parallel_executor.cc

/* Copyright (c) 2016 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/parallel_executor.h"

#include <algorithm>
#include <memory>
#include <string>
#include <tuple>
#include <utility>
#include <vector>

#include "paddle/fluid/framework/details/async_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/fast_threaded_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/multi_devices_helper.h"
#include "paddle/fluid/framework/details/op_handle_base.h"
#include "paddle/fluid/framework/details/parallel_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/scope_buffered_ssa_graph_executor.h"
#include "paddle/fluid/framework/details/threaded_ssa_graph_executor.h"
#include "paddle/fluid/framework/ir/graph.h"
#include "paddle/fluid/framework/ir/graph_helper.h"
#include "paddle/fluid/framework/ir/memory_optimize_pass/memory_optimization_var_info.h"
#include "paddle/fluid/framework/ir/memory_optimize_pass/reference_count_pass_helper.h"
#include "paddle/fluid/framework/ir/multi_devices_graph_pass/set_reader_device_info_utils.h"
#include "paddle/fluid/platform/event.h"
#include "paddle/fluid/platform/profiler.h"

#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cuda_device_guard.h"
#endif

DECLARE_double(eager_delete_tensor_gb);

#ifdef WITH_GPERFTOOLS
#include "gperftools/profiler.h"
#endif
DEFINE_string(pe_profile_fname, "",
              "Profiler filename for PE, which generated by gperftools."
              "Only valid when compiled `WITH_PRIFILER=ON`. Empty if disable.");
DEFINE_bool(enable_parallel_graph, false,
            "Force disable parallel graph execution mode if set false.");

namespace paddle {
namespace framework {

static std::once_flag gProfileOnce;
#ifdef WITH_GPERFTOOLS
static bool gProfileStarted = false;
#endif

#ifdef PADDLE_WITH_CUDA
std::once_flag p2p_init_flag;
#endif

using UseDevice = paddle::framework::details::ExecutionStrategy::UseDevice;

class ParallelExecutorPrivate {
 public:
  ParallelExecutorPrivate(const std::vector<platform::Place> &places,
                          Scope *global_scope)
      : places_(places), global_scope_(global_scope) {
    if (!FLAGS_pe_profile_fname.empty()) {
      std::call_once(gProfileOnce, [] {
#ifdef WITH_GPERFTOOLS
        ProfilerStart(FLAGS_pe_profile_fname.c_str());
        gProfileStarted = true;
#else
        LOG(WARNING) << "Paddle is not compiled with gperftools. "
          "FLAGS_pe_profile_fname will be ignored";
#endif
      });
    }
  }

  ~ParallelExecutorPrivate() {
    if (own_local_scope_) {
      for (size_t i = 1; i < local_scopes_.size(); ++i) {
        // Skip the first scope, since it is the global scope.
        Scope *local_scope = local_scopes_[i];
        if (global_scope_->HasKid(local_scope)) {
          global_scope_->DeleteScope(local_scope);
        }
      }
    }
  }

  bool IsUseCUDA(UseDevice use_device);

  void SetHasFeed(size_t dev_idx, bool has_feed = true);

  bool AllowPartialFeed() const;

  ir::Graph *ApplyMemoryOptimizePass(ir::Graph *graph);

  inline bool HasGarbageCollectors() const { return !gcs_.empty(); }

  /**
   * NOTE(zengjinle): the fed variables of users should not be reused,
   * because users may feed them into another network. Changing the fed
   * variables that users can visit may cause calculation wrong, which is
   * a very subtle bug when traning networks. However, these variables
   * can be garbage collected.
   *
   * ParallelExecutor provides 2 methods to feed variables:
   *
   *  - FeedTensorsIntoLocalScopes: this method would share memory of fed
   *                                variables, so we have to skip these.
   *
   *  - FeedAndSplitTensorIntoLocalScopes: this method would copy data of fed
   *                                       variables, so we do not need to skip
   *                                       them.
   */
  inline void SetSkipMemoryReuse(size_t scope_idx, const std::string &name) {
    if (mem_opt_var_infos_.size() == 0) {
      VLOG(4) << "The mem_opt_var_infos_ is empty, maybe no memory "
                 "optimization strategy is enabled";
      return;
    }
    auto iter = mem_opt_var_infos_[scope_idx].find(name);
    if (iter != mem_opt_var_infos_[scope_idx].end()) {
      iter->second->SetSkipMemoryReuse(true);
    }
  }

#if defined(PADDLE_WITH_NCCL)
  void InitNCCLCtxs(framework::Scope *scope, const BuildStrategy &bst) {
    VLOG(1) << "nccl comm num:" << bst.nccl_comm_num_ << ", nranks:" << nranks_
            << ", num_trainers:" << bst.num_trainers_
            << ", trainer_id:" << bst.trainer_id_;

    if (bst.use_hierarchical_allreduce_) {
      VLOG(1) << ", use_hierarchical_allreduce:"
              << bst.use_hierarchical_allreduce_ << ", inter_trainers_num:"
              << bst.hierarchical_allreduce_inter_nranks_
              << ", exter_trainers_num:"
              << bst.hierarchical_allreduce_exter_nranks_;
    }

    std::vector<ncclUniqueId *> flat_nccl_ids;
    if (nranks_ == 1) {
      // FIXME(gongwb): need not to create ncclid when nranks==1
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
      return;
    }

    if (bst.enable_parallel_graph_) {
      VLOG(1) << "use only one ncclid in pg model";

      ncclUniqueId *nccl_id = nullptr;

      std::string var_name = platform::GetFlatNCCLVarName(0);
      auto nccl_id_var = scope->FindVar(var_name);
      if (nccl_id_var) {
        nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
        VLOG(10) << "find nccl_id_var:" << var_name << ", nccl_id:" << nccl_id;
      } else {
        nccl_id = new ncclUniqueId();
        PADDLE_ENFORCE_EQ(
            platform::dynload::ncclGetUniqueId(nccl_id), ncclSuccess,
            platform::errors::PreconditionNotMet("Get NCCL unique ID failed."));
        VLOG(10) << "can't find nccl_id_var:" << var_name
                 << ", nccl_id:" << nccl_id;
      }

      flat_nccl_ids.push_back(nccl_id);

      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
      VLOG(1) << "init bst nccl context complete!";
      return;
    }

    // num_trainers ==1 && places > 1
    if (bst.num_trainers_ == 1) {
      nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                               bst.trainer_id_);
      return;
    }

    for (int i = 0; i < static_cast<int>(bst.nccl_comm_num_); i++) {
      std::string var_name = platform::GetFlatNCCLVarName(i);
      auto nccl_id_var = scope->FindVar(var_name);
      PADDLE_ENFORCE_NOT_NULL(
          nccl_id_var,
          platform::errors::NotFound("Can't find nccl_id_var '%s'.", var_name));
      auto nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
      flat_nccl_ids.push_back(nccl_id);
    }

    nccl_ctxs_->InitFlatCtxs(places_, flat_nccl_ids, bst.num_trainers_,
                             bst.trainer_id_);

    if (bst.use_hierarchical_allreduce_) {
      std::vector<ncclUniqueId *> inter_nccl_ids;
      for (int i = 0; i < static_cast<int>(bst.nccl_comm_num_); i++) {
        std::string var_name = platform::GetHierarchicalInterNCCLVarName(i);
        auto nccl_id_var = scope->FindVar(var_name);
        PADDLE_ENFORCE_NOT_NULL(nccl_id_var,
                                platform::errors::NotFound(
                                    "Can't find nccl_id_var '%s'.", var_name));
        auto inter_nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
        inter_nccl_ids.push_back(inter_nccl_id);
      }

      std::vector<ncclUniqueId *> exter_nccl_ids;
      for (int i = 0; i < static_cast<int>(bst.nccl_comm_num_); i++) {
        std::string var_name = platform::GetHierarchicalExterNCCLVarName(i);
        auto nccl_id_var = scope->FindVar(var_name);
        PADDLE_ENFORCE_NOT_NULL(nccl_id_var,
                                platform::errors::NotFound(
                                    "Can't find nccl_id_var '%s'.", var_name));
        auto nccl_id = nccl_id_var->GetMutable<ncclUniqueId>();
        exter_nccl_ids.push_back(nccl_id);
      }

      nccl_ctxs_->InitHierarchicalCtxs(
          places_, inter_nccl_ids, exter_nccl_ids, bst.num_trainers_,
          bst.trainer_id_, bst.hierarchical_allreduce_inter_nranks_,
          bst.hierarchical_allreduce_exter_nranks_);
    }
  }

  void InitOrGetNCCLCommunicator(framework::Scope *scope, BuildStrategy *bst) {
    const std::string var_name = "NCCLCommunicator";
    auto var = scope->FindVar(var_name);
    if (var != nullptr) {
      PADDLE_ENFORCE_EQ(var->IsInitialized(), true,
                        platform::errors::PreconditionNotMet(
                            "if %s exists, it must be initialized", var_name));
      VLOG(1) << "find " << var_name
              << " in scope, so use it and does not recreate!";
      nccl_ctxs_ = var->GetMutable<platform::NCCLCommunicator>();
      return;
    }

    if (bst->use_hierarchical_allreduce_) {
      PADDLE_ENFORCE_GT(
          bst->num_trainers_, 1,
          platform::errors::PreconditionNotMet(
              "The num_trainers should be greater than 1, but received %llu.",
              bst->num_trainers_));
      PADDLE_ENFORCE_GT(
          bst->hierarchical_allreduce_inter_nranks_, 1,
          platform::errors::PreconditionNotMet(
              "The inter_nranks should be greater than 1, but received %d.",
              bst->hierarchical_allreduce_inter_nranks_));
      PADDLE_ENFORCE_EQ(
          bst->num_trainers_ % bst->hierarchical_allreduce_inter_nranks_, 0,
          platform::errors::PreconditionNotMet(
              "num_trainers:%llu mod inter_nranks:%d != 0", bst->num_trainers_,
              bst->hierarchical_allreduce_inter_nranks_));

      bst->hierarchical_allreduce_exter_nranks_ =
          bst->num_trainers_ / bst->hierarchical_allreduce_inter_nranks_;
    }

    VLOG(1) << "not find " << var_name << " in scope, so recreate it!";
    nccl_ctxs_ = scope->Var(var_name)->GetMutable<platform::NCCLCommunicator>();
    InitNCCLCtxs(scope, *bst);
  }
#endif

  inline bool IsPersistable(const std::string &name) const {
    auto iter = is_persistable_.find(name);
    return iter != is_persistable_.end() && iter->second;
  }

  BuildStrategy build_strategy_;
  std::vector<platform::Place> places_;
  std::vector<Scope *> local_scopes_;
  std::vector<Scope *> local_exec_scopes_;
  Scope *global_scope_;  // not owned
  std::unique_ptr<details::SSAGraphExecutor> executor_;

  std::unordered_map<std::string, bool> is_persistable_;

#if defined(PADDLE_WITH_NCCL)
  platform::NCCLCommunicator *nccl_ctxs_{nullptr};
#endif
  bool own_local_scope_;
  UseDevice use_device_;
  bool use_all_reduce_;
  size_t nranks_;

  ir::MemOptVarInfoMapList mem_opt_var_infos_;
  ir::GarbageCollectorMap gcs_;

  details::ParallelSSAGraphExecutor *inference_executor_{nullptr};
};

bool ParallelExecutorPrivate::IsUseCUDA(UseDevice use_device) {
  return use_device == UseDevice::kCUDA;
}

void ParallelExecutorPrivate::SetHasFeed(size_t dev_idx, bool has_feed) {
  if (inference_executor_) {
    inference_executor_->SetHasFeed(dev_idx, has_feed);
  }
}

bool ParallelExecutorPrivate::AllowPartialFeed() const {
  return inference_executor_ && inference_executor_->SupportPartialFeed();
}

ir::Graph *ParallelExecutorPrivate::ApplyMemoryOptimizePass(ir::Graph *graph) {
  /**
   * NOTE(zengjinle): If BuildStrategy.memory_optimize = None in Python,
   * set BuildStrategy.memory_optimize according to whether gc is enabled.
   * If gc is enabled, BuildStrategy.memory_optimize = False.
   * If gc is disabled, BuildStrategy.memory_optimize = True.
   * This is because gc+memory_optimize is worse than gc only.
   *
   * As an option, users can enable BuildStrategy.memory_optimize forcely
   * by setting True, and disable it forcely by setting False.
   */
  bool is_gc_enabled = (GetEagerDeletionThreshold() >= 0);
  if (!build_strategy_.memory_optimize_) {
    build_strategy_.memory_optimize_ = !is_gc_enabled;
  }

  bool need_mem_opt = build_strategy_.enable_inplace_ ||
                      build_strategy_.enable_addto_ ||
                      build_strategy_.memory_optimize_.get() || is_gc_enabled;

  if (!need_mem_opt) return graph;

  std::vector<ir::LastLiveOpsOfVars> last_live_ops_of_vars;

  auto ref_cnt_pass = ir::PassRegistry::Instance().Get("reference_count_pass");
  ref_cnt_pass->SetNotOwned(ir::kMemOptVarInfoMapList, &mem_opt_var_infos_);
  ref_cnt_pass->SetNotOwned(ir::kLastLiveOpsOfVars, &last_live_ops_of_vars);
  graph = ref_cnt_pass->Apply(graph);
  VLOG(10) << "ReferenceCountPass Applied";

  if (build_strategy_.enable_addto_) {
    auto addto_pass = ir::PassRegistry::Instance().Get("inplace_addto_op_pass");
    addto_pass->SetNotOwned(ir::kMemOptVarInfoMapList, &mem_opt_var_infos_);
    addto_pass->SetNotOwned(ir::kLastLiveOpsOfVars, &last_live_ops_of_vars);
    addto_pass->Set(ir::kUseCuda, new bool(use_device_ == UseDevice::kCUDA));
    VLOG(10) << "Start to apply inplace_addto_op_pass";
    graph = addto_pass->Apply(graph);
    VLOG(10) << "inplace_addto_op_pass Applied";
  }

  if (build_strategy_.enable_inplace_) {
    auto inplace_pass =
        ir::PassRegistry::Instance().Get("buffer_shared_inplace_pass");
    inplace_pass->SetNotOwned(ir::kMemOptVarInfoMapList, &mem_opt_var_infos_);
    inplace_pass->SetNotOwned(ir::kLastLiveOpsOfVars, &last_live_ops_of_vars);
    inplace_pass->Set(ir::kUseCuda, new bool(use_device_ == UseDevice::kCUDA));
    VLOG(10) << "Start to apply buffer_shared_inplace_pass";
    graph = inplace_pass->Apply(graph);
    VLOG(10) << "buffer_shared_inplace_pass Applied";
    VLOG(1) << "Inplace strategy is enabled, when "
               "build_strategy.enable_inplace = True";
  }

  if (build_strategy_.memory_optimize_.get()) {
    auto cross_op_memory_reuse_pass = ir::PassRegistry::Instance().Get(
        "buffer_shared_cross_op_memory_reuse_pass");
    cross_op_memory_reuse_pass->SetNotOwned(ir::kMemOptVarInfoMapList,
                                            &mem_opt_var_infos_);
    cross_op_memory_reuse_pass->SetNotOwned(ir::kLastLiveOpsOfVars,
                                            &last_live_ops_of_vars);
    cross_op_memory_reuse_pass->Set(ir::kUseCuda,
                                    new bool(use_device_ == UseDevice::kCUDA));
    VLOG(10) << "Start to apply buffer_shared_cross_op_memory_reuse_pass";
    graph = cross_op_memory_reuse_pass->Apply(graph);
    VLOG(10) << "buffer_shared_cross_op_memory_reuse_pass Applied";
    LOG(INFO) << "Cross op memory reuse strategy is enabled, when "
                 "build_strategy.memory_optimize = True or garbage collection "
                 "strategy is disabled, which is not recommended";
  }

  if (!is_gc_enabled) {
    return graph;
  }
  size_t max_memory_size = static_cast<size_t>(GetEagerDeletionThreshold());

  for (size_t i = 0; i < places_.size(); ++i) {
    auto &place = places_[i];
    if (gcs_.count(place) > 0) {
      continue;
    }
    std::unique_ptr<GarbageCollector> gc;
    if (platform::is_gpu_place(place)) {
#ifdef PADDLE_WITH_CUDA
      if (IsFastEagerDeletionModeEnabled()) {
        gc.reset(new UnsafeFastGPUGarbageCollector(
            BOOST_GET_CONST(platform::CUDAPlace, place), max_memory_size));
      } else {
        gc.reset(new StreamGarbageCollector(
            BOOST_GET_CONST(platform::CUDAPlace, place), max_memory_size));
      }
      VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
#else
      PADDLE_THROW(platform::errors::PermissionDenied(
          "Paddle can't use CUDA device since it's not compiled with CUDA,"
          "Please recompile or reinstall Paddle with GPU support."));
#endif
    } else if (platform::is_xpu_place(place)) {
#if defined(PADDLE_WITH_XPU)
      gc.reset(new XPUGarbageCollector(
          BOOST_GET_CONST(platform::XPUPlace, place), max_memory_size));
      VLOG(10) << "Created " << i << "-th GarbageCollector at " << place;
#else
      PADDLE_THROW(platform::errors::PermissionDenied(
          "Paddle can't use XPU device since it's not compiled with XPU,"
          "Please recompile or reinstall Paddle with XPU support."));
#endif
    } else if (platform::is_cpu_place(place)) {
      gc.reset(new CPUGarbageCollector(
          BOOST_GET_CONST(platform::CPUPlace, place), max_memory_size));
      VLOG(10) << "Created GarbageCollector at " << place;
    } else {
      PADDLE_THROW(platform::errors::PreconditionNotMet(
          "Unsupported place for garbage collection"));
    }
    gcs_.emplace(place, std::move(gc));
  }

  if (!gcs_.empty()) {
    auto eager_deletion_pass =
        ir::PassRegistry::Instance().Get("eager_deletion_pass");
    eager_deletion_pass->SetNotOwned(ir::kMemOptVarInfoMapList,
                                     &mem_opt_var_infos_);
    eager_deletion_pass->SetNotOwned(ir::kGarbageCollector, &gcs_);
    eager_deletion_pass->SetNotOwned(ir::kLastLiveOpsOfVars,
                                     &last_live_ops_of_vars);
    eager_deletion_pass->SetNotOwned(ir::kAllPlaces, &places_);
    graph = eager_deletion_pass->Apply(graph);
    VLOG(10) << "EagerDeletionPass Applied";
    VLOG(1) << "Garbage collection strategy is enabled, when "
            << "FLAGS_eager_delete_tensor_gb = "
            << FLAGS_eager_delete_tensor_gb;
  }
  return graph;
}

class ResetHasFeedGuard {
 public:
  explicit ResetHasFeedGuard(ParallelExecutorPrivate *pe_member)
      : pe_member_(pe_member) {}

  ~ResetHasFeedGuard() {
    for (size_t i = 0; i < pe_member_->places_.size(); ++i) {
      pe_member_->SetHasFeed(i, false);
    }
  }

 private:
  ParallelExecutorPrivate *pe_member_;
};

size_t ParallelExecutor::DeviceCount() const { return member_->places_.size(); }

std::vector<Scope *> &ParallelExecutor::GetLocalScopes() {
  return member_->local_scopes_;
}

void ParallelExecutor::DropLocalExeScopes() {
  auto executor = dynamic_cast<details::ScopeBufferedSSAGraphExecutor *>(
      member_->executor_.get());
  if (executor) {
    executor->DropLocalExeScopes();
  }
}

bool ParallelExecutor::NeedCreateLocalExeScope() {
  auto executor = dynamic_cast<details::ScopeBufferedSSAGraphExecutor *>(
      member_->executor_.get());
  return executor && executor->NeedCreateLocalExeScope();
}

void InitP2P(const std::vector<platform::Place> &places) {
#ifdef PADDLE_WITH_CUDA
  std::call_once(p2p_init_flag, [&]() {
    int count = places.size();
    if (count <= 1) return;

    std::vector<int> devices;
    for (int i = 0; i < count; i++) {
      if (!is_gpu_place(places[i])) return;

      platform::CUDAPlace device =
          BOOST_GET_CONST(platform::CUDAPlace, places[i]);
      devices.push_back(device.GetDeviceId());
    }

    for (int i = 0; i < count; ++i) {
      for (int j = 0; j < count; ++j) {
        if (devices[i] == devices[j]) continue;
        int can_acess = -1;
        cudaError_t ret =
            cudaDeviceCanAccessPeer(&can_acess, devices[i], devices[j]);
        if (ret != cudaSuccess || can_acess != 1) {
          LOG(WARNING) << "Cannot enable P2P access from " << devices[i]
                       << " to " << devices[j];
        } else {
          platform::CUDADeviceGuard guard(devices[i]);
          cudaDeviceEnablePeerAccess(devices[j], 0);
        }
      }
    }
    VLOG(1) << "init p2p";
  });
#endif
}

ParallelExecutor::ParallelExecutor(const std::vector<platform::Place> &places,
                                   const std::vector<std::string> &bcast_vars,
                                   const std::string &loss_var_name,
                                   Scope *scope,
                                   const std::vector<Scope *> &local_scopes,
                                   const ExecutionStrategy &exec_strategy,
                                   const BuildStrategy &build_strategy,
                                   ir::Graph *graph)
    : member_(new ParallelExecutorPrivate(places, scope)) {
  InitP2P(places);
  ir::InitReaderQueueDeviceCount(graph, *(member_->global_scope_),
                                 member_->places_.size());
  member_->use_device_ = exec_strategy.use_device_;
  member_->build_strategy_ = build_strategy;
  member_->use_all_reduce_ = member_->build_strategy_.reduce_ ==
                             BuildStrategy::ReduceStrategy::kAllReduce;
  member_->nranks_ = build_strategy.num_trainers_ * places.size();
  if (!member_->use_all_reduce_ && member_->nranks_ == 1) {
    LOG(INFO) << "If you set build_strategy.reduce with 'Reduce',"
                 "the number of places should be greater than 1.";
    member_->build_strategy_.reduce_ =
        BuildStrategy::ReduceStrategy::kAllReduce;
    member_->use_all_reduce_ = true;
  }
#if defined(PADDLE_WITH_CUDA) && defined(_WIN32)
  if (member_->IsUseCUDA(member_->use_device_)) {
    PADDLE_ENFORCE_EQ(
        places.size(), 1,
        platform::errors::Unavailable("Windows can support Single GPU only."));
  }
#endif

#if defined(PADDLE_WITH_CUDA) && !defined(PADDLE_WITH_NCCL)
  if (member_->IsUseCUDA(member_->use_device_)) {
    PADDLE_ENFORCE_EQ(
        places.size(), 1,
        platform::errors::PermissionDenied(
            "Your machine has multiple cards, "
            "but the WITH_NCCL option is not turned on during compilation, "
            "and you cannot use multi-card training or prediction. "
            "Please recompile and turn on the WITH_NCCL option."));
  }
#endif

  std::string device_name;
  if (member_->use_device_ == UseDevice::kCPU) {
    device_name = "CPU";
  } else if (member_->use_device_ == UseDevice::kCUDA) {
    device_name = "CUDA";
  } else {
    device_name = "XPU";
  }

  VLOG(1) << string::Sprintf(
      "The Program will be executed on %s using ParallelExecutor, %lu "
      "cards are used, so %lu programs are executed in parallel.",
      device_name, places.size(), places.size());

  // Step 1. Bcast the bcast_vars to devs.
  // Create local scopes
  if (local_scopes.empty()) {
    member_->own_local_scope_ = true;
    member_->local_scopes_.emplace_back(member_->global_scope_);
    for (size_t i = 1; i < member_->places_.size(); ++i) {
      member_->local_scopes_.emplace_back(&scope->NewScope());
    }
  } else {
    member_->own_local_scope_ = false;
    PADDLE_ENFORCE_EQ(member_->places_.size(), local_scopes.size(),
                      platform::errors::PreconditionNotMet(
                          "member_->places_.size() = %d is not equal to "
                          "local_scopes.size() = %d",
                          member_->places_.size(), local_scopes.size()));
    for (size_t i = 0; i < member_->places_.size(); ++i) {
      member_->local_scopes_.emplace_back(&local_scopes[i]->NewScope());
    }
  }

  std::vector<ir::Graph *> graphs;
  if (member_->build_strategy_.async_mode_) {
    PADDLE_ENFORCE_EQ(member_->IsUseCUDA(member_->use_device_), false,
                      platform::errors::Unavailable(
                          "gpu mode does not support async_mode_ now!"));
    graphs.push_back(graph);
    for (size_t i = 1; i < places.size(); ++i) {
      auto *tmp_graph = new ir::Graph(graph->OriginProgram());
      async_graphs_.emplace_back(tmp_graph);
      graphs.push_back(tmp_graph);
    }
  }

  // FIXME(Yancey1989): parallel graph mode get better performance
  // in GPU allreduce distributed training. Need an elegant way to
  // choice the execution strategy.
  member_->build_strategy_.enable_parallel_graph_ =
      EnableParallelGraphExecution(*graph, exec_strategy,
                                   member_->build_strategy_);
  if (member_->build_strategy_.enable_parallel_graph_) {
    LOG(INFO) << "The Executor would execute the graph by ParallelGraph "
                 "Execution which can get better performance,"
              << "you can force it off by env FLAGS_enable_parallel_graph=0";
  }

  if (member_->IsUseCUDA(member_->use_device_) && member_->nranks_ > 1) {
#if defined(PADDLE_WITH_NCCL)
    member_->InitOrGetNCCLCommunicator(scope, &member_->build_strategy_);

    // Initialize device context's nccl comm, will be used by normal
    // Operators like sync_batch_norm, and collective ops.
    // NOTE: more than one ParallelExecutor with same place, the nccl comm will
    // be rewrite and there will be some problem.
    // NOTE: NCCL group-calls and non-group-calls can not use the same
    // NCCL communicator, so for ParallelGraph and Multi-Process mode, re-use
    // same communicators.
    auto *nccl_ctxs =
        member_->nccl_ctxs_->GetSyncBatchNormCtx(scope, member_->places_);
    auto &pool = platform::DeviceContextPool::Instance();
    for (size_t dev_id = 0; dev_id < member_->places_.size(); ++dev_id) {
      auto *dev_ctx = static_cast<platform::CUDADeviceContext *>(
          pool.Get(member_->places_[dev_id]));
      auto &nccl_ctx = nccl_ctxs->at(member_->places_[dev_id]);
      dev_ctx->set_nccl_comm(nccl_ctx.comm());
    }
#endif
  }
  // broadcast parameters from the 0th device to others:
  auto need_broadcast = [&]() -> bool {
    if (member_->build_strategy_.num_trainers_ > 1) {
      // 1. num_tariners would be grater than 1 for nccl distributed training.
      return true;
    } else if (member_->local_scopes_.size() != 1 && local_scopes.empty()) {
      // 2. Only one trainer process, but ParallelExecutor hold multiple
      // devices.
      return true;
    }
    return false;
  };
  // Bcast Parameters to all GPUs
  if (need_broadcast()) {
    BCastParamsToDevices(bcast_vars, member_->build_strategy_.trainer_id_);
  }

  // Startup Program has been run. All local scopes has correct parameters.

  // Step 2. Convert main_program to SSA form and dependency graph. Also, insert
  // ncclOp
  std::vector<ir::Graph *> async_graphs(places.size());
#if defined(PADDLE_WITH_NCCL)
  if (member_->build_strategy_.async_mode_) {
    VLOG(3) << "use local async mode";
    graph = member_->build_strategy_.Apply(
        graph, {member_->places_[0]}, loss_var_name,
        {member_->local_scopes_[0]}, 1,
        member_->IsUseCUDA(member_->use_device_), member_->nccl_ctxs_);
    for (size_t i = 1; i < member_->places_.size(); ++i) {
      graphs[i] = member_->build_strategy_.Apply(
          graphs[i], {member_->places_[i]}, loss_var_name,
          {member_->local_scopes_[i]}, 1,
          member_->IsUseCUDA(member_->use_device_), member_->nccl_ctxs_);
      async_graphs[i] = graphs[i];
    }
  } else {
    graph = member_->build_strategy_.Apply(
        graph, member_->places_, loss_var_name, member_->local_scopes_,
        member_->nranks_, member_->IsUseCUDA(member_->use_device_),
        member_->nccl_ctxs_);
  }
#else
  if (member_->build_strategy_.async_mode_) {
    VLOG(3) << "use local async mode";
    graph = member_->build_strategy_.Apply(
        graph, {member_->places_[0]}, loss_var_name,
        {member_->local_scopes_[0]}, 1,
        member_->IsUseCUDA(member_->use_device_));
    for (size_t i = 1; i < member_->places_.size(); ++i) {
      graphs[i] = member_->build_strategy_.Apply(
          graphs[i], {member_->places_[i]}, loss_var_name,
          {member_->local_scopes_[i]}, 1,
          member_->IsUseCUDA(member_->use_device_));
      async_graphs[i] = graphs[i];
    }
  } else {
    graph = member_->build_strategy_.Apply(
        graph, member_->places_, loss_var_name, member_->local_scopes_,
        member_->nranks_, member_->IsUseCUDA(member_->use_device_));
  }
#endif

  graph = member_->ApplyMemoryOptimizePass(graph);

  async_graphs[0] = graph;

  // Step 3. Create vars in each scope. Passes may also create new vars.
  //         skip control vars and empty vars
  std::vector<details::VariableInfo> var_infos;
  for (auto &node : graph->Nodes()) {
    if (node->IsVar() && !node->IsCtrlVar() && node->Var()) {
      var_infos.emplace_back();
      var_infos.back().name_ = node->Var()->Name();
      var_infos.back().type_ = node->Var()->GetType();
      var_infos.back().persistable_ = node->Var()->Persistable();

      member_->is_persistable_.emplace(node->Var()->Name(),
                                       node->Var()->Persistable());
    }
  }

  std::unordered_map<Scope *, Scope *> scope_map;
  for (auto *scope : member_->local_scopes_) {
    auto &local_exec_scope = scope->NewScope();
    member_->local_exec_scopes_.emplace_back(&local_exec_scope);
    scope_map.emplace(scope, &local_exec_scope);
  }

  PADDLE_ENFORCE_EQ(
      member_->local_scopes_.size(), member_->local_exec_scopes_.size(),
      platform::errors::PreconditionNotMet(
          "member_->local_scopes_.size() = %d is not equal to "
          "member_->local_exec_scopes_.size() = %d",
          member_->local_scopes_.size(), member_->local_exec_scopes_.size()));

  std::vector<ir::Graph *> final_graphs;

  if (member_->build_strategy_.async_mode_) {
    VLOG(3) << "use AsyncSSAGraphExecutor";
    member_->executor_.reset(new details::AsyncSSAGraphExecutor(
        exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
        member_->places_, async_graphs));
    final_graphs = async_graphs;
  } else if (member_->build_strategy_.enable_parallel_graph_) {
    VLOG(3) << "use ParallelSSAGraphExecutor";
#ifdef PADDLE_WITH_CUDA
    // TODO(Yancey1989): Remove passing in the main_program when
    // allreduce_seq_pass doesn't need it as the attr.
    bool is_inference = details::IsDataParallelInferenceGraph(*graph);
    bool has_drop_last_read_op = details::HasDropLastReadOp(*graph);

    auto *pg_exe = new details::ParallelSSAGraphExecutor(
        exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
        member_->places_, graph);
    final_graphs = pg_exe->Graphs();
    member_->executor_.reset(pg_exe);

    if (is_inference && member_->places_.size() > 1) {
      member_->inference_executor_ = pg_exe;
      if (!has_drop_last_read_op) {
        VLOG(5) << "Enable partial feed support in inference phase";
        pg_exe->EnablePartialFeedSupport();
      }
    }
#else
    PADDLE_THROW(platform::errors::PreconditionNotMet(
        "Paddle should be compiled with CUDA for ParallelGraph Execution."));
#endif
  } else {
    bool has_drop_last_read_op = details::HasDropLastReadOp(*graph);
    auto possible_inference_graphs =
        details::TrySeparateToMultipleSingleDeviceGraphs(graph);
    if (!possible_inference_graphs.empty()) {
      VLOG(5) << "Use ParallelSSAGraphExecutor in inference phase";
      auto *pg_exe = new details::ParallelSSAGraphExecutor(
          exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
          member_->places_, std::move(possible_inference_graphs));
      if (!has_drop_last_read_op) {
        VLOG(5) << "Enable partial feed support in inference phase";
        pg_exe->EnablePartialFeedSupport();
      }
      final_graphs = pg_exe->Graphs();
      member_->executor_.reset(pg_exe);
      member_->inference_executor_ = pg_exe;
    } else {
      LOG_IF(WARNING, details::HasKeepLastReadOp(*graph))
          << "drop_last=False for DataLoader is not supported in training "
             "network. It is automatically turned to drop_last=True.";
      if (exec_strategy.type_ == ExecutionStrategy::kDefault) {
        VLOG(3) << "use ThreadedSSAGraphExecutor";
        member_->executor_.reset(new details::ThreadedSSAGraphExecutor(
            exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
            member_->places_, graph));
      } else {
        VLOG(3) << "use FastThreadedSSAGraphExecutor";
        member_->executor_.reset(new details::FastThreadedSSAGraphExecutor(
            exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
            member_->places_, graph));
      }
      final_graphs.emplace_back(graph);
    }
  }

  VLOG(3) << "use ScopeBufferedSSAGraphExecutor";
  if (!member_->build_strategy_.async_mode_) {
    member_->executor_.reset(new details::ScopeBufferedSSAGraphExecutor(
        exec_strategy, member_->local_scopes_, member_->local_exec_scopes_,
        std::move(var_infos), member_->places_, std::move(member_->executor_)));
  }

  for (auto *g : final_graphs) {
    auto ops = ir::FilterByNodeWrapper<details::OpHandleBase>(*g);
    for (auto *op : ops) {
      op->SetLocalExecScopes(scope_map);
    }
  }

  if (final_graphs.size() == 1) {
    ir::SetReaderOpDeviceInfo(final_graphs[0], member_->places_.size());
  } else {
    for (size_t i = 0; i < final_graphs.size(); ++i) {
      ir::SetReaderOpDeviceInfo(final_graphs[i], member_->places_.size(), i);
    }
  }
}

void ParallelExecutor::BCastParamsToDevices(
    const std::vector<std::string> &vars, int trainer_id) const {
  VLOG(3) << "BCastParamsToDevices";
  // the initializing bcast, all vars would be bcast from device(0).
  for (auto &var : vars) {
    framework::Variable *main_var = member_->local_scopes_[0]->FindVar(var);
    if (main_var == nullptr || !main_var->IsType<LoDTensor>()) {
      continue;
    }

    auto &main_tensor = main_var->Get<LoDTensor>();
    if (!main_tensor.IsInitialized()) {
      VLOG(3) << "one in var not inited, return!";
      continue;
    }
    auto &dims = main_tensor.dims();
    if (paddle::platform::is_gpu_place(main_tensor.place())) {
#if defined(PADDLE_WITH_NCCL)
      std::vector<void *> buffers;
      buffers.reserve(member_->places_.size());
      size_t numel = main_tensor.numel();
      ncclDataType_t data_type = platform::ToNCCLDataType(main_tensor.type());
      for (size_t i = 0; i < member_->places_.size(); ++i) {
        auto place = member_->places_[i];
        void *buffer;

        if (i == 0 && trainer_id == 0) {
          buffer = const_cast<void *>(main_tensor.data<void>());
        } else {
          auto local_scope = member_->local_scopes_[i];
          auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();
          t->Resize(dims);
          buffer = t->mutable_data(place, main_tensor.type());
        }
        buffers.push_back(buffer);
      }

      PADDLE_ENFORCE_EQ(member_->places_.size(), buffers.size(),
                        platform::errors::PreconditionNotMet(
                            "variables' buffer size to bcast is %d, which is "
                            "NOT equal to places size %d",
                            buffers.size(), member_->places_.size()));
      {
        auto *nccl_ctxs = member_->nccl_ctxs_->DefaultFlatCtx();
        platform::NCCLGroupGuard guard;
        for (size_t i = 0; i < member_->places_.size(); ++i) {
          auto &nccl_ctx = nccl_ctxs->at(member_->places_[i]);
          platform::dynload::ncclBcast(buffers[i], numel, data_type, 0,
                                       nccl_ctx.comm_, nccl_ctx.stream());
        }
        nccl_ctxs->WaitAll();
      }
#endif
    } else {
      platform::CPUPlace cpu;
      for (size_t i = 1; i < member_->places_.size(); ++i) {
        auto local_scope = member_->local_scopes_[i];
        auto *t = local_scope->Var(var)->GetMutable<LoDTensor>();

        auto copy_memory = [&] {
          t->Resize(dims);
          t->mutable_data(cpu, main_tensor.type());
          paddle::framework::TensorCopy(main_tensor, cpu, t);
        };

        auto share_memory = [&] { t->ShareDataWith(main_tensor); };

        // FIXME(zcd): LR_DECAY_COUNTER should not be shared. This is a hot fix.
        if (member_->build_strategy_.async_mode_) {
          share_memory();
        } else if (member_->use_all_reduce_ ||
                   member_->IsUseCUDA(member_->use_device_) ||
                   var == "@LR_DECAY_COUNTER@") {
          copy_memory();
        } else {
          share_memory();
        }
      }
    }
  }
}

FetchResultType ParallelExecutor::Run(
    const std::vector<std::string> &fetch_tensors, bool return_merged) {
  VLOG(3) << "enter ParallelExecutor Run";
  platform::RecordEvent parallel_executor_event(
      "ParallelExecutor::Run", paddle::platform::EventRole::kSpecial);
#ifdef WITH_GPERFTOOLS
  if (gProfileStarted) {
    ProfilerFlush();
  }
#endif

  platform::RecordBlock b(0);

  ResetHasFeedGuard reset_has_feed_guard(member_);

  ir::SkipMemOptVarsGuard guard(&(member_->mem_opt_var_infos_), fetch_tensors,
                                member_->HasGarbageCollectors());

  VLOG(3) << "ParallelExecutor begin to run member_->executor_->Run";
  auto fetch_data = member_->executor_->Run(fetch_tensors, return_merged);
  return fetch_data;
}

void ParallelExecutor::FeedTensorsIntoLocalScopes(
    const std::vector<std::unordered_map<std::string, LoDTensor>> &tensors) {
  if (!member_->AllowPartialFeed()) {
    PADDLE_ENFORCE_EQ(tensors.size(), member_->local_scopes_.size(),
                      platform::errors::Unimplemented(
                          "The feed data number %d does not match the device "
                          "number %d. If you are using DataLoader to feed "
                          "data, this may be because you set drop_last=False "
                          "in training network. Currently, drop_last=False for "
                          "DataLoader is not supported for training network. "
                          "Please set drop_last=True when defining DataLoader.",
                          tensors.size(), member_->local_scopes_.size()));
  } else {
    PADDLE_ENFORCE_GE(member_->local_scopes_.size(), tensors.size(),
                      platform::errors::InvalidArgument(
                          "The feed tensor number exceeds the device number"));
  }

  size_t feed_num = 0;
  for (size_t i = 0; i < tensors.size(); ++i) {
    auto &map = tensors[i];
    if (map.empty()) {
      continue;
    }

    member_->SetHasFeed(i);
    ++feed_num;
    for (auto &pair : map) {
      bool is_persistable = member_->IsPersistable(pair.first);
      if (!is_persistable) {
        member_->SetSkipMemoryReuse(i, pair.first);
      }
      auto *feed_scope = is_persistable ? member_->local_scopes_[i]
                                        : member_->local_exec_scopes_[i];
      auto *feed_var = feed_scope->Var(pair.first);

      auto *trg = feed_var->GetMutable<LoDTensor>();
      trg->ShareDataWith(pair.second);
      trg->set_lod(pair.second.lod());
    }
  }

  if (!member_->AllowPartialFeed()) {
    PADDLE_ENFORCE_EQ(feed_num, member_->local_scopes_.size(),
                      platform::errors::Unimplemented(
                          "The feed data number %d does not match the device "
                          "number %d. If you are using DataLoader to feed "
                          "data, this may be because you set drop_last=False "
                          "in training network. Currently, drop_last=False for "
                          "DataLoader is not supported for training network. "
                          "Please set drop_last=True when defining DataLoader.",
                          feed_num, member_->local_scopes_.size()));
  }
}

void ParallelExecutor::FeedAndSplitTensorIntoLocalScopes(
    const std::unordered_map<std::string, LoDTensor> &tensors) {
  size_t num_places = member_->places_.size();
  bool allow_partial_feed = member_->AllowPartialFeed();

  size_t persistable_feed_len = -1UL;
  size_t non_persistable_feed_len = -1UL;

  for (auto &pair : tensors) {
    bool is_persistable = member_->IsPersistable(pair.first);
    VLOG(3) << "Split " << (is_persistable ? "persistable" : "no persistable")
            << " data (" << pair.first << "), dim:" << pair.second.dims()
            << ", place: " << pair.second.place();
    auto lod_tensors = pair.second.SplitLoDTensor(member_->places_);
    bool is_cpu_place = platform::is_cpu_place(member_->places_.front());
    if (!is_persistable && num_places != lod_tensors.size() &&
        !allow_partial_feed) {
      auto error_info = string::Sprintf(
          "The number(%d) of samples[%s] of current batch is less than the "
          "count(%d) of devices(%s), currently, it is not allowed. ",
          lod_tensors.size(), pair.first, num_places,
          (is_cpu_place ? "CPU" : "GPU"));
      if (is_cpu_place) {
        error_info +=
            "You should set the environment variable CPU_NUM in the system "
            "to determine the number of devices you need.";
      }
      PADDLE_THROW(platform::errors::PreconditionNotMet(error_info));
    } else if (is_persistable) {
      if (lod_tensors.size() == 1) {
        lod_tensors.reserve(num_places);
        auto &tensor = lod_tensors.front();
        PADDLE_ENFORCE_EQ(
            tensor.dims(), pair.second.dims(),
            platform::errors::PreconditionNotMet("The dim doesn't match."));
        PADDLE_ENFORCE_EQ(
            tensor.place(), member_->places_.at(0),
            platform::errors::PreconditionNotMet("The place doesn't match."));
        for (size_t i = 1; i < num_places; ++i) {
          lod_tensors.emplace_back();
          auto &tmp = lod_tensors.back();
          framework::TensorCopy(pair.second, member_->places_.at(i), &tmp);
        }
      }
      if (lod_tensors.size() != num_places && !allow_partial_feed) {
        auto error_info = string::Sprintf(
            "The number(%d) of samples[%s] of the current batch does not match "
            "the count(%d) of devices(%s). Because that %s is a persistable "
            "variable, you can feed just one sample, in that case, the input "
            "sample will be copied in %d copies and be sent to different "
            "places separately. If you need that different place has different "
            "value, you should feed %d samples.",
            lod_tensors.size(), pair.first, num_places,
            (is_cpu_place ? "CPU" : "GPU"), pair.first, num_places, num_places);
        PADDLE_THROW(platform::errors::PreconditionNotMet(error_info));
      }
    }

    if (allow_partial_feed) {
      if (is_persistable) {
        if (persistable_feed_len == -1UL) {
          persistable_feed_len = lod_tensors.size();
        } else {
          PADDLE_ENFORCE_EQ(
              persistable_feed_len, lod_tensors.size(),
              platform::errors::InvalidArgument(
                  "The feeded number of different persistable variables "
                  "should be the same"));
        }
      } else {
        if (non_persistable_feed_len == -1UL) {
          non_persistable_feed_len = lod_tensors.size();
        } else {
          PADDLE_ENFORCE_EQ(
              non_persistable_feed_len, lod_tensors.size(),
              platform::errors::InvalidArgument(
                  "The feeded number of different non-persistable variables "
                  "should be the same"));
        }
      }
    }

    for (size_t j = 0; j < lod_tensors.size(); ++j) {
      auto *feed_scope = is_persistable ? member_->local_scopes_[j]
                                        : member_->local_exec_scopes_[j];
      auto *feed_var = feed_scope->Var(pair.first);

      auto t = feed_var->GetMutable<LoDTensor>();
      t->ShareDataWith(lod_tensors[j]);
      t->set_lod(lod_tensors[j].lod());
    }
  }

  if (allow_partial_feed && persistable_feed_len != -1UL &&
      non_persistable_feed_len != -1UL) {
    VLOG(10) << "Persistable len " << persistable_feed_len;
    VLOG(10) << "Non persistable len " << non_persistable_feed_len;
    PADDLE_ENFORCE_GE(persistable_feed_len, non_persistable_feed_len,
                      platform::errors::InvalidArgument(
                          "The feeded number of persistable variables should "
                          "not be less than non-persistable variables"));
  }

  if (non_persistable_feed_len != -1UL) {
    for (size_t i = 0; i < non_persistable_feed_len; ++i) {
      member_->SetHasFeed(i);
    }
  }
}

ParallelExecutor::~ParallelExecutor() {
  for (auto &p : member_->places_) {
    platform::DeviceContextPool::Instance().Get(p)->Wait();
  }
  delete member_;
}

bool ParallelExecutor::EnableParallelGraphExecution(
    const ir::Graph &graph, const ExecutionStrategy &exec_strategy,
    const BuildStrategy &build_strategy) const {
  if (!FLAGS_enable_parallel_graph) {
    return false;
  }

  bool enable_parallel_graph = true;

  for (ir::Node *node : graph.Nodes()) {
    if (node->IsVar() && node->Var()) {
      // TODO(Yancey1989): support sparse update in ParallelGraph mode.
      if (node->Var()->GetType() == proto::VarType::SELECTED_ROWS) {
        enable_parallel_graph = false;
        break;
      }
    } else if (node->IsOp() && node->Op()) {
      // TODO(Yancey1989): support pserver mode
      if (node->Op()->Type() == "send" || node->Op()->Type() == "recv") {
        enable_parallel_graph = false;
        break;
      }
    }
  }

  if (!member_->use_all_reduce_ || !member_->IsUseCUDA(member_->use_device_)) {
    if (build_strategy.enable_sequential_execution_ ||
        exec_strategy.type_ == ExecutionStrategy::ExecutorType::kExperimental) {
      enable_parallel_graph = false;
    }
  }

#ifdef WIN32
  VLOG(1) << "Windows has no support to parallel graph, enable_parallel_graph "
             "would be forced to false.";
  enable_parallel_graph = false;
#endif

  return enable_parallel_graph;
}

const ir::Graph &ParallelExecutor::Graph() const {
  return member_->executor_->Graph();
}

}  // namespace framework
}  // namespace paddle

USE_PASS(reference_count_pass);
USE_PASS(eager_deletion_pass);
USE_PASS(buffer_shared_inplace_pass);
USE_PASS(buffer_shared_cross_op_memory_reuse_pass);
USE_PASS(inplace_addto_op_pass);