fix env variable settting bug

test=develop

paddle/fluid/framework/CMakeLists.txt

@@ -63,7 +63,7 @@ cc_library(lod_tensor SRCS lod_tensor.cc DEPS ddim place tensor framework_proto
 cc_test(lod_tensor_test SRCS lod_tensor_test.cc DEPS lod_tensor memory)
 nv_test(lod_tensor_gpu_test SRCS lod_tensor_test.cu DEPS lod_tensor)
 
-cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory gflags)
+cc_library(garbage_collector SRCS garbage_collector.cc DEPS device_context memory gflags glog)
 
 cc_library(reader SRCS reader.cc DEPS lod_tensor ddim)
 cc_test(reader_test SRCS reader_test.cc DEPS reader)

paddle/fluid/framework/details/eager_deletion_pass.cc

@@ -22,14 +22,9 @@
 #include "paddle/fluid/framework/details/computation_op_handle.h"
 #include "paddle/fluid/framework/details/eager_deletion_op_handle.h"
 #include "paddle/fluid/framework/details/multi_devices_helper.h"
+#include "paddle/fluid/framework/garbage_collector.h"
 #include "paddle/fluid/framework/ir/graph_helper.h"
 
-DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
-              "Fraction of eager deletion. If less than 1.0, all variables in "
-              "the program would be sorted according to its memory size, and "
-              "only the FLAGS_memory_fraction_of_eager_deletion of the largest "
-              "variables would be deleted.");
-
 namespace paddle {
 namespace framework {
 namespace details {
@@ -206,8 +201,9 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     }
   }
 
-  op_vars_map = ShrinkGCVars(op_vars_map, vars, places,
-                             FLAGS_memory_fraction_of_eager_deletion);
+  double memory_fraction = framework::GetEagerDeletionMemoryFraction();
+
+  op_vars_map = ShrinkGCVars(op_vars_map, vars, places, memory_fraction);
 
   for (auto &pair : op_vars_map) {
     auto *op = pair.first;
@@ -239,8 +235,7 @@ std::unique_ptr<ir::Graph> EagerDeletionPass::ApplyImpl(
     eager_deletion_op->AddOutput(dummy_leaf);
   }
 
-  VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = "
-           << FLAGS_memory_fraction_of_eager_deletion;
+  VLOG(10) << "FLAGS_memory_fraction_of_eager_deletion = " << memory_fraction;
   VLOG(10) << "Create " << op_vars_map.size() << " EagerDeletionOpHandle(s)";
 
   auto while_op_eager_deletion_pass =

paddle/fluid/framework/executor_gc_helper.cc

@@ -93,7 +93,6 @@ std::unordered_map<OperatorBase *, std::vector<std::string>> GetUnusedVars(
     const BlockDesc &block,
     const std::vector<std::unique_ptr<OperatorBase>> &ops,
     const std::vector<std::string> &skip_var_list) {
-  UseGarbageCollectorGFlags();
   std::unordered_set<std::string> skip_vars(skip_var_list.begin(),
                                             skip_var_list.end());
 

paddle/fluid/framework/garbage_collector.cc

@@ -22,6 +22,7 @@
 #include "paddle/fluid/platform/cuda_device_guard.h"
 #endif
 #include "gflags/gflags.h"
+#include "glog/logging.h"
 #include "paddle/fluid/framework/garbage_collector.h"
 
 namespace paddle {
@@ -36,6 +37,12 @@ DEFINE_bool(fast_eager_deletion_mode, true,
             "Fast eager deletion mode. If enabled, memory would release "
             "immediately without waiting GPU kernel ends.");
 
+DEFINE_double(memory_fraction_of_eager_deletion, 1.0,
+              "Fraction of eager deletion. If less than 1.0, all variables in "
+              "the program would be sorted according to its memory size, and "
+              "only the FLAGS_memory_fraction_of_eager_deletion of the largest "
+              "variables would be deleted.");
+
 GarbageCollector::GarbageCollector(const platform::Place &place,
                                    size_t max_memory_size)
     : max_memory_size_((std::max)(max_memory_size, static_cast<size_t>(1))) {
@@ -101,8 +108,6 @@ void StreamGarbageCollector::ClearCallback(
 }
 #endif
 
-void UseGarbageCollectorGFlags() {}
-
 int64_t GetEagerDeletionThreshold() {
   return FLAGS_eager_delete_tensor_gb < 0
              ? -1
@@ -111,5 +116,16 @@ int64_t GetEagerDeletionThreshold() {
 }
 
 bool IsFastEagerDeletionModeEnabled() { return FLAGS_fast_eager_deletion_mode; }
+
+void SetEagerDeletionMode(double threshold, double fraction, bool fast_mode) {
+  FLAGS_eager_delete_tensor_gb = threshold;
+  FLAGS_memory_fraction_of_eager_deletion = fraction;
+  FLAGS_fast_eager_deletion_mode = fast_mode;
+}
+
+double GetEagerDeletionMemoryFraction() {
+  return FLAGS_memory_fraction_of_eager_deletion;
+}
+
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/framework/garbage_collector.h

@@ -19,6 +19,7 @@
 #include <memory>
 #include <mutex>  // NOLINT
 #include <utility>
+#include "gflags/gflags.h"
 #include "paddle/fluid/platform/device_context.h"
 
 namespace paddle {
@@ -130,7 +131,9 @@ void GarbageCollector::Add(Container &&objs, Callback &&callback) {
 int64_t GetEagerDeletionThreshold();
 bool IsFastEagerDeletionModeEnabled();
 
-extern void UseGarbageCollectorGFlags();
+void SetEagerDeletionMode(double threshold, double fraction, bool fast_mode);
+
+double GetEagerDeletionMemoryFraction();
 
 }  // namespace framework
 }  // namespace paddle

paddle/fluid/operators/batch_norm_op.cc

@@ -567,6 +567,7 @@ std::unique_ptr<framework::OpDesc> BatchNormGradMaker::Apply() const {
   op->SetInput(framework::GradVarName("Y"), OutputGrad("Y"));
 
   op->SetInput("Scale", Input("Scale"));
+  op->SetInput("Bias", Input("Bias"));
   op->SetInput("SavedMean", Output("SavedMean"));
   op->SetInput("SavedVariance", Output("SavedVariance"));
 

paddle/fluid/operators/concat_op.cc

@@ -152,13 +152,28 @@ class ConcatOpGrad : public framework::OperatorWithKernel {
 DECLARE_NO_NEED_BUFFER_VARS_INFERENCE(ConcatOpGradNoNeedBufferVarInference,
                                       "X");
 
+class ConcatGradOpDescMaker : public framework::SingleGradOpDescMaker {
+ public:
+  using framework::SingleGradOpDescMaker::SingleGradOpDescMaker;
+
+ protected:
+  std::unique_ptr<framework::OpDesc> Apply() const override {
+    std::unique_ptr<framework::OpDesc> op(new framework::OpDesc());
+    op->SetType("concat_grad");
+    op->SetInput("X", Input("X"));
+    op->SetInput(framework::GradVarName("Out"), OutputGrad("Out"));
+    op->SetOutput(framework::GradVarName("X"), InputGrad("X", false));
+    op->SetAttrMap(Attrs());
+    return op;
+  }
+};
+
 }  // namespace operators
 }  // namespace paddle
 
 namespace ops = paddle::operators;
 REGISTER_OPERATOR(concat, ops::ConcatOp, ops::ConcatOpMaker,
-                  paddle::framework::DefaultGradOpDescMaker<
-                      false> /* set false to disable empty grad */);
+                  ops::ConcatGradOpDescMaker);
 REGISTER_OPERATOR(concat_grad, ops::ConcatOpGrad,
                   ops::ConcatOpGradNoNeedBufferVarInference);
 REGISTER_OP_CPU_KERNEL(

paddle/fluid/pybind/pybind.cc

@@ -76,10 +76,6 @@ DEFINE_bool(reader_queue_speed_test_mode, false,
             "If set true, the queue.pop will only get data from queue but not "
             "remove the data from queue for speed testing");
 
-DECLARE_double(eager_delete_tensor_gb);
-DECLARE_bool(fast_eager_deletion_mode);
-DECLARE_double(memory_fraction_of_eager_deletion);
-
 // disable auto conversion to list in Python
 PYBIND11_MAKE_OPAQUE(paddle::framework::LoDTensorArray);
 
@@ -144,7 +140,6 @@ PYBIND11_MODULE(core, m) {
   paddle::platform::CpuTotalPhysicalMemory();
 
   paddle::memory::allocation::UseAllocatorStrategyGFlag();
-  paddle::framework::UseGarbageCollectorGFlags();
 
   m.doc() = "C++ core of PaddlePaddle";
 
@@ -163,12 +158,7 @@ PYBIND11_MODULE(core, m) {
   // NOTE(zjl): ctest would load environment variables at the beginning even
   // though we have not `import paddle.fluid as fluid`. So we add this API
   // to enable eager deletion mode in unittest.
-  m.def("_set_eager_deletion_mode",
-        [](double threshold, double fraction, bool fast_mode) {
-          FLAGS_eager_delete_tensor_gb = threshold;
-          FLAGS_memory_fraction_of_eager_deletion = fraction;
-          FLAGS_fast_eager_deletion_mode = fast_mode;
-        });
+  m.def("_set_eager_deletion_mode", &paddle::framework::SetEagerDeletionMode);
 
   m.add_object("_cleanup",
                py::capsule([]() { ScopePool::Instance().Clear(); }));

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

@@ -14,6 +14,7 @@
 
 import os
 import unittest
+import paddle.fluid as fluid
 
 fluid.core._set_eager_deletion_mode(0.0, 1.0, True)
 

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

@@ -14,6 +14,7 @@
 
 import os
 import unittest
+import paddle.fluid as fluid
 
 fluid.core._set_eager_deletion_mode(0.0, 1.0, True)