diff --git a/paddle/fluid/operators/lod_reset_op.h b/paddle/fluid/operators/lod_reset_op.h
index bd19d8908e35e51872d324ea5aa6bb02110d5a92..d36aa0ce025a1c0f717913131fcc75040d16afac 100644
--- a/paddle/fluid/operators/lod_reset_op.h
+++ b/paddle/fluid/operators/lod_reset_op.h
@@ -46,8 +46,7 @@ class LoDResetKernel : public framework::OpKernel<T> {
         auto* lod = lod_t->data<int>();
         if (platform::is_gpu_place(ctx.GetPlace())) {
           framework::Tensor lod_cpu;
-          framework::TensorCopy(*lod_t, platform::CPUPlace(),
-                                ctx.device_context(), &lod_cpu);
+          framework::TensorCopySync(*lod_t, platform::CPUPlace(), &lod_cpu);
           lod = lod_cpu.data<int>();
         }
         level0 = std::vector<int>(lod, lod + lod_t->numel());
diff --git a/paddle/fluid/operators/math/concat_test.cc b/paddle/fluid/operators/math/concat_test.cc
index f0847aafae78f17eb28745bd224d45ec86497030..a46f2d51ca64501a622b5b48b424dffa16efc5b4 100644
--- a/paddle/fluid/operators/math/concat_test.cc
+++ b/paddle/fluid/operators/math/concat_test.cc
@@ -69,8 +69,8 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a);
+    paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b);
   }
 
   std::vector<paddle::framework::Tensor> input;
@@ -86,8 +86,8 @@ void testConcat() {
 
   int* out_ptr;
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
-                                  &out_cpu);
+    paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(),
+                                      &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();
@@ -142,8 +142,8 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a);
+    paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b);
   }
 
   input.clear();
@@ -157,8 +157,8 @@ void testConcat() {
   PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
 
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
-                                  &out_cpu);
+    paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(),
+                                      &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();
@@ -215,8 +215,8 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a);
+    paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b);
   }
 
   input.clear();
@@ -230,8 +230,8 @@ void testConcat() {
   PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
 
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
-                                  &out_cpu);
+    paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(),
+                                      &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();
@@ -290,8 +290,8 @@ void testConcat() {
   }
 
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(input_a_cpu, Place(), *context, &input_a);
-    paddle::framework::TensorCopy(input_b_cpu, Place(), *context, &input_b);
+    paddle::framework::TensorCopySync(input_a_cpu, Place(), &input_a);
+    paddle::framework::TensorCopySync(input_b_cpu, Place(), &input_b);
   }
 
   input.clear();
@@ -305,8 +305,8 @@ void testConcat() {
   PADDLE_ENFORCE_EQ(input_b.dims(), dim_b);
 
   if (paddle::platform::is_gpu_place(Place())) {
-    paddle::framework::TensorCopy(out, paddle::platform::CPUPlace(), *context,
-                                  &out_cpu);
+    paddle::framework::TensorCopySync(out, paddle::platform::CPUPlace(),
+                                      &out_cpu);
     out_ptr = out_cpu.data<int>();
   } else {
     out_ptr = out.data<int>();
diff --git a/paddle/fluid/operators/math/sequence_padding_test.cc b/paddle/fluid/operators/math/sequence_padding_test.cc
index b9a1b9ae4d6e8c82aa782190d9e145e7a2b502be..b0c201db0ccbe81d8f57cd984d2cdfd2f6a48f25 100644
--- a/paddle/fluid/operators/math/sequence_padding_test.cc
+++ b/paddle/fluid/operators/math/sequence_padding_test.cc
@@ -41,7 +41,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
   if (paddle::platform::is_cpu_place(*place)) {
     seq = cpu_seq;
   } else {
-    TensorCopy(cpu_seq, *place, *context, &seq);
+    TensorCopySync(cpu_seq, *place, &seq);
     seq.set_lod(lod);
   }
 
@@ -64,7 +64,7 @@ void TestSequencePadding(const paddle::framework::LoD& lod,
   if (paddle::platform::is_cpu_place(*place)) {
     cpu_seq_back = seq_back;
   } else {
-    TensorCopy(seq_back, paddle::platform::CPUPlace(), *context, &cpu_seq_back);
+    TensorCopySync(seq_back, paddle::platform::CPUPlace(), &cpu_seq_back);
     cpu_seq_back.set_lod(lod);
   }
 
diff --git a/paddle/fluid/operators/multiplex_op.cu b/paddle/fluid/operators/multiplex_op.cu
index 45a2550793511f7cb8c20644ac79e9e88629ce7b..2f8a602f3c5c0a7c262235f99943ce336e20a7b4 100644
--- a/paddle/fluid/operators/multiplex_op.cu
+++ b/paddle/fluid/operators/multiplex_op.cu
@@ -33,7 +33,7 @@ class MultiplexGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    TensorCopy(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
+    TensorCopySync(*ids, platform::CPUPlace(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
     auto stream = ctx.cuda_device_context().stream();
     platform::CUDAPlace place = boost::get<platform::CUDAPlace>(ctx.GetPlace());
@@ -69,7 +69,7 @@ class MultiplexGradGPUKernel : public framework::OpKernel<T> {
     auto cols = ins[0]->numel() / rows;
     // copy index to cpu
     Tensor index_t_cpu;
-    TensorCopy(*ids, platform::CPUPlace(), ctx.device_context(), &index_t_cpu);
+    TensorCopySync(*ids, platform::CPUPlace(), &index_t_cpu);
     auto* index = index_t_cpu.data<int32_t>();
 
     auto stream = ctx.cuda_device_context().stream();
diff --git a/paddle/fluid/operators/sequence_slice_op.h b/paddle/fluid/operators/sequence_slice_op.h
index b9c565cac9581a2e830697c1136919062eef345c..b5ea6ff49bbb29571f9a6ef6358ef881acd9be9e 100644
--- a/paddle/fluid/operators/sequence_slice_op.h
+++ b/paddle/fluid/operators/sequence_slice_op.h
@@ -66,13 +66,11 @@ class SequenceSliceOpKernel : public framework::OpKernel<T> {
 
     if (platform::is_gpu_place(ctx.GetPlace())) {
       offset_cpu.mutable_data<T>(offset->dims(), platform::CPUPlace());
-      framework::TensorCopy(*offset, platform::CPUPlace(), ctx.device_context(),
-                            &offset_cpu);
+      framework::TensorCopySync(*offset, platform::CPUPlace(), &offset_cpu);
       offset_data = offset_cpu.data<int64_t>();
 
       length_cpu.mutable_data<T>(length->dims(), platform::CPUPlace());
-      framework::TensorCopy(*length, platform::CPUPlace(), ctx.device_context(),
-                            &length_cpu);
+      framework::TensorCopySync(*length, platform::CPUPlace(), &length_cpu);
       length_data = length_cpu.data<int64_t>();
     }
 
@@ -127,13 +125,11 @@ class SequenceSliceGradOpKernel : public framework::OpKernel<T> {
 
     if (platform::is_gpu_place(ctx.GetPlace())) {
       offset_cpu.mutable_data<T>(offset->dims(), platform::CPUPlace());
-      framework::TensorCopy(*offset, platform::CPUPlace(), ctx.device_context(),
-                            &offset_cpu);
+      framework::TensorCopySync(*offset, platform::CPUPlace(), &offset_cpu);
       offset_data = offset_cpu.data<int64_t>();
 
       length_cpu.mutable_data<T>(length->dims(), platform::CPUPlace());
-      framework::TensorCopy(*length, platform::CPUPlace(), ctx.device_context(),
-                            &length_cpu);
+      framework::TensorCopySync(*length, platform::CPUPlace(), &length_cpu);
       length_data = length_cpu.data<int64_t>();
     }