framework::Scope scope;
   paddle::platform::CPUPlace place;
 
@@ -169,7 +170,7 @@ TEST(SaveLoadCombineFP16Op, CPU) {
       20, 50, lod4, "test_var4", place, &scope, &expect_lod4);
 
   // Set attributes
-  std::string filename = "check_tensor_fp16.ls";
+  std::string filename = "check_tensor_fp16_save.ls";
   paddle::framework::AttributeMap attrs;
   attrs.insert({"file_path", std::string(filename)});
   attrs.insert({"save_as_fp16", true});
@@ -216,6 +217,89 @@ TEST(SaveLoadCombineFP16Op, CPU) {
                                                 actual_lod4, numel4);
 }
 
+// FP16 version of SaveLoadCombineOp Test, only altering the loading aspect
+// to load tensors with FP16 precision.
+TEST(LoadCombineFP16Op, CPU) {
+  paddle::framework::Scope scope;
+  paddle::platform::CPUPlace place;
+
+  std::vector<int> lod1 = {0, 1, 2, 3, 10};
+  int numel1 = 100;
+  paddle::framework::LoD expect_lod1;
+  float* expect1 = CreateForSaveCombineOp<float, paddle::platform::float16>(
+      10, 10, lod1, "test_var1", place, &scope, &expect_lod1);
+
+  std::vector<int> lod2 = {0, 2, 5, 10};
+  int numel2 = 200;
+  paddle::framework::LoD expect_lod2;
+  float* expect2 = CreateForSaveCombineOp<float, paddle::platform::float16>(
+      10, 20, lod2, "test_var2", place, &scope, &expect_lod2);
+
+  std::vector<int> lod3 = {0, 20};
+  int numel3 = 4000;
+  paddle::framework::LoD expect_lod3;
+  float* expect3 = CreateForSaveCombineOp<float, paddle::platform::float16>(
+      20, 200, lod3, "test_var3", place, &scope, &expect_lod3);
+
+  std::vector<int> lod4 = {0, 1, 20};
+  int numel4 = 1000;
+  paddle::framework::LoD expect_lod4;
+  float* expect4 = CreateForSaveCombineOp<float, paddle::platform::float16>(
+      20, 50, lod4, "test_var4", place, &scope, &expect_lod4);
+
+  // Set attributes
+  std::string filename = "check_tensor_fp16_load.ls";
+  paddle::framework::AttributeMap attrs;
+  attrs.insert({"file_path", std::string(filename)});
+
+  // Run the save_combine_op
+  auto save_combine_op = paddle::framework::OpRegistry::CreateOp(
+      "save_combine",
+      {{"X", {"test_var1", "test_var2", "test_var3", "test_var4"}}}, {}, attrs);
+  save_combine_op->Run(scope, place);
+
+  // Set up output vars
+  auto load_var1 = scope.Var("out_var1");
+  auto load_var2 = scope.Var("out_var2");
+  auto load_var3 = scope.Var("out_var3");
+  auto load_var4 = scope.Var("out_var4");
+
+  attrs.insert({"load_as_fp16", true});
+  // Run the load_combine_op
+  auto load_combine_op = paddle::framework::OpRegistry::CreateOp(
+      "load_combine", {},
+      {{"Out", {"out_var1", "out_var2", "out_var3", "out_var4"}}}, attrs);
+  load_combine_op->Run(scope, place);
+
+  auto* target1 = load_var1->GetMutable<paddle::framework::LoDTensor>();
+  auto* target2 = load_var2->GetMutable<paddle::framework::LoDTensor>();
+  auto* target3 = load_var3->GetMutable<paddle::framework::LoDTensor>();
+  auto* target4 = load_var4->GetMutable<paddle::framework::LoDTensor>();
+
+  paddle::framework::LoD actual_lod1, actual_lod2, actual_lod3, actual_lod4;
+  paddle::platform::float16* actual1 =
+      GetValuesAfterLoadCombineOp<paddle::platform::float16>(target1, scope,
+                                                             &actual_lod1);
+  paddle::platform::float16* actual2 =
+      GetValuesAfterLoadCombineOp<paddle::platform::float16>(target2, scope,
+                                                             &actual_lod2);
+  paddle::platform::float16* actual3 =
+      GetValuesAfterLoadCombineOp<paddle::platform::float16>(target3, scope,
+                                                             &actual_lod3);
+  paddle::platform::float16* actual4 =
+      GetValuesAfterLoadCombineOp<paddle::platform::float16>(target4, scope,
+                                                             &actual_lod4);
+
+  CheckValues<float, paddle::platform::float16>(expect1, actual1, expect_lod1,
+                                                actual_lod1, numel1);
+  CheckValues<float, paddle::platform::float16>(expect2, actual2, expect_lod2,
+                                                actual_lod2, numel2);
+  CheckValues<float, paddle::platform::float16>(expect3, actual3, expect_lod3,
+                                                actual_lod3, numel3);
+  CheckValues<float, paddle::platform::float16>(expect4, actual4, expect_lod4,
+                                                actual_lod4, numel4);
+}
+
 // Test with original SaveLoadTest
 TEST(SaveLoadTestWithCombineOp, CPU) {
   paddle::framework::Scope scope;

paddle/fluid/platform/nccl_helper.h

@@ -53,7 +53,7 @@ class NCCLGroupGuard {
   }
 
   inline ~NCCLGroupGuard() {
-    PADDLE_ENFORCE(dynload::ncclGroupEnd());
+    CHECK_EQ(dynload::ncclGroupEnd(), ncclSuccess);
     NCCLMutex().unlock();
   }
 };

paddle/fluid/pybind/pybind.cc

@@ -494,23 +494,61 @@ All parameter, weight, gradient are variables in Paddle.
   m.def("disable_profiler", platform::DisableProfiler);
   m.def("reset_profiler", platform::ResetProfiler);
 
-  py::class_<ParallelExecutor>(m, "ParallelExecutor")
-      .def("__init__",
-           [](ParallelExecutor &self, size_t num_threads, bool use_event,
-              const std::vector<platform::Place> &places,
-              const std::unordered_set<std::string> &params,
-              const std::unordered_set<std::string> &bcast_vars,
-              const ProgramDesc &main_program, const std::string &loss_var_name,
-              Scope *scope, std::vector<Scope *> &local_scopes,
-              bool allow_op_delay, bool use_default_grad_scale,
-              bool balance_parameter_opt_between_cards, size_t num_trainers,
-              size_t trainer_id) {
-             new (&self) ParallelExecutor(
-                 num_threads, use_event, places, params, bcast_vars,
-                 main_program, loss_var_name, scope, local_scopes,
-                 allow_op_delay, use_default_grad_scale,
-                 balance_parameter_opt_between_cards, num_trainers, trainer_id);
-           })
+  // -- python binds for parallel executor.
+  py::class_<ParallelExecutor> pe(m, "ParallelExecutor");
+  py::class_<ExecutionStrategy>(pe, "ExecutionStrategy")
+      .def(py::init())
+      .def_property(
+          "num_threads",
+          [](const ExecutionStrategy &self) { return self.num_threads_; },
+          [](ExecutionStrategy &self, size_t num_threads) {
+            self.num_threads_ = num_threads;
+          })
+      .def_property(
+          "use_event",
+          [](const ExecutionStrategy &self) { return self.use_event_; },
+          [](ExecutionStrategy &self, bool use_event) {
+            self.use_event_ = use_event;
+          })
+      .def_property(
+          "allow_op_delay",
+          [](const ExecutionStrategy &self) { return self.allow_op_delay_; },
+          [](ExecutionStrategy &self, bool allow_op_delay) {
+            self.allow_op_delay_ = allow_op_delay;
+          });
+  py::class_<BuildStrategy> build_strategy(pe, "BuildStrategy");
+
+  py::enum_<BuildStrategy::ReduceStrategy>(build_strategy, "ReduceStrategy")
+      .value("Reduce", BuildStrategy::ReduceStrategy::kReduce)
+      .value("AllReduce", BuildStrategy::ReduceStrategy::kAllReduce);
+  py::enum_<BuildStrategy::GradientScaleStrategy>(build_strategy,
+                                                  "GradientScaleStrategy")
+      .value("CoeffNumDevice",
+             BuildStrategy::GradientScaleStrategy::kCoeffNumDevice)
+      .value("One", BuildStrategy::GradientScaleStrategy::kOne)
+      .value("Customized", BuildStrategy::GradientScaleStrategy::kCustomized);
+
+  build_strategy.def(py::init())
+      .def_property(
+          "reduce_strategy",
+          [](const BuildStrategy &self) { return self.reduce_; },
+          [](BuildStrategy &self, BuildStrategy::ReduceStrategy strategy) {
+            self.reduce_ = strategy;
+          })
+      .def_property(
+          "gradient_scale_strategy",
+          [](const BuildStrategy &self) { return self.gradient_scale_; },
+          [](BuildStrategy &self,
+             BuildStrategy::GradientScaleStrategy strategy) {
+            self.gradient_scale_ = strategy;
+          });
+
+  pe.def(py::init<const std::vector<platform::Place> &,
+                  const std::unordered_set<std::string> &,
+                  const std::unordered_set<std::string> &, const ProgramDesc &,
+                  const std::string &, Scope *, std::vector<Scope *> &,
+                  const ExecutionStrategy &, const BuildStrategy &, size_t,
+                  size_t>())
       .def("bcast_params", &ParallelExecutor::BCastParamsToGPUs)
       // NOTE: even we return a vec<Scope*>* to Python use reference policy.
       // We still cannot get local_scope from this vector, since the element

python/paddle/fluid/__init__.py

@@ -44,42 +44,44 @@ import transpiler
 from param_attr import ParamAttr, WeightNormParamAttr
 from data_feeder import DataFeeder
 from core import LoDTensor, CPUPlace, CUDAPlace, CUDAPinnedPlace
-from transpiler import DistributeTranspiler, SimpleDistributeTranspiler, InferenceTranspiler, memory_optimize, release_memory
+from transpiler import DistributeTranspiler, SimpleDistributeTranspiler, \
+    InferenceTranspiler, memory_optimize, release_memory
 from concurrency import (Go, make_channel, channel_send, channel_recv,
                          channel_close, Select)
 import clip
 import profiler
 import unique_name
 import recordio_writer
-from parallel_executor import ParallelExecutor
+import parallel_executor
+from parallel_executor import *
 
 Tensor = LoDTensor
 
-__all__ = framework.__all__ + executor.__all__ + concurrency.__all__ +\
-          trainer.__all__ + inferencer.__all__ + transpiler.__all__ + [
-    'io',
-    'initializer',
-    'layers',
-    'transpiler'
-    'nets',
-    'optimizer',
-    'learning_rate_decay',
-    'backward',
-    'regularizer',
-    'LoDTensor',
-    'CPUPlace',
-    'CUDAPlace',
-    'CUDAPinnedPlace',
-    'Tensor',
-    'ParamAttr',
-    'WeightNormParamAttr',
-    'DataFeeder',
-    'clip',
-    'profiler',
-    'unique_name',
-    'recordio_writer',
-    'ParallelExecutor',
-]
+__all__ = framework.__all__ + executor.__all__ + concurrency.__all__ + \
+          trainer.__all__ + inferencer.__all__ + transpiler.__all__ + \
+          parallel_executor.__all__ + [
+              'io',
+              'initializer',
+              'layers',
+              'transpiler'
+              'nets',
+              'optimizer',
+              'learning_rate_decay',
+              'backward',
+              'regularizer',
+              'LoDTensor',
+              'CPUPlace',
+              'CUDAPlace',
+              'CUDAPinnedPlace',
+              'Tensor',
+              'ParamAttr',
+              'WeightNormParamAttr',
+              'DataFeeder',
+              'clip',
+              'profiler',
+              'unique_name',
+              'recordio_writer',
+          ]
 
 
 def __bootstrap__():

python/paddle/fluid/backward.py

@@ -480,6 +480,8 @@ def append_backward(loss, parameter_list=None, no_grad_set=None,
 
     program.current_block_idx = current_block_idx
     program.sync_with_cpp()
+    # FIXME(zcd): prevent loss.grad optimized by mem_opt.
+    loss.block.var(_append_grad_suffix_(loss.name)).persistable = True
 
     if parameter_list is not None:
         parameters = parameter_list

python/paddle/fluid/inferencer.py

@@ -13,29 +13,35 @@
 # limitations under the License.
 
 import core
-import framework
+
 import executor
+import framework
 import io
+import unique_name
 from trainer import check_and_get_place
 
 __all__ = ['Inferencer', ]
 
 
 class Inferencer(object):
-    def __init__(self, param_path, place=None):
+    def __init__(self, infer_func, param_path, place=None):
         """
-        :param param_path: the path where the inference model is saved by fluid.io.save_inference_model
+        :param infer_func: a function that will return predict Variable
+        :param param_path: the path where the inference model is saved by fluid.io.save_params
         :param place: place to do the inference
         """
         self.param_path = param_path
         self.scope = core.Scope()
 
+        self.inference_program = framework.Program()
+        with framework.program_guard(self.inference_program):
+            with unique_name.guard():
+                self.predict_var = infer_func()
+
         self.exe = executor.Executor(check_and_get_place(place))
         with executor.scope_guard(self.scope):
             # load params from param_path into scope
-            [self.inference_program, _,
-             self.fetch_targets] = io.load_inference_model(
-                 executor=self.exe, dirname=param_path)
+            io.load_params(self.exe, param_path, self.inference_program)
 
     def infer(self, inputs, return_numpy=True):
         """
@@ -51,7 +57,7 @@ class Inferencer(object):
         with executor.scope_guard(self.scope):
             results = self.exe.run(self.inference_program,
                                    feed=inputs,
-                                   fetch_list=self.fetch_targets,
+                                   fetch_list=[self.predict_var],
                                    return_numpy=return_numpy)
 
         return results

python/paddle/fluid/parallel_executor.py

@@ -19,7 +19,10 @@ import executor
 import warnings
 import sys
 
-__all__ = ['ParallelExecutor']
+__all__ = ['ParallelExecutor', 'ExecutionStrategy', 'BuildStrategy']
+
+ExecutionStrategy = core.ParallelExecutor.ExecutionStrategy
+BuildStrategy = core.ParallelExecutor.BuildStrategy
 
 
 class ParallelExecutor(object):
@@ -27,13 +30,12 @@ class ParallelExecutor(object):
                  use_cuda,
                  loss_name=None,
                  main_program=None,
-                 num_threads=None,
-                 allow_op_delay=False,
                  share_vars_from=None,
-                 use_default_grad_scale=True,
-                 balance_parameter_opt_between_cards=False,
+                 exec_strategy=None,
+                 build_strategy=None,
                  num_trainers=1,
-                 trainer_id=0):
+                 trainer_id=0,
+                 **kwargs):
         """
         ParallelExecutor can run program in parallel.
 
@@ -42,21 +44,8 @@ class ParallelExecutor(object):
             loss_name(str, default None): The loss name must set in training.
             main_program(Program, default None): The program that need to run,
                 if not provided, then default_main_program will be used.
-            num_threads(int, default None): How many threads are used for
-                training.
-            allow_op_delay(bool, default False): Whether to delay and buffer
-                some operators together for scheduling or not, which may
-                improve performance in some cases, default False.
             share_vars_from(ParallelExecutor, default None): If provied,
                 it will share variables from the specified ParallelExecutor.
-            use_default_grad_scale(bool, default True): If set True, a default
-                scale value equal to `1./device_count` would be multiplied to
-                gradients of each device and scaled gradients would be
-                aggregated. Otherwise, a customized scale value should be fed
-                to the network.
-            balance_parameter_opt_between_cards(bool, default True): Whether
-                updating different gradients on different cards. Currently, it
-                is not recommended.
             num_trainers(int, default 1): If greater than 1, NCCL will be
                 initialized with multpile rank of nodes, each node should have
                 same number of GPUs. Distributed training will be enabled then.
@@ -83,6 +72,25 @@ class ParallelExecutor(object):
               train_loss, = train_exe.run([loss.name], feed=feed_dict)
               test_loss, = test_exe.run([loss.name], feed=feed_dict)
         """
+        if len(kwargs) != 0:
+            err_msg = ""
+            for key in kwargs:
+                if key in dir(ExecutionStrategy):
+                    err_msg += \
+                        "Setting {0} by constructor is deprecated. Use " \
+                        "strategy=ExecutionStrategy(); strategy.{0}=xxx; " \
+                        "pe=ParallelExecutor(exec_strategy=strategy) " \
+                        "instead.\n ".format(key)
+                elif key in dir(BuildStrategy):
+                    err_msg += \
+                        "Setting {0} by constructor is deprecated. Use " \
+                        "strategy=BuildStrategy(); See help(" \
+                        "paddle.fluid.ParallelExecutor.BuildStrategy) \n".format(
+                            key)
+                else:
+                    err_msg += "Setting {0} by constructor is deprecated. Use strategy.\n".format(
+                        key)
+            raise ValueError(err_msg)
 
         self._places = []
         self._act_places = []
@@ -100,15 +108,25 @@ class ParallelExecutor(object):
                 self._places.append(p)
         assert self._places, "no place for execution"
 
-        if num_threads is None:
+        if exec_strategy is None:
+            exec_strategy = ExecutionStrategy()
+            if use_cuda:
+                exec_strategy.use_event = True
+            else:
+                exec_strategy.use_event = False
+
+        if exec_strategy.num_threads == 0:
             if use_cuda:
                 # Experiments on se-resnext shows that too many threads hurt
                 # performance. Worth tunning for other models in the future.
-                num_threads = len(self._places) * 2
+                exec_strategy.num_threads = len(self._places) * 2
             else:
-                num_threads = min(
+                exec_strategy.num_threads = min(
                     len(self._places) * 2, multiprocessing.cpu_count())
 
+        if build_strategy is None:
+            build_strategy = BuildStrategy()
+
         main = main_program
         main = main if main else framework.default_main_program()
         scope = executor.global_scope()
@@ -127,23 +145,14 @@ class ParallelExecutor(object):
         ]
 
         self.executor = core.ParallelExecutor(
-            num_threads,
-            True if use_cuda else False,  # use_event
             self._places,
             set([
                 p.name for p in main.global_block().iter_parameters()
                 if not p.stop_gradient
             ]),
-            set(self.persistable_vars),
-            main.desc,
-            loss_name if loss_name else '',
-            scope,
-            local_scopes,
-            allow_op_delay,
-            use_default_grad_scale,
-            balance_parameter_opt_between_cards,
-            num_trainers,
-            trainer_id)
+            set(self.persistable_vars), main.desc, loss_name
+            if loss_name else '', scope, local_scopes, exec_strategy,
+            build_strategy, num_trainers, trainer_id)
         self.scope = scope
 
     def run(self, fetch_list, feed=None, feed_dict=None):

python/paddle/fluid/tests/book/high-level-api/fit_a_line/test_fit_a_line.py

@@ -48,12 +48,11 @@ def linear():
     return avg_loss
 
 
-def train(use_cuda, save_dirname):
+def train(use_cuda, train_program, save_dirname):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
     trainer = fluid.Trainer(
-        train_func=linear,
-        infer_func=inference_program,
+        train_func=train_program,
         place=place,
         optimizer=fluid.optimizer.SGD(learning_rate=0.001))
 
@@ -72,11 +71,7 @@ def train(use_cuda, save_dirname):
             '''
             if float(test_metrics[0]) < 20.0:
                 if save_dirname is not None:
-                    # NOT clear yet
-                    # fluid.io.save_inference_model(save_dirname, ['x'], [y_predict])
-                    # trainer.save_params(save_dirname)
-                    # https://github.com/PaddlePaddle/Paddle/pull/10445
-                    trainer.save_inference_model(save_dirname)
+                    trainer.save_params(save_dirname)
             return
 
     trainer.train(
@@ -87,12 +82,13 @@ def train(use_cuda, save_dirname):
 
 
 # infer
-def infer(use_cuda, save_dirname=None):
+def infer(use_cuda, inference_program, save_dirname=None):
     if save_dirname is None:
         return
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    inferencer = fluid.Inferencer(param_path=save_dirname, place=place)
+    inferencer = fluid.Inferencer(
+        infer_func=inference_program, param_path=save_dirname, place=place)
 
     batch_size = 10
     tensor_x = numpy.random.uniform(0, 10, [batch_size, 13]).astype("float32")
@@ -108,8 +104,8 @@ def main(use_cuda):
     # Directory for saving the trained model
     save_dirname = "fit_a_line.inference.model"
 
-    train(use_cuda, save_dirname)
-    infer(use_cuda, save_dirname)
+    train(use_cuda, linear, save_dirname)
+    infer(use_cuda, inference_program, save_dirname)
 
 
 class TestFitALine(unittest.TestCase):

python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_conv.py

@@ -53,48 +53,40 @@ def train_program():
     predict = inference_program()
     cost = fluid.layers.cross_entropy(input=predict, label=label)
     avg_cost = fluid.layers.mean(cost)
-    # acc = fluid.layers.accuracy(input=predict, label=label)
-    # return avg_cost, acc
-    return avg_cost
+    acc = fluid.layers.accuracy(input=predict, label=label)
+    return [avg_cost, acc]
 
 
-def train(use_cuda, save_dirname):
+def train(use_cuda, train_program, save_dirname):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
 
     trainer = fluid.Trainer(
-        train_func=train_program,
-        infer_func=inference_program,
-        place=place,
-        optimizer=optimizer)
+        train_func=train_program, place=place, optimizer=optimizer)
 
     def event_handler(event):
         if isinstance(event, fluid.EndEpochEvent):
-            # if (event.epoch + 1) % 10 == 0:
-            # trainer.save_params(save_dirname)
-            trainer.save_inference_model(save_dirname)
-
-            # TODO: Uncomment this part once we are sure that .train is working
-            # test_reader = paddle.batch(
-            #     paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
-            # test_metrics = trainer.test(reader=test_reader)
-            # avg_cost_set = test_metrics[0]
-            # acc_set = test_metrics[1]
-            #
-            # # get test acc and loss
-            # acc = numpy.array(acc_set).mean()
-            # avg_cost = numpy.array(avg_cost_set).mean()
-            #
-            # print("avg_cost: %s" % avg_cost)
-            # print("acc     : %s" % acc)
-            #
-            # if float(acc) > 0.2:  # Smaller value to increase CI speed
-            #     trainer.save_params(save_dirname)
-            # else:
-            #     print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
-            #         event.epoch + 1, float(avg_cost), float(acc)))
-            #     if math.isnan(float(avg_cost)):
-            #         sys.exit("got NaN loss, training failed.")
+            test_reader = paddle.batch(
+                paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
+            test_metrics = trainer.test(
+                reader=test_reader, feed_order=['img', 'label'])
+            avg_cost_set = test_metrics[0]
+            acc_set = test_metrics[1]
+
+            # get test acc and loss
+            acc = numpy.array(acc_set).mean()
+            avg_cost = numpy.array(avg_cost_set).mean()
+
+            print("avg_cost: %s" % avg_cost)
+            print("acc     : %s" % acc)
+
+            if float(acc) > 0.2:  # Smaller value to increase CI speed
+                trainer.save_params(save_dirname)
+            else:
+                print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
+                    event.epoch + 1, float(avg_cost), float(acc)))
+                if math.isnan(float(avg_cost)):
+                    sys.exit("got NaN loss, training failed.")
 
     train_reader = paddle.batch(
         paddle.reader.shuffle(
@@ -108,10 +100,11 @@ def train(use_cuda, save_dirname):
         feed_order=['img', 'label'])
 
 
-def infer(use_cuda, save_dirname=None):
+def infer(use_cuda, inference_program, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
-    inferencer = fluid.Inferencer(param_path=save_dirname, place=place)
+    inferencer = fluid.Inferencer(
+        infer_func=inference_program, param_path=save_dirname, place=place)
 
     batch_size = 1
     tensor_img = numpy.random.uniform(-1.0, 1.0,
@@ -126,8 +119,14 @@ def main(use_cuda):
     save_dirname = "recognize_digits_conv.inference.model"
 
     # call train() with is_local argument to run distributed train
-    train(use_cuda=use_cuda, save_dirname=save_dirname)
-    infer(use_cuda=use_cuda, save_dirname=save_dirname)
+    train(
+        use_cuda=use_cuda,
+        train_program=train_program,
+        save_dirname=save_dirname)
+    infer(
+        use_cuda=use_cuda,
+        inference_program=inference_program,
+        save_dirname=save_dirname)
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/book/high-level-api/recognize_digits/test_recognize_digits_mlp.py

@@ -40,47 +40,40 @@ def train_program():
     predict = inference_program()
     cost = fluid.layers.cross_entropy(input=predict, label=label)
     avg_cost = fluid.layers.mean(cost)
-    # acc = fluid.layers.accuracy(input=predict, label=label)
-    # return avg_cost, acc
-    return avg_cost
+    acc = fluid.layers.accuracy(input=predict, label=label)
+    return [avg_cost, acc]
 
 
-def train(use_cuda, save_dirname):
+def train(use_cuda, train_program, save_dirname):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
 
     trainer = fluid.Trainer(
-        train_func=train_program,
-        infer_func=inference_program,
-        place=place,
-        optimizer=optimizer)
+        train_func=train_program, place=place, optimizer=optimizer)
 
     def event_handler(event):
         if isinstance(event, fluid.EndEpochEvent):
-            # if (event.epoch + 1) % 10 == 0:
-            trainer.save_inference_model(save_dirname)
-
-            # TODO: Uncomment this part once we are sure that .train is working
-            # test_reader = paddle.batch(
-            #     paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
-            # test_metrics = trainer.test(reader=test_reader)
-            # avg_cost_set = test_metrics[0]
-            # acc_set = test_metrics[1]
-            #
-            # # get test acc and loss
-            # acc = numpy.array(acc_set).mean()
-            # avg_cost = numpy.array(avg_cost_set).mean()
-            #
-            # print("avg_cost: %s" % avg_cost)
-            # print("acc     : %s" % acc)
-            #
-            # if float(acc) > 0.2:  # Smaller value to increase CI speed
-            #     trainer.save_params(save_dirname)
-            # else:
-            #     print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
-            #         event.epoch + 1, float(avg_cost), float(acc)))
-            #     if math.isnan(float(avg_cost)):
-            #         sys.exit("got NaN loss, training failed.")
+            test_reader = paddle.batch(
+                paddle.dataset.mnist.test(), batch_size=BATCH_SIZE)
+            test_metrics = trainer.test(
+                reader=test_reader, feed_order=['img', 'label'])
+            avg_cost_set = test_metrics[0]
+            acc_set = test_metrics[1]
+
+            # get test acc and loss
+            acc = numpy.array(acc_set).mean()
+            avg_cost = numpy.array(avg_cost_set).mean()
+
+            print("avg_cost: %s" % avg_cost)
+            print("acc     : %s" % acc)
+
+            if float(acc) > 0.2:  # Smaller value to increase CI speed
+                trainer.save_params(save_dirname)
+            else:
+                print('BatchID {0}, Test Loss {1:0.2}, Acc {2:0.2}'.format(
+                    event.epoch + 1, float(avg_cost), float(acc)))
+                if math.isnan(float(avg_cost)):
+                    sys.exit("got NaN loss, training failed.")
 
     train_reader = paddle.batch(
         paddle.reader.shuffle(
@@ -94,10 +87,11 @@ def train(use_cuda, save_dirname):
         feed_order=['img', 'label'])
 
 
-def infer(use_cuda, save_dirname=None):
+def infer(use_cuda, inference_program, save_dirname=None):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 
-    inferencer = fluid.Inferencer(param_path=save_dirname, place=place)
+    inferencer = fluid.Inferencer(
+        infer_func=inference_program, param_path=save_dirname, place=place)
 
     batch_size = 1
     tensor_img = numpy.random.uniform(-1.0, 1.0,
@@ -112,8 +106,14 @@ def main(use_cuda):
     save_dirname = "recognize_digits_mlp.inference.model"
 
     # call train() with is_local argument to run distributed train
-    train(use_cuda=use_cuda, save_dirname=save_dirname)
-    infer(use_cuda=use_cuda, save_dirname=save_dirname)
+    train(
+        use_cuda=use_cuda,
+        train_program=train_program,
+        save_dirname=save_dirname)
+    infer(
+        use_cuda=use_cuda,
+        inference_program=inference_program,
+        save_dirname=save_dirname)
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/book/high-level-api/word2vec/no_test_word2vec_new_api.py

@@ -90,7 +90,7 @@ def train_program(is_sparse):
     return avg_cost
 
 
-def train(use_cuda, is_sparse, save_path):
+def train(use_cuda, train_program, save_path):
     train_reader = paddle.batch(
         paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE)
     test_reader = paddle.batch(
@@ -105,23 +105,21 @@ def train(use_cuda, is_sparse, save_path):
             print("loss= ", avg_cost)
 
             if avg_cost < 5.0:
-                trainer.save_inference_model(save_path)
+                trainer.save_params(save_path)
                 return
             if math.isnan(avg_cost):
                 sys.exit("got NaN loss, training failed.")
 
     trainer = fluid.Trainer(
-        partial(train_program, is_sparse),
-        partial(inference_program, is_sparse),
-        fluid.optimizer.SGD(learning_rate=0.001),
-        place=place)
+        train_program, fluid.optimizer.SGD(learning_rate=0.001), place=place)
     trainer.train(
         reader=train_reader, num_epochs=1, event_handler=event_handler)
 
 
-def infer(use_cuda, is_sparse, save_path):
+def infer(use_cuda, inference_program, save_path):
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
-    inferencer = fluid.Inferencer(param_path=save_path, place=place)
+    inferencer = fluid.Inferencer(
+        infer_func=inference_program, param_path=save_path, place=place)
 
     lod = [0, 1]
     first_word = create_random_lodtensor(lod, place, low=0, high=dict_size - 1)
@@ -144,9 +142,9 @@ def main(use_cuda, is_sparse):
     if use_cuda and not fluid.core.is_compiled_with_cuda():
         return
 
-    save_path = "word2vec.inference.model"
-    train(use_cuda, is_sparse, save_path)
-    infer(use_cuda, is_sparse, save_path)
+    save_path = "word2vec.params"
+    train(use_cuda, partial(train_program, is_sparse), save_path)
+    infer(use_cuda, partial(inference_program, is_sparse), save_path)
 
 
 if __name__ == '__main__':

python/paddle/fluid/tests/book/notest_understand_sentiment.py

@@ -170,7 +170,7 @@ def train(word_dict,
         assert save_dirname is None
 
     adagrad = fluid.optimizer.Adagrad(learning_rate=0.002)
-    optimize_ops, params_grads = adagrad.minimize(cost)
+    adagrad.minimize(cost)
 
     train_data = paddle.batch(
         paddle.reader.shuffle(

python/paddle/fluid/tests/book/test_fit_a_line.py

@@ -33,7 +33,7 @@ def train(use_cuda, save_dirname, is_local):
     avg_cost = fluid.layers.mean(cost)
 
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     BATCH_SIZE = 20
 

python/paddle/fluid/tests/book/test_image_classification.py

@@ -125,7 +125,7 @@ def train(net_type, use_cuda, save_dirname, is_local):
     test_program = fluid.default_main_program().clone(for_test=True)
 
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    optimize_ops, params_grads = optimizer.minimize(avg_cost)
+    optimizer.minimize(avg_cost)
 
     BATCH_SIZE = 128
     PASS_NUM = 1

python/paddle/fluid/tests/book/test_label_semantic_roles.py

@@ -175,7 +175,7 @@ def train(use_cuda, save_dirname=None, is_local=True):
             decay_steps=100000,
             decay_rate=0.5,
             staircase=True))
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     # TODO(qiao)
     # add dependency track and move this config before optimizer

python/paddle/fluid/tests/book/test_machine_translation.py

@@ -185,7 +185,7 @@ def train_main(use_cuda, is_sparse, is_local=True):
         learning_rate=1e-4,
         regularization=fluid.regularizer.L2DecayRegularizer(
             regularization_coeff=0.1))
-    optimize_ops, params_grads = optimizer.minimize(avg_cost)
+    optimizer.minimize(avg_cost)
 
     train_data = paddle.batch(
         paddle.reader.shuffle(

python/paddle/fluid/tests/book/test_recognize_digits.py

@@ -95,7 +95,7 @@ def train(nn_type,
     test_program = fluid.default_main_program().clone(for_test=True)
 
     optimizer = fluid.optimizer.Adam(learning_rate=0.001)
-    optimize_ops, params_grads = optimizer.minimize(avg_loss)
+    optimizer.minimize(avg_loss)
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 

python/paddle/fluid/tests/book/test_recommender_system.py

@@ -160,7 +160,7 @@ def train(use_cuda, save_dirname, is_local=True):
     test_program = fluid.default_main_program().clone(for_test=True)
 
     sgd_optimizer = SGDOptimizer(learning_rate=0.2)
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     place = fluid.CUDAPlace(0) if use_cuda else fluid.CPUPlace()
 

python/paddle/fluid/tests/book/test_word2vec.py

@@ -101,7 +101,7 @@ def train(use_cuda, is_sparse, is_parallel, save_dirname, is_local=True):
         avg_cost = fluid.layers.mean(pd())
 
     sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.001)
-    optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+    sgd_optimizer.minimize(avg_cost)
 
     train_reader = paddle.batch(
         paddle.dataset.imikolov.train(word_dict, N), BATCH_SIZE)

python/paddle/fluid/tests/unittests/test_dist_transpiler.py

+#   Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# 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.
+
+import unittest
+
+import paddle.fluid as fluid
+import paddle.fluid.core as core
+import paddle.fluid.layers as layers
+from paddle.fluid.transpiler.distribute_transpiler import delete_ops
+import numpy
+
+
+class TestDistTranspiler(unittest.TestCase):
+    def setUp(self):
+        self.trainer_id = 0
+        self.trainers = 2
+        self.pservers = 2
+        self.pserver_eps = "127.0.0.1:6174,127.0.0.1:6175"
+        self.current_pserver_ep = "127.0.0.1:6174"
+
+    def net_conf(self):
+        x = fluid.layers.data(name='x', shape=[1000], dtype='float32')
+
+        y_predict = fluid.layers.fc(input=x,
+                                    size=1000,
+                                    act=None,
+                                    param_attr=fluid.ParamAttr(name='fc_w'))
+
+        y = fluid.layers.data(name='y', shape=[1], dtype='float32')
+
+        cost = fluid.layers.square_error_cost(input=y_predict, label=y)
+        avg_cost = fluid.layers.mean(cost)
+        sgd_optimizer = fluid.optimizer.SGD(learning_rate=0.1)
+
+        optimize_ops, params_grads = sgd_optimizer.minimize(avg_cost)
+        return optimize_ops, params_grads
+
+    def test_transpiler(self):
+        trainer = self.get_trainer()
+        pserver, startup = self.get_pserver(self.current_pserver_ep)
+
+        self.assertEqual([op.type for op in trainer.global_block().ops],
+                         self.get_expect_trainer_ops())
+
+        self.assertEqual(len(pserver.blocks), 3)
+        # block0: listen_and_serv
+        self.assertEqual([op.type for op in pserver.blocks[0].ops],
+                         ["listen_and_serv"])
+        # block2: optimize pass
+        self.assertEqual([op.type for op in pserver.blocks[1].ops],
+                         ["sum", "scale", "sgd"])
+
+        # confirm startup program
+
+        self.assertEqual([op.type for op in startup.global_block().ops], [
+            "fill_constant", "fill_constant", "uniform_random", "uniform_random"
+        ])
+
+        # the variable #fc_w will be split into two blocks 
+        fc_w_var = startup.global_block().var("fc_w.block1")
+        self.assertEqual(fc_w_var.shape, (500, 1000))
+
+    def get_main_program(self):
+        main = fluid.Program()
+
+        with fluid.program_guard(main):
+            self.net_conf()
+
+        return main
+
+    def get_expect_trainer_ops(self):
+        trainer = fluid.Program()
+
+        with fluid.program_guard(trainer):
+            optimize_ops, params_grads = self.net_conf()
+
+        delete_ops(trainer.global_block(), optimize_ops)
+        return [op.type for op in trainer.global_block().ops
+                ] + ["split_byref", "send", "concat"]
+
+    def get_trainer(self):
+        return self._transpiler_instance().get_trainer_program()
+
+    def get_pserver(self, ep):
+        t = self._transpiler_instance()
+        pserver = t.get_pserver_program(ep)
+        startup = t.get_startup_program(ep, pserver)
+        return pserver, startup
+
+    def _transpiler_instance(self):
+        main = self.get_main_program()
+        t = fluid.DistributeTranspiler()
+        t.transpile(
+            self.trainer_id,
+            program=main,
+            pservers=self.pserver_eps,
+            trainers=self.trainers)
+        return t
+
+
+if __name__ == "__main__":
+    unittest.main()

python/paddle/fluid/tests/unittests/test_parallel_executor.py

@@ -232,14 +232,18 @@ class TestParallelExecutorBase(unittest.TestCase):
             place = fluid.CUDAPlace(0)
             startup_exe = fluid.Executor(place)
             startup_exe.run(startup)
+            exec_strategy = fluid.ExecutionStrategy()
+            exec_strategy.allow_op_delay = allow_op_delay
+
+            build_strategy = fluid.BuildStrategy()
+            build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce if balance_parameter_opt_between_cards else fluid.BuildStrategy.ReduceStrategy.AllReduce
 
             if use_parallel_executor:
                 exe = fluid.ParallelExecutor(
                     True,
                     loss_name=loss.name,
-                    allow_op_delay=allow_op_delay,
-                    balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
-                )
+                    exec_strategy=exec_strategy,
+                    build_strategy=build_strategy)
             else:
                 exe = fluid.Executor(place=place)
 
@@ -548,7 +552,7 @@ class TestTransformer(TestParallelExecutorBase):
 
 
 class ParallelExecutorTestingDuringTraining(unittest.TestCase):
-    def check_network_convergence(self, balance_parameter_opt_between_cards):
+    def check_network_convergence(self, build_strategy=None):
         main = fluid.Program()
         startup = fluid.Program()
         with fluid.program_guard(main, startup):
@@ -571,15 +575,13 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
                 use_cuda=True,
                 loss_name=loss.name,
                 main_program=main,
-                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
-            )
+                build_strategy=build_strategy)
 
             test_exe = fluid.ParallelExecutor(
                 use_cuda=True,
                 main_program=test_program,
                 share_vars_from=train_exe,
-                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
-            )
+                build_strategy=build_strategy)
 
             for i in xrange(5):
                 test_loss, = test_exe.run([loss.name], feed=feed_dict)
@@ -594,10 +596,14 @@ class ParallelExecutorTestingDuringTraining(unittest.TestCase):
                     str(test_loss))
 
     def test_parallel_testing(self):
-        self.check_network_convergence(False)
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
+        self.check_network_convergence(build_strategy)
 
     def test_parallel_testing_with_new_strategy(self):
-        self.check_network_convergence(True)
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
+        self.check_network_convergence(build_strategy)
 
 
 import paddle.dataset.conll05 as conll05
@@ -617,7 +623,7 @@ embedding_name = 'emb'
 
 
 def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
-            is_sparse, balance_parameter_opt_between_cards, **ignored):
+            is_sparse, **ignored):
     # 8 features
     predicate_embedding = fluid.layers.embedding(
         input=predicate,
@@ -686,9 +692,7 @@ def db_lstm(word, predicate, ctx_n2, ctx_n1, ctx_0, ctx_p1, ctx_p2, mark,
 
 
 class TestCRFModel(unittest.TestCase):
-    def check_network_convergence(self,
-                                  is_sparse,
-                                  balance_parameter_opt_between_cards=False):
+    def check_network_convergence(self, is_sparse, build_strategy=None):
         main = fluid.Program()
         startup = fluid.Program()
         with fluid.program_guard(main, startup):
@@ -739,8 +743,7 @@ class TestCRFModel(unittest.TestCase):
             pe = fluid.ParallelExecutor(
                 use_cuda=True,
                 loss_name=avg_cost.name,
-                balance_parameter_opt_between_cards=balance_parameter_opt_between_cards
-            )
+                build_strategy=build_strategy)
 
             feeder = fluid.DataFeeder(
                 feed_list=[
@@ -756,19 +759,29 @@ class TestCRFModel(unittest.TestCase):
                           pe.run(feed=feeder.feed(cur_batch),
                                  fetch_list=[avg_cost.name]))[0]
 
-    def test_update_sparse_parameter(self):
-        self.check_network_convergence(is_sparse=True)
+    def test_update_sparse_parameter_all_reduce(self):
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
+        self.check_network_convergence(
+            is_sparse=True, build_strategy=build_strategy)
 
-    def test_update_dense_parameter(self):
-        self.check_network_convergence(is_sparse=False)
+    def test_update_dense_parameter_all_reduce(self):
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.AllReduce
+        self.check_network_convergence(
+            is_sparse=False, build_strategy=build_strategy)
 
-    def test_update_sparse_parameter_with_new_strategy(self):
+    def test_update_sparse_parameter_reduce(self):
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
         self.check_network_convergence(
-            is_sparse=False, balance_parameter_opt_between_cards=True)
+            is_sparse=False, build_strategy=build_strategy)
 
-    def test_update_dense_parameter_with_new_strategy(self):
+    def test_update_dense_parameter_reduce(self):
+        build_strategy = fluid.BuildStrategy()
+        build_strategy.reduce_strategy = fluid.BuildStrategy.ReduceStrategy.Reduce
         self.check_network_convergence(
-            is_sparse=False, balance_parameter_opt_between_cards=True)
+            is_sparse=False, build_strategy=build_strategy)
 
 
 # test fetch all the variables of global_block
@@ -836,7 +849,8 @@ class TestFetchOp(unittest.TestCase):
                     assert not math.isnan(np.sum(ret[i])) and \
                            not math.isinf(np.sum(ret[i]))
 
-    def test_update_sparse_parameter(self):
+    @unittest.skip("this test is buggy")
+    def test_feed(self):
         tst_reader = paddle.batch(flowers.test(use_xmap=False), batch_size=16)
         tst_reader_iter = tst_reader()
 

python/paddle/fluid/tests/unittests/test_split_var.py

@@ -21,15 +21,7 @@ import random
 
 
 class TestSplitVar(unittest.TestCase):
-    def test_check_output(self):
-        # split below shapes to 10 servers
-        shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]]
-        expected_sizes = [
-            [15], [1024],
-            [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784],
-            [2040, 2040, 2040, 2040],
-            [1150, 1150, 1150, 1150, 1150, 1150, 1100]
-        ]
+    def check_split_output(self, shapes, expected_sizes, min_size):
         var_list = []
         program = fluid.Program()
         for shape in shapes:
@@ -39,7 +31,7 @@ class TestSplitVar(unittest.TestCase):
                 # dtype=core.VarDesc.VarType.LOD_TENSOR,
                 shape=shape)
             var_list.append(var)
-        blocks = split_dense_variable(var_list, 10)
+        blocks = split_dense_variable(var_list, 10, min_size)
         all_sizes = []
         for s in expected_sizes:
             for s2 in s:
@@ -48,6 +40,25 @@ class TestSplitVar(unittest.TestCase):
             varname, block_id, size = block_str.split(":")
             self.assertEqual(int(size), all_sizes[i])
 
+    def test_1k(self):
+        shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10]]
+        expected_sizes = [
+            [15], [1024],
+            [2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 2352, 784],
+            [2040, 2040, 2040, 2040],
+            [1150, 1150, 1150, 1150, 1150, 1150, 1100]
+        ]
+
+        self.check_split_output(shapes, expected_sizes, 1024)
+
+    def test_check_output_8k(self):
+        shapes = [[3, 5], [1024], [28, 784], [8, 1020], [800, 10],
+                  [6, 33, 33, 33]]
+        expected_sizes = [[15], [1024], [10976, 10976], [8160], [8000],
+                          [35937, 35937, 35937, 35937, 35937, 35937]]
+
+        self.check_split_output(shapes, expected_sizes, 8192)
+
 
 if __name__ == '__main__':
     unittest.main()

python/paddle/fluid/trainer.py

@@ -92,19 +92,13 @@ class Trainer(object):
         place: The device place of this trainer.
     """
 
-    def __init__(self,
-                 train_func,
-                 infer_func,
-                 optimizer,
-                 param_path=None,
-                 place=None):
+    def __init__(self, train_func, optimizer, param_path=None, place=None):
         # 1. we need to generate a framework.Program by calling
         # program_func. Reference: fluid.program_guard in
         # test_word2vec.py
         if not isinstance(optimizer, opt_module.Optimizer):
             raise TypeError("The optimizer should be an instance of Optimizer")
 
-        self.infer_func = infer_func
         self.scope = core.Scope()
 
         self.startup_program = framework.Program()
@@ -178,9 +172,9 @@ class Trainer(object):
     def train(self,
               num_epochs,
               event_handler,
-              reader=None,
-              parallel=False,
-              feed_order=None):
+              reader,
+              feed_order,
+              parallel=False):
         """
         Train the model.
 
@@ -208,7 +202,7 @@ class Trainer(object):
 
         self._train_by_executor(num_epochs, event_handler, reader, feed_order)
 
-    def test(self, reader, feed_order=None):
+    def test(self, reader, feed_order):
         """
         Test the model on given test data
 
@@ -226,15 +220,6 @@ class Trainer(object):
             exe = executor.Executor(self.place)
             io.save_persistables(exe, dirname=param_path)
 
-    def save_inference_model(self, model_path):
-        inference_program = framework.Program()
-        with framework.program_guard(inference_program):
-            with unique_name.guard():
-                predict_var = self.infer_func()
-        predict_var = self.train_program.block(0).var(predict_var.name)
-        exe = executor.Executor(self.place)
-        io.save_inference_model(model_path, [], [predict_var], exe)
-
     @contextlib.contextmanager
     def _prog_and_scope_guard(self):
         with framework.program_guard(
@@ -291,12 +276,7 @@ def build_feed_var_list(program, feed_order):
     if not isinstance(program, framework.Program):
         raise TypeError("The 'program' should be an object of Program")
 
-    if feed_order is None:
-        feed_var_list = [
-            var for var in program.global_block().vars.itervalues()
-            if var.is_data
-        ]
-    elif isinstance(feed_order, list):
+    if isinstance(feed_order, list):
         feed_var_list = [
             program.global_block().var(var_name) for var_name in feed_order
         ]

python/paddle/fluid/transpiler/__init__.py

@@ -11,6 +11,7 @@
 # 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.
+
 from distribute_transpiler import DistributeTranspiler
 from inference_transpiler import InferenceTranspiler
 from memory_optimization_transpiler import memory_optimize, release_memory

python/paddle/fluid/transpiler/distribute_transpiler.py

@@ -17,7 +17,7 @@ from __future__ import print_function
 import math
 
 import distributed_splitter as splitter
-from .. import core
+from .. import core, framework
 from ..framework import Program, default_main_program, \
                         default_startup_program, \
                         Variable, Parameter, grad_var_name
@@ -93,30 +93,33 @@ def same_or_split_var(p_name, var_name):
     return p_name == var_name or p_name.startswith(var_name + ".block")
 
 
-def split_dense_variable(var_list,
-                         pserver_count,
-                         min_block_size=1024,
-                         max_block_size=1048576):
+def split_dense_variable(var_list, service_count, min_block_size=8192):
     """
-        We may need to split dense tensor to one or more blocks and put
-        them equally onto parameter server. One block is a sub-tensor
-        aligned by dim[0] of the tensor.
-
-        We need to have a minimal block size so that the calculations in
-        the parameter server side can gain better performance. By default
-        minimum block size is 1024. The max block size is used to prevent
-        very large blocks that may cause send error.
-        :return: A list of VarBlocks. Each VarBlock specifies a shard of
-           the var.
+    We may need to split dense tensor to one or more blocks and put
+    them equally onto parameter server. One block is a sub-tensor
+    aligned by dim[0] of the tensor.
+
+    We need to have a minimal block size so that the calculations in
+    the parameter server side can gain better performance. By default
+    minimum block size 8K elements (maybe 16bit or 32bit or 64bit). 
+
+    Args:
+        var_list (list): List of variables.
+        service_count (int): Numel of pserver services. A pserver may have two
+            or more listening ports.
+        min_block_size (int): Minimum splitted block size.
+    Returns:
+        blocks (list[(varname, block_id, current_block_size)]): A list 
+            of VarBlocks. Each VarBlock specifies a shard of the var.
     """
     blocks = []
     for var in var_list:
-        split_count = pserver_count
+        split_count = service_count
         var_numel = reduce(lambda x, y: x * y, var.shape)
         max_pserver_count = int(math.floor(var_numel / float(min_block_size)))
         if max_pserver_count == 0:
             max_pserver_count = 1
-        if max_pserver_count < pserver_count:
+        if max_pserver_count < service_count:
             split_count = max_pserver_count
         block_size = int(math.ceil(var_numel / float(split_count)))
 
@@ -270,6 +273,7 @@ class DistributeTranspiler:
         grad_var_mapping = self._append_split_op(program, grad_blocks)
         param_var_mapping = self._create_vars_from_blocklist(program,
                                                              param_blocks)
+
         # step3: Add gradients as send op inputs and parameters as send
         # op outputs.
         send_inputs = []
@@ -277,9 +281,11 @@ class DistributeTranspiler:
         for b in grad_blocks:  # append by order
             varname, block_id, _ = b.split(":")
             send_inputs.append(grad_var_mapping[varname][int(block_id)])
+
         for b in param_blocks:
             varname, block_id, _ = b.split(":")
             send_outputs.append(param_var_mapping[varname][int(block_id)])
+
         # let send_op know which endpoint to send which var to, eplist has the same
         # order as send_inputs.
         eplist = split_method(send_inputs, pserver_endpoints)
@@ -417,7 +423,7 @@ class DistributeTranspiler:
         def __append_optimize_op__(op, block, grad_to_block_id):
             if self._is_opt_op(op):
                 self._append_pserver_ops(block, op, endpoint, grad_to_block_id,
-                                         default_main_program())
+                                         self.origin_program)
             else:
                 self._append_pserver_non_opt_ops(block, op)
 
@@ -751,9 +757,18 @@ class DistributeTranspiler:
         Create vars for each split.
         NOTE: only grads need to be named for different trainers, use
               add_trainer_suffix to rename the grad vars.
-        :return: A dict mapping from original var name to each var split.
+        Args:
+            program (ProgramDesc): ProgramDesc which gradients blong.
+            block_list (list[(varname, block_id, block_size)]): List of gradient blocks.
+            add_trainer_suffix (Bool): Add trainer suffix to new variable's name if set True.
+        Returns: 
+            var_mapping (dict(varname->[new_varname_variable])):A dict mapping 
+                from original var name to each var split.
         """
+
+        # varname->[(block_id, current_block_size)]
         block_map = dict()
+
         var_mapping = dict()
         for block_str in block_list:
             varname, offset, size = block_str.split(":")
@@ -824,7 +839,16 @@ class DistributeTranspiler:
             persistable=persistable)
 
     def _append_split_op(self, program, gradblocks):
-        # Split variables that need to be split and append respective ops
+        """
+        Split variables that need to be split and append respective ops
+        Args:
+            program (ProgramDesc): ProgramDesc that gradients blong.
+            gradblocks (list[(varname, block_id, block_size)]): List of gradient blocks.
+        Returns:
+            var_mapping (dict(varname->[new_splitted_variable])):A dict mapping 
+                from original var name to each var split.
+        """
+
         add_suffix = False
         if self.trainer_num > 1:
             add_suffix = True
@@ -1148,6 +1172,12 @@ class DistributeTranspiler:
         return lr_ops
 
     def _get_optimize_pass(self):
+        """
+        Get optimizer operators, paramters and gradients from origin_program
+        Returns:
+            opt_ops (list): optimize operators.
+            params_grads (dict): paramter->gradient.
+        """
         block = self.origin_program.global_block()
         opt_ops = []
         params_grads = []

tools/manylinux1/README.md

@@ -28,3 +28,38 @@ git clone https://github.com/paddlepaddle/paddle
 cd paddle/tools/manylinux1
 REPO=[yourrepo] ./build_all.sh
 ```
+
+## Build PaddlePaddle for the different Python ABIs
+
+Choose one of the following Python ABI and set the correct environment variables.
+
+- cp27-cp27m
+
+  ```bash
+  export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs2/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs4/lib:}
+  export PATH=/opt/python/cp27-cp27m/bin/:${PATH}
+  export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27m/bin/python
+        -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27m/include/python2.7
+        -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs2/lib/libpython2.7.so"
+  ```
+
+- cp27-cp27mu
+
+  ```bash
+  export LD_LIBRARY_PATH=/opt/_internal/cpython-2.7.11-ucs4/lib:${LD_LIBRARY_PATH#/opt/_internal/cpython-2.7.11-ucs2/lib:}
+  export PATH=/opt/python/cp27-cp27mu/bin/:${PATH}
+  export PYTHON_FLAGS="-DPYTHON_EXECUTABLE:FILEPATH=/opt/python/cp27-cp27mu/bin/python
+        -DPYTHON_INCLUDE_DIR:PATH=/opt/python/cp27-cp27mu/include/python2.7
+        -DPYTHON_LIBRARIES:FILEPATH=/opt/_internal/cpython-2.7.11-ucs4/lib/libpython2.7.so"
+  ```
+
+And then add the `PYTHON_FLAGS` as your cmake flags:
+
+```bash
+cmake ..
+  ${PYTHON_FLAGS} \
+  -DWITH_GPU=OFF \
+  ...
+```
+
+You can find more details about cmake flags at [here](http://www.paddlepaddle.org/docs/develop/documentation/fluid/en/build_and_install/build_from_source_en.html#appendix-build-options)