diff --git a/paddle/fluid/framework/attribute.h b/paddle/fluid/framework/attribute.h
index 66b988ee1f1fb6486423b0d4196c883979ee6fe3..079ef7e2a2468cfa819adc98d5a09c90224c17eb 100644
--- a/paddle/fluid/framework/attribute.h
+++ b/paddle/fluid/framework/attribute.h
@@ -208,7 +208,8 @@ Attribute GetAttrValue(const proto::OpDesc::Attr& attr_desc);
 
 class AttrReader {
  public:
-  explicit AttrReader(const AttributeMap& attrs) : attrs_(attrs) {}
+  explicit AttrReader(const AttributeMap& attrs, const AttributeMap& default_attrs = {} ) 
+      : attrs_(attrs), default_attrs_(default_attrs) {}
 
   template <typename T>
   inline const T& Get(const std::string& name) const {
@@ -224,6 +225,7 @@ class AttrReader {
 
  private:
   const AttributeMap& attrs_;
+  const AttributeMap& default_attrs_;
 };
 
 // check whether a value(attribute) fit a certain limit
@@ -406,6 +408,14 @@ class OpAttrChecker {
     return default_values_map;
   }
 
+  void InitDefaultMap() {
+    for (const auto& checker : attr_checkers_) {
+      checker(&default_values_map_, true);
+    }
+  }
+
+  const AttributeMap& default_attr_map() const { return default_values_map_; }
+
   void RecordExplicitCheckerNum() {
     explicit_checker_num_ = attr_checkers_.size();
   }
@@ -413,6 +423,8 @@ class OpAttrChecker {
  private:
   std::vector<AttrChecker> attr_checkers_;
 
+  AttributeMap default_values_map_;
+
   // in order to improve the efficiency of dynamic graph mode,
   // we divede the attribute into explicit type and implicit type.
   // for explicit attribute, we mean the attribute added in the customized
diff --git a/paddle/fluid/framework/hogwild_worker.cc b/paddle/fluid/framework/hogwild_worker.cc
index d8639643f2c8a773246d9cd88637a7f9f478e8a8..5de8d0fcdde91993d69c16cea805a7e4042c9645 100644
--- a/paddle/fluid/framework/hogwild_worker.cc
+++ b/paddle/fluid/framework/hogwild_worker.cc
@@ -194,11 +194,14 @@ void HogwildWorker::TrainFilesWithProfiler() {
 
 void HogwildWorker::TrainFiles() {
   platform::SetNumThreads(1);
-
+    
+  std::cerr << "1!!!!!" << std::endl;
   // how to accumulate fetched values here
   device_reader_->Start();
   int cur_batch;
+  int i = 0;
   while ((cur_batch = device_reader_->Next()) > 0) {
+    i++;
     for (auto &op : ops_) {
       bool need_skip = false;
       for (auto t = 0u; t < skip_ops_.size(); ++t) {
@@ -215,6 +218,7 @@ void HogwildWorker::TrainFiles() {
     PrintFetchVars();
     thread_scope_->DropKids();
   }
+  std::cerr << "total bacth " << i << std::endl;
 #if defined PADDLE_WITH_PSCORE
   if (thread_barrier_) {
     paddle::distributed::Communicator::GetInstance()->BarrierTriggerDecrement();
diff --git a/paddle/fluid/framework/multi_trainer.cc b/paddle/fluid/framework/multi_trainer.cc
index ff8e71b92e0ac5ea53aa883587d6c55bdde395ec..a9e15ee1758070794e7d933d2e9c093186ac60a5 100644
--- a/paddle/fluid/framework/multi_trainer.cc
+++ b/paddle/fluid/framework/multi_trainer.cc
@@ -124,6 +124,7 @@ Scope* MultiTrainer::GetWorkerScope(int thread_id) {
 
 void MultiTrainer::Run() {
   VLOG(3) << "Going to run";
+  LOG(ERROR) << "multi run " << thread_num_ << "\t" << debug_;
   for (int thidx = 0; thidx < thread_num_; ++thidx) {
     if (!debug_) {
       threads_.push_back(
diff --git a/paddle/fluid/framework/op_proto_maker.cc b/paddle/fluid/framework/op_proto_maker.cc
index 0b9fd0a47e22c76fa9612d4e0ff3632448197a98..d86b555d9a843d9aae2e9b99216d61f0960d8268 100644
--- a/paddle/fluid/framework/op_proto_maker.cc
+++ b/paddle/fluid/framework/op_proto_maker.cc
@@ -66,6 +66,7 @@ void OpProtoAndCheckerMaker::operator()(proto::OpProto* proto,
   op_checker_ = attr_checker;
   Make();
   op_checker_->RecordExplicitCheckerNum();
+  op_checker_->InitDefaultMap();
 
   AddAttr<int>(OpRoleAttrName(), "The role of this operator")
       .InEnum(
diff --git a/paddle/fluid/imperative/CMakeLists.txt b/paddle/fluid/imperative/CMakeLists.txt
index a24c0ac09c75876a406cd9f6bd40ba5296fa6a89..5cfa2c2278662d6164464450e0993d3fe676382c 100644
--- a/paddle/fluid/imperative/CMakeLists.txt
+++ b/paddle/fluid/imperative/CMakeLists.txt
@@ -28,4 +28,6 @@ endif(NOT WIN32)
 
 cc_library(gradient_accumulator SRCS gradient_accumulator.cc DEPS blas operator lod_tensor selected_rows selected_rows_functor var_type_traits layer math_function)
 
+cc_binary(tracer_test SRCS tracer_test.cc DEPS tracer layer  op_registry python pybind ${GLOB_OP_LIB} ${GLOB_OPERATOR_DEPS}  profiler )
+
 add_subdirectory(tests)
diff --git a/paddle/fluid/imperative/execution_context.h b/paddle/fluid/imperative/execution_context.h
index 398b1292e2ffe05beef3fea50c7b676625cab5bd..885c5d5ed0de80b1ec3320d2b2f76faf518bcffe 100644
--- a/paddle/fluid/imperative/execution_context.h
+++ b/paddle/fluid/imperative/execution_context.h
@@ -35,11 +35,13 @@ class DygraphExecutionContext : public framework::ExecutionContext {
                           const framework::RuntimeContext& ctx,
                           const NameVarMap<VarType>& var_base_map_in,
                           const NameVarMap<VarType>& var_base_map_out,
-                          const framework::AttributeMap& attrs)
+                          const framework::AttributeMap& attrs,
+                          const framework::AttributeMap& default_attrs)
       : ExecutionContext(op, scope, device_context, ctx),
         var_base_map_in_(var_base_map_in),
         var_base_map_out_(var_base_map_out),
-        attrs_(attrs) {}
+        attrs_(attrs),
+        default_attrs_(default_attrs){}
 
   std::string InputName(const std::string& name) const override {
     auto it = var_base_map_in_.find(name);
@@ -92,16 +94,22 @@ class DygraphExecutionContext : public framework::ExecutionContext {
   }
 
   bool HasAttr(const std::string& name) const override {
-    return attrs_.count(name) != 0;
+    return attrs_.count(name) != 0 || default_attrs_.count( name );
   }
 
   const framework::AttributeMap& Attrs() const override { return attrs_; }
 
   const framework::Attribute& GetAttr(const std::string& name) const override {
     auto it = attrs_.find(name);
-
+    
+    bool find = ( it != attrs_.end() );
+    if( it == attrs_.end() )
+    {
+        it = default_attrs_.find( name );
+        find = (  it != default_attrs_.end() );
+    }
     PADDLE_ENFORCE_NE(
-        it, attrs_.end(),
+        find, false,
         platform::errors::NotFound("can not find [%s] in attrs", name));
 
     return it->second;
@@ -192,6 +200,7 @@ class DygraphExecutionContext : public framework::ExecutionContext {
   const NameVarMap<VarType>& var_base_map_in_;
   const NameVarMap<VarType>& var_base_map_out_;
   const framework::AttributeMap& attrs_;
+  const framework::AttributeMap& default_attrs_;
 };
 
 }  // namespace imperative
diff --git a/paddle/fluid/imperative/prepared_operator.cc b/paddle/fluid/imperative/prepared_operator.cc
index 2a3b6424d4a14e1cd6345cf24594582bd19f51d4..5c05a5642a4381c1fae8f05b22fcbb858a3db132 100644
--- a/paddle/fluid/imperative/prepared_operator.cc
+++ b/paddle/fluid/imperative/prepared_operator.cc
@@ -104,10 +104,11 @@ PreparedOp PrepareImpl(const NameVarMap<VarType>& ins,
   }
 #endif
 
+  //auto *attr_checker = op_->Info().Checker();
   // 1. get expected kernel key
   auto expected_kernel_key =
       op.GetExpectedKernelType(DygraphExecutionContext<VarType>(
-          op, framework::Scope(), *dev_ctx, ctx, ins, outs, attrs));
+          op, framework::Scope(), *dev_ctx, ctx, ins, outs, attrs, {} )) ;
   VLOG(3) << "expected_kernel_key:" << expected_kernel_key;
 
   // 2. check if op[type] has kernel registered.
@@ -172,8 +173,10 @@ static void PreparedOpRunImpl(
   static_cast<const framework::OperatorWithKernel&>(op).InferShape(
       &infer_shape_ctx);
 
+  auto *attr_checker = op.Info().Checker();
   func(DygraphExecutionContext<VarType>(op, scope, *dev_ctx, ctx, ins, outs,
-                                        attrs));
+                                        attrs, 
+                                        attr_checker->default_attr_map() ));
 
   /**
    * [ Why need handle complex gradient to real gradient? ]
diff --git a/paddle/fluid/imperative/tests/test_layer.cc b/paddle/fluid/imperative/tests/test_layer.cc
index 4a30ffb7e3d01ffa90a42278e2e5ef5271045d8a..93602ddaa343f8e4c6a46bbc7524b254a859430e 100644
--- a/paddle/fluid/imperative/tests/test_layer.cc
+++ b/paddle/fluid/imperative/tests/test_layer.cc
@@ -358,7 +358,7 @@ TEST(test_layer, test_dygraph_execution_context) {
   framework::Scope scope;
 
   DygraphExecutionContext<imperative::VarBase> dy_exe_context(
-      *(op.get()), scope, *dev_ctx, ctx, ins, outs, concat_att_map);
+      *(op.get()), scope, *dev_ctx, ctx, ins, outs, concat_att_map, {});
 
   ASSERT_EQ(dy_exe_context.InputSize("X"), 1u);
   ASSERT_EQ(dy_exe_context.InputName("X"), "vin");
diff --git a/paddle/fluid/imperative/tracer.cc b/paddle/fluid/imperative/tracer.cc
index 608cc407d5b776cf6e88a89b18de8c2389c965d1..c9567a98a73bf213c70bbfbd2c9a39b5efb38d2e 100644
--- a/paddle/fluid/imperative/tracer.cc
+++ b/paddle/fluid/imperative/tracer.cc
@@ -149,11 +149,16 @@ void Tracer::TraceOp(const std::string& type, const NameVarBaseMap& ins,
     }
   }
   auto op = framework::OpRegistry::CreateOp(type, {}, {}, {}, false);
+  
+  
   const auto& op_info = op->Info();
   auto* attr_checker = op_info.Checker();
+  
   if (attr_checker) {
     attr_checker->Check(&attrs, true);
   }
+  
+  
 
   NameVarBaseMap new_ins = ins;
   if (enable_autocast_) {
diff --git a/paddle/fluid/operators/elementwise/elementwise_op.h b/paddle/fluid/operators/elementwise/elementwise_op.h
index 6ec73b02ade11a03f230e85a3cb54efb04f3305b..3cec5388d11f9764e3ab6d64d944ea2a9a0ed439 100644
--- a/paddle/fluid/operators/elementwise/elementwise_op.h
+++ b/paddle/fluid/operators/elementwise/elementwise_op.h
@@ -109,6 +109,7 @@ class ElementwiseOp : public framework::OperatorWithKernel {
     auto input_data_type =
         OperatorWithKernel::IndicateOrPromoteVarDataTypes(ctx, "X", "Y");
 
+/*
 #ifdef PADDLE_WITH_MKLDNN
     if (this->CanMKLDNNBeUsed(ctx, input_data_type)) {
       return framework::OpKernelType(input_data_type, ctx.GetPlace(),
@@ -116,6 +117,7 @@ class ElementwiseOp : public framework::OperatorWithKernel {
                                      framework::LibraryType::kMKLDNN);
     }
 #endif
+*/
     return framework::OpKernelType(input_data_type, ctx.GetPlace());
   }
 
diff --git a/paddle/fluid/operators/sync_batch_norm_op.cu.h b/paddle/fluid/operators/sync_batch_norm_op.cu.h
index d08a34ade77f289810a181d611ef8d2801c1d56d..e785e89aebaa5cce7c5dab0abb44ef2d7a4393d4 100644
--- a/paddle/fluid/operators/sync_batch_norm_op.cu.h
+++ b/paddle/fluid/operators/sync_batch_norm_op.cu.h
@@ -186,12 +186,14 @@ void SyncBatchNormFunctor(const framework::ExecutionContext &ctx,
                    framework::DataLayout::kNHWC><<<grid, threads, 0, stream>>>(
           x_d, N, H * W * D, C, stats);
     }
-
+    
+    /*
     Tensor c_g_st;
     auto *c_g_st_d = c_g_st.mutable_data<BatchNormParamType<T>>(
         {2 * C + 1}, platform::CPUPlace());
     auto gplace = BOOST_GET_CONST(platform::CUDAPlace, ctx.GetPlace());
     memory::Copy(platform::CPUPlace(), c_g_st_d, gplace, stats, bytes, 0);
+    */
 
 #if defined(PADDLE_WITH_NCCL) || defined(PADDLE_WITH_RCCL)
     auto *comm = dev_ctx.nccl_comm();
diff --git a/paddle/fluid/pybind/op_function.h b/paddle/fluid/pybind/op_function.h
index 0c457531211b906d1d6e43363df4511aea5c7435..4e5ddac3ec9f82ac6a1860bc79b88df6016502b2 100644
--- a/paddle/fluid/pybind/op_function.h
+++ b/paddle/fluid/pybind/op_function.h
@@ -177,4 +177,4 @@ static inline void HandleViewBetweenInputAndOutput(
 }  // namespace paddle
 
 // This include must be the last line
-#include "paddle/fluid/pybind/op_function_impl.h"
+#include "paddle/fluid/pybind/op_function_impl_new.h"