enhance gc

test=develop

paddle/fluid/framework/CMakeLists.txt

@@ -174,7 +174,7 @@ else()
   cc_test(test_naive_executor SRCS naive_executor_test.cc DEPS naive_executor elementwise_add_op)
 endif()
 
-target_link_libraries(executor garbage_collector)
+target_link_libraries(executor garbage_collector while_op_helper)
 
 cc_library(parallel_executor SRCS parallel_executor.cc DEPS
         threaded_ssa_graph_executor scope_buffered_ssa_graph_executor parallel_ssa_graph_executor

paddle/fluid/framework/details/CMakeLists.txt

@@ -61,7 +61,8 @@ cc_library(inplace_op_pass SRCS inplace_op_pass.cc DEPS memory_optimize_pass op_
 cc_library(modify_op_lock_and_record_event_pass SRCS modify_op_lock_and_record_event_pass.cc DEPS computation_op_handle op_graph_view multi_devices_helper)
 cc_library(reference_count_pass_helper SRCS reference_count_pass_helper.cc DEPS garbage_collector computation_op_handle)
 cc_library(eager_deletion_op_handle SRCS eager_deletion_op_handle.cc DEPS lod_tensor selected_rows reference_count_pass_helper)
-cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass)
+cc_library(while_op_eager_deletion_pass SRCS while_op_eager_deletion_pass.cc DEPS while_op_helper graph_helper pass computation_op_handle)
+cc_library(eager_deletion_pass SRCS eager_deletion_pass.cc DEPS computation_op_handle eager_deletion_op_handle graph graph_helper pass while_op_eager_deletion_pass)
 cc_library(reference_count_pass SRCS reference_count_pass.cc DEPS computation_op_handle graph graph_helper pass op_graph_view reference_count_pass_helper)
 
 cc_library(sequential_execution_pass SRCS sequential_execution_pass.cc DEPS graph graph_helper pass)

paddle/fluid/framework/details/computation_op_handle.h

@@ -31,6 +31,8 @@ class ComputationOpHandle : public OpHandleBase {
   ComputationOpHandle(ir::Node *node, Scope *scope, platform::Place place,
                       size_t scope_idx);
 
+  OperatorBase *GetOp() { return op_.get(); }
+
   std::string Name() const override;
 
   const Scope *GetScope() const { return scope_; }

paddle/fluid/framework/details/eager_deletion_op_handle.cc

@@ -25,8 +25,6 @@ namespace paddle {
 namespace framework {
 namespace details {
 
-static const std::string kEagerDeletionOpName{"eager_deletion"};  // NOLINT
-
 EagerDeletionOpHandle::EagerDeletionOpHandle(
     ir::Node *node, const Scope *scope, const platform::Place &place,
     const std::unordered_set<std::string> &var_names, GarbageCollector *gc,
@@ -61,10 +59,9 @@ EagerDeletionOpHandle::~EagerDeletionOpHandle() {
 #endif
 }
 
-std::string EagerDeletionOpHandle::Name() const { return kEagerDeletionOpName; }
+std::string EagerDeletionOpHandle::Name() const { return "eager_deletion"; }
 
 void EagerDeletionOpHandle::RunImpl() {
-  platform::RecordEvent event(kEagerDeletionOpName, nullptr);
   Scope *exec_scope = nullptr;
   std::deque<std::shared_ptr<memory::Allocation>> garbages;
   for (auto &name : var_names_) {

paddle/fluid/framework/details/eager_deletion_pass.cc

@@ -21,35 +21,42 @@
 
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
-#include "paddle/fluid/framework/details/eager_deletion_pass.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 
-DEFINE_double(fraction_of_eager_deletion, 1.0, "Fraction of eager deletion");
-DEFINE_bool(eager_delete_tensor_only, false, "");
+DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
+              "Fraction of eager deletion");
 
 namespace paddle {
 namespace framework {
 namespace details {
 
-namespace {  // NOLINT
+// op -> variables which can be deleted after op runs
 using OpToVarNameSetMap =
     std::unordered_map<ComputationOpHandle *, std::unordered_set<std::string>>;
-}  // NOLINT
 
+// Check whether the variable is LoDTensor based on static VarDesc info
 static bool IsLoDTensor(VarDesc *var) {
   return var->Proto()->type().type() == proto::VarType::LOD_TENSOR;
 }
 
-static int64_t GetNumel(const GraphVars &vars, const std::string &var_name,
-                        size_t scope_idx) {
-  auto *var_desc = TryGetLatestVarDesc(vars[scope_idx].at(var_name));
+// Get memory size of LoDTensor
+static int64_t GetMemorySize(
+    const std::unordered_map<std::string, std::vector<VarHandle *>> &vars,
+    const std::string &var_name) {
+  auto *var_desc = TryGetLatestVarDesc(vars.at(var_name));
+  PADDLE_ENFORCE_NOT_NULL(var_desc);
   PADDLE_ENFORCE(IsLoDTensor(var_desc));
   auto dims = var_desc->GetShape();
-  return std::accumulate(dims.begin(), dims.end(), static_cast<int64_t>(1),
+  return SizeOfType(var_desc->GetDataType()) *
+         std::accumulate(dims.begin(), dims.end(), static_cast<int64_t>(1),
                          std::multiplies<int64_t>());
 }
 
+// Split all variables in the graph into LoDTensor and Non-LoDTensor (e.g.
+// SelectedRows, LoDTensorArray)
+// Since partial GC is based on static analysis of memory size of each variable
+// So we should skip SelectedRows and LoDTensorArray here
 static void SplitIntoLoDTensorAndNonLoDTensorVars(
     const OpToVarNameSetMap &m, const GraphVars &vars,
     OpToVarNameSetMap *lod_tensors, OpToVarNameSetMap *other_vars) {
@@ -69,76 +76,106 @@ static void SplitIntoLoDTensorAndNonLoDTensorVars(
   }
 }
 
-static OpToVarNameSetMap ShrinkGCVars(const OpToVarNameSetMap &m,
-                                      const GraphVars &vars,
-                                      double fraction_of_memory_size,
-                                      bool delete_lod_tensor_only = false) {
-  // Do not perform gc
+struct GCVarInfo {
+  GCVarInfo(const std::string &name, int64_t memory_size,
+            ComputationOpHandle *op, size_t scope_idx)
+      : name_(name),
+        memory_size_(memory_size),
+        op_(op),
+        scope_idx_(scope_idx) {}
+
+  std::string name_;         // variable name
+  int64_t memory_size_;      // memory size
+  ComputationOpHandle *op_;  // op after which the variable could be deleted
+  size_t scope_idx_;         // scope index where the variable locates
+
+  int64_t AbsMemorySize() const { return std::abs(memory_size_); }
+};
+
+// Delete delete_lod_tensor_only is not used currently
+static OpToVarNameSetMap ShrinkGCVars(
+    const OpToVarNameSetMap &m, const GraphVars &vars,
+    const std::vector<platform::Place> &places, double fraction_of_memory_size,
+    bool delete_lod_tensor_only = false) {
+  // Do not perform gc when fraction_of_memory_size = 0
   if (fraction_of_memory_size <= 0.0) return {};
 
-  // Perform complete gc
+  /**
+   * Step 1: Split all variables into LoDTensor and Non-LoDTensor.
+   * We can only calculate memory size of LoDTensors
+   */
+  OpToVarNameSetMap lod_tensors, other_vars;
+  SplitIntoLoDTensorAndNonLoDTensorVars(m, vars, &lod_tensors, &other_vars);
+
+  // Perform complete gc when fraction_of_memory_size >= 1
   if (fraction_of_memory_size >= 1.0) {
-    if (delete_lod_tensor_only) {
-      OpToVarNameSetMap lod_tensors, other_vars;
-      SplitIntoLoDTensorAndNonLoDTensorVars(m, vars, &lod_tensors, &other_vars);
-      return lod_tensors;
-    } else {
-      return m;
-    }
+    return delete_lod_tensor_only ? lod_tensors : m;
   }
 
-  // Perform partial gc
-  OpToVarNameSetMap lod_tensors, other_vars;
-  SplitIntoLoDTensorAndNonLoDTensorVars(m, vars, &lod_tensors, &other_vars);
+  /**
+   * Step 2: build GCVarInfos, and calculate total memory sizes of each device
+   */
 
-  using TupleType = std::tuple<std::string, ComputationOpHandle *, int64_t>;
+  // place -> variable info (name, memory size, place, scope_idx)
+  std::map<platform::Place, std::vector<GCVarInfo>> place_to_vars;
 
-  std::unordered_map<size_t, std::vector<TupleType>> place_to_vars;
-  std::unordered_map<size_t, int64_t> total_memory_size;
+  // place -> total memory sizes
+  std::map<platform::Place, int64_t> place_to_size;
   for (auto &op_vars_pair : lod_tensors) {
-    auto scope_idx = op_vars_pair.first->GetScopeIdx();
-    int64_t size = 0;
-    for (auto &var_name : op_vars_pair.second) {
-      auto var_size = GetNumel(vars, var_name, scope_idx);
-      size += std::abs(var_size);
-      place_to_vars[scope_idx].emplace_back(var_name, op_vars_pair.first,
-                                            var_size);
+    auto *op = op_vars_pair.first;
+    auto &var_names = op_vars_pair.second;
+    auto scope_idx = op->GetScopeIdx();
+    auto &place = places[scope_idx];
+
+    for (auto &var_name : var_names) {
+      auto var_size = GetMemorySize(vars[scope_idx], var_name);
+      GCVarInfo var_info(var_name, var_size, op, scope_idx);
+      place_to_size[place] += var_info.AbsMemorySize();
+      place_to_vars[place].emplace_back(std::move(var_info));
     }
-    total_memory_size.emplace(scope_idx, size);
   }
 
-  for (auto &pair : place_to_vars) {
-    std::sort(pair.second.begin(), pair.second.end(),
-              [](const TupleType &t1, const TupleType &t2) {
-                return std::abs(std::get<2>(t1)) > std::abs(std::get<2>(t2));
+  /**
+   * Step 3: sort GCVarInfos, and only delete the largest variables.
+   */
+  OpToVarNameSetMap partial_vars;
+  for (auto &place_to_var_pair : place_to_vars) {
+    auto &place = place_to_var_pair.first;
+    auto &gc_vars = place_to_var_pair.second;
+    std::sort(gc_vars.begin(), gc_vars.end(),
+              [](const GCVarInfo &var1, const GCVarInfo &var2) {
+                return var1.AbsMemorySize() > var2.AbsMemorySize();
               });
-  }
 
-  OpToVarNameSetMap ret;
-  for (auto &pair : place_to_vars) {
-    auto desired_delete_size = static_cast<int64_t>(
-        fraction_of_memory_size * total_memory_size.at(pair.first));
-    int64_t cur_size = 0;
-    for (size_t i = 0; i < pair.second.size() && cur_size < desired_delete_size;
+    int64_t accumulated_size = 0;
+    int64_t size_threshold =
+        static_cast<int64_t>(fraction_of_memory_size * place_to_size[place]);
+    for (size_t i = 0; i < gc_vars.size() && accumulated_size < size_threshold;
          ++i) {
-      auto &var_name = std::get<0>(pair.second[i]);
-      auto *op = std::get<1>(pair.second[i]);
-      cur_size += std::get<2>(pair.second[i]);
-      ret[op].insert(var_name);
+      partial_vars[gc_vars[i].op_].insert(gc_vars[i].name_);
+      accumulated_size += gc_vars[i].AbsMemorySize();
     }
   }
 
+  /**
+   * Step 4: Combine other vars (SelectedRows, LoDTensorArray)
+   */
   if (!delete_lod_tensor_only) {
     for (auto &op_vars_pair : other_vars) {
-      for (auto &var_name : op_vars_pair.second) {
-        ret[op_vars_pair.first].insert(var_name);
-      }
+      partial_vars[op_vars_pair.first].insert(op_vars_pair.second.begin(),
+                                              op_vars_pair.second.end());
     }
   }
 
-  return ret;
+  return partial_vars;
 }
 
+class EagerDeletionPass : public ir::Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override;
+};
+
 std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     std::unique_ptr<ir::Graph> graph) const {
   auto &ref_cnts =
@@ -166,9 +203,8 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     }
   }
 
-  op_vars_map =
-      ShrinkGCVars(op_vars_map, vars, FLAGS_fraction_of_eager_deletion,
-                   FLAGS_eager_delete_tensor_only);
+  op_vars_map = ShrinkGCVars(op_vars_map, vars, places,
+                             FLAGS_memory_fraction_of_eager_deletion);
 
   for (auto &pair : op_vars_map) {
     auto *op = pair.first;
@@ -200,12 +236,13 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     eager_deletion_op->AddOutput(dummy_leaf);
   }
 
-  VLOG(10) << "FLAGS_fraction_of_eager_deletion = "
-           << FLAGS_fraction_of_eager_deletion;
-  VLOG(10) << "FLAGS_eager_delete_tensor_only = "
-           << FLAGS_eager_delete_tensor_only;
+  VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = "
+           << FLAGS_memory_fraction_of_eager_deletion;
   VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
-  return graph;
+
+  auto while_op_eager_deletion_pass =
+      ir::PassRegistry::Instance().Get("while_op_eager_deletion_pass");
+  return while_op_eager_deletion_pass->Apply(std::move(graph));
 }
 
 }  // namespace details
@@ -218,3 +255,5 @@ REGISTER_PASS(eager_deletion_pass,
     .RequirePassAttr(paddle::framework::details::kLastLiveOpsOfVars)
     .RequirePassAttr(paddle::framework::details::kAllPlaces)
     .RequirePassAttr(paddle::framework::details::kGarbageCollector);
+
+USE_PASS(while_op_eager_deletion_pass);

paddle/fluid/framework/details/eager_deletion_pass.h

@@ -1,32 +0,0 @@
-// 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
-
-#include "paddle/fluid/framework/ir/graph.h"
-#include "paddle/fluid/framework/ir/pass.h"
-
-namespace paddle {
-namespace framework {
-namespace details {
-
-class EagerDeletionPass : public ir::Pass {
- protected:
-  std::unique_ptr<ir::Graph> ApplyImpl(
-      std::unique_ptr<ir::Graph> graph) const override;
-};
-
-}  // namespace details
-}  // namespace framework
-}  // namespace paddle

paddle/fluid/framework/details/while_op_eager_deletion_pass.cc

@@ -0,0 +1,62 @@
+// Copyright (c) 2019 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/details/computation_op_handle.h"
+#include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/ir/graph_helper.h"
+#include "paddle/fluid/operators/controlflow/while_op_helper.h"
+
+namespace paddle {
+namespace framework {
+namespace details {
+
+class WhileOpEagerDeletionPass : public ir::Pass {
+ protected:
+  std::unique_ptr<ir::Graph> ApplyImpl(
+      std::unique_ptr<ir::Graph> graph) const override {
+    auto all_ops = ir::FilterByNodeWrapper<OpHandleBase>(*graph);
+
+    // Find all while_op and while_grad_op
+    std::unordered_map<size_t, std::pair<std::vector<OperatorBase *>,
+                                         std::vector<OperatorBase *>>>
+        target_ops;
+    for (auto *op : all_ops) {
+      auto compute_op = dynamic_cast<ComputationOpHandle *>(op);
+      if (compute_op == nullptr) continue;
+
+      if (compute_op->Name() == "while") {
+        target_ops[compute_op->GetScopeIdx()].first.emplace_back(
+            compute_op->GetOp());
+      } else if (compute_op->Name() == "while_grad") {
+        target_ops[compute_op->GetScopeIdx()].second.emplace_back(
+            compute_op->GetOp());
+      }
+    }
+
+    for (auto &ops_pair : target_ops) {
+      auto &while_ops = ops_pair.second.first;
+      auto &while_grad_ops = ops_pair.second.second;
+      operators::PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(
+          while_ops, while_grad_ops);
+    }
+    return graph;
+  }
+};
+
+}  // namespace details
+}  // namespace framework
+}  // namespace paddle
+
+REGISTER_PASS(while_op_eager_deletion_pass,
+              paddle::framework::details::WhileOpEagerDeletionPass);

paddle/fluid/framework/executor.cc

@@ -23,6 +23,7 @@ limitations under the License. */
 #include "paddle/fluid/framework/threadpool.h"
 #include "paddle/fluid/framework/transfer_scope_cache.h"
 #include "paddle/fluid/framework/variable_helper.h"
+#include "paddle/fluid/operators/controlflow/while_op_helper.h"
 #include "paddle/fluid/operators/distributed/distributed.h"
 #include "paddle/fluid/platform/place.h"
 #include "paddle/fluid/platform/profiler.h"
@@ -409,8 +410,7 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 
   int64_t max_memory_size = GetEagerDeletionThreshold();
   std::unique_ptr<GarbageCollector> gc;
-  // skip while_op and while_grad_op temporarily
-  if (max_memory_size >= 0 && !keep_kids) {
+  if (max_memory_size >= 0) {
     ctx->ResetReferenceCount();
 #ifdef PADDLE_WITH_CUDA
     if (platform::is_gpu_place(place_)) {
@@ -428,6 +428,10 @@ void Executor::RunPreparedContext(ExecutorPrepareContext* ctx, Scope* scope,
 #ifdef PADDLE_WITH_CUDA
     }
 #endif
+    if (gc) {
+      operators::PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(ctx->block_id_,
+                                                                 ctx->ops_);
+    }
   }
 
   for (auto& op : ctx->ops_) {

paddle/fluid/operators/controlflow/CMakeLists.txt

@@ -1,4 +1,5 @@
 include(operators)
 register_operators(DEPS naive_executor)
+cc_library(while_op_helper SRCS while_op_helper.cc DEPS operator) 
 
 file(APPEND ${pybind_file} "USE_OP(less_than);\nUSE_OP(logical_and);\nUSE_NO_KERNEL_OP(read_from_array);\n")

paddle/fluid/operators/controlflow/while_op.cc

@@ -18,6 +18,7 @@
 #include "paddle/fluid/framework/op_registry.h"
 #include "paddle/fluid/framework/operator.h"
 #include "paddle/fluid/framework/var_type.h"
+#include "paddle/fluid/operators/controlflow/while_op_helper.h"
 #include "paddle/fluid/operators/detail/safe_ref.h"
 
 namespace paddle {
@@ -26,14 +27,6 @@ namespace operators {
 using StepScopeVar = std::vector<framework::Scope *>;
 using LoDTensor = framework::LoDTensor;
 
-static constexpr char kStepBlock[] = "sub_block";
-static constexpr char kCondition[] = "Condition";
-static constexpr char kStepScopes[] = "StepScopes";
-static constexpr char kX[] = "X";
-static constexpr char kXGRAD[] = "X@GRAD";
-static constexpr char kOutputs[] = "Out";
-static constexpr char kSkipEagerDeletionVars[] = "skip_eager_deletion_vars";
-
 namespace {  // NOLINT
 static std::string GetSkipEagerDeletionVarsDebugString(
     const std::vector<std::string> &vars) {

paddle/fluid/operators/controlflow/while_op_helper.h

@@ -0,0 +1,43 @@
+// Copyright (c) 2019 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
+
+#include <memory>
+#include <string>
+#include <vector>
+#include "paddle/fluid/framework/operator.h"
+#include "paddle/fluid/platform/variant.h"
+
+namespace paddle {
+namespace operators {
+
+static constexpr char kStepBlock[] = "sub_block";
+static constexpr char kCondition[] = "Condition";
+static constexpr char kStepScopes[] = "StepScopes";
+static constexpr char kX[] = "X";
+static constexpr char kXGRAD[] = "X@GRAD";
+static constexpr char kOutputs[] = "Out";
+static constexpr char kSkipEagerDeletionVars[] = "skip_eager_deletion_vars";
+
+void PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(
+    int block_id,
+    const std::vector<std::unique_ptr<framework::OperatorBase>> &all_ops);
+
+void PrepareSafeEagerDeletionOnWhileOpAndWhileGradOp(
+    const std::vector<framework::OperatorBase *> &while_ops,
+    const std::vector<framework::OperatorBase *> &while_grad_ops);
+
+}  // namespace operators
+}  // namespace paddle

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

@@ -0,0 +1,153 @@
+# 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.
+
+from __future__ import print_function
+
+import os
+os.environ['CPU_NUM'] = '2'
+os.environ['FLAGS_eager_delete_tensor_gb'] = '0.0'
+os.environ['FLAGS_fast_eager_deletion_mode'] = '1'
+
+import unittest
+import paddle.fluid as fluid
+import paddle.fluid.layers as layers
+from paddle.fluid.executor import Executor
+import paddle.fluid.core as core
+from paddle.fluid.backward import append_backward
+import paddle.fluid.compiler as compiler
+import numpy
+import multiprocessing
+
+
+class TestEagerDeletionWhileOpBase(unittest.TestCase):
+    def test_main(self):
+        places = [core.CPUPlace(), ]
+        if core.is_compiled_with_cuda():
+            places.append(core.CUDAPlace(0))
+
+        for p in places:
+            for with_data_parallel in [False, True]:
+                with fluid.program_guard(fluid.Program(), fluid.Program()):
+                    with fluid.scope_guard(fluid.Scope()):
+                        self.run_main(p, with_data_parallel)
+
+    def run_main(self, place, with_data_parallel):
+        self.place = place
+        self.with_data_parallel = with_data_parallel
+
+        if not core.is_compiled_with_cuda() and isinstance(self.place,
+                                                           core.CUDPlace):
+            return
+
+        if isinstance(self.place, core.CUDAPlace):
+            device_cnt = core.get_cuda_device_count(
+            ) if self.with_data_parallel else 1
+        else:
+            device_cnt = int(
+                os.environ['CPU_NUM'],
+                multiprocessing.cpu_count()) if self.with_data_parallel else 1
+
+        d0 = layers.data(
+            "d0", shape=[10], append_batch_size=False, dtype='float32')
+        d1 = layers.data(
+            "d1", shape=[10], append_batch_size=False, dtype='float32')
+        d2 = layers.data(
+            "d2", shape=[10], append_batch_size=False, dtype='float32')
+
+        i = layers.zeros(shape=[1], dtype='int64')
+        i.stop_gradient = True
+
+        init = layers.zeros(shape=[10], dtype='float32')
+        mem_array = layers.array_write(x=init, i=i)
+        data_array = layers.array_write(x=d0, i=i)
+
+        i = layers.increment(i)
+        layers.array_write(d1, i, array=data_array)
+
+        i = layers.increment(i)
+        layers.array_write(d2, i, array=data_array)
+
+        i = layers.zeros(shape=[1], dtype='int64')
+        i.stop_gradient = True
+
+        array_len = layers.fill_constant(shape=[1], dtype='int64', value=1)
+        array_len.stop_gradient = True
+        cond = layers.less_than(x=i, y=array_len)
+
+        j = layers.fill_constant(shape=[1], dtype='int64', value=1)
+        j.stop_gradient = True
+
+        array_len2 = layers.fill_constant(shape=[1], dtype='int64', value=3)
+        array_len2.stop_gradient = True
+        cond2 = layers.less_than(x=j, y=array_len2)
+
+        while_op = layers.While(cond=cond)
+        while_op2 = layers.While(cond=cond2)
+        with while_op.block():
+            d = layers.array_read(array=data_array, i=i)
+            prev = layers.array_read(array=mem_array, i=i)
+            d = layers.reshape(d, shape=[10])
+            prev = layers.reshape(prev, shape=[10])
+            result = layers.sums(input=[d, prev])
+
+            i = layers.increment(x=i, in_place=True)
+            layers.array_write(result, i=i, array=mem_array)
+            layers.less_than(x=i, y=array_len, cond=cond)
+            with while_op2.block():
+                d2 = layers.array_read(array=data_array, i=j)
+                prev2 = layers.array_read(array=mem_array, i=j)
+                d2 = layers.reshape(d2, shape=[10])
+                prev2 = layers.reshape(prev2, shape=[10])
+                result2 = layers.sums(input=[d2, prev2])
+
+                j = layers.increment(x=j, in_place=True)
+                layers.array_write(result2, i=j, array=mem_array)
+                layers.less_than(x=j, y=array_len2, cond=cond2)
+
+        sum_result = layers.array_read(array=mem_array, i=j)
+        sum_result.persistable = True
+        tmp = layers.unsqueeze(sum_result, axes=[0])
+        tmp = layers.expand(tmp, expand_times=[10, 1])
+        fc = layers.fc(tmp, size=256)
+        loss = layers.mean(sum_result)
+
+        optim = fluid.optimizer.Adam(learning_rate=1e-3)
+        optim.minimize(loss)
+
+        exe = Executor(self.place)
+        exe.run(fluid.default_startup_program())
+
+        prog = compiler.CompiledProgram(fluid.default_main_program())
+        if self.with_data_parallel:
+            prog = prog.with_data_parallel()
+
+        for _ in range(5):
+            d = []
+            for i in range(3):
+                tmp = numpy.random.random(size=[10]).astype('float32')
+                if not self.with_data_parallel:
+                    d.append(tmp)
+                else:
+                    d.append(numpy.array([tmp] * device_cnt))
+
+            outs = exe.run(program=prog,
+                           feed={'d0': d[0],
+                                 'd1': d[1],
+                                 'd2': d[2]},
+                           fetch_list=[sum_result])
+            self.assertAlmostEqual(numpy.sum(d), numpy.sum(outs[0]), delta=0.01)
+
+
+if __name__ == '__main__':
+    unittest.main()

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

@@ -0,0 +1,26 @@
+# 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 os
+import unittest
+os.environ['FLAGS_eager_delete_tensor_gb'] = "0.0"
+os.environ['FLAGS_memory_fraction_of_eager_deletion'] = "0.55"
+
+os.environ[
+    'RECORDIO_FILENAME'] = '/tmp/eager_deletion_transformer.wmt16.recordio'
+
+from test_parallel_executor_transformer import TestTransformer
+
+if __name__ == '__main__':
+    unittest.main()