diff --git a/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.cc b/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.cc
index 640607b9e54..7ef1cb31571 100644
--- a/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.cc
+++ b/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.cc
@@ -21,6 +21,7 @@
 #include <utility>
 #include <functional>
 #include <unordered_map>
+#include <set>
 #include "kernel/kernel.h"
 #include "device/cpu/cpu_device_address.h"
 #include "utils/context/ms_context.h"
@@ -139,8 +140,12 @@ DeviceAddressPtr CPUKernelRuntime::CreateDeviceAddress(void *device_ptr, size_t
   return std::make_shared<CPUDeviceAddress>(device_ptr, device_size, format, type_id);
 }
 
-BaseRef CPUKernelRuntime::CreatTensorForOutput(const AnfNodePtr &input_node, size_t index,
-                                               const std::unordered_map<AnfNode *, tensor::TensorPtr> &input_map) {
+BaseRef CPUKernelRuntime::CreatTensorForOutput(const session::KernelWithIndex &kernel_with_index,
+                                               const std::unordered_map<AnfNode *, tensor::TensorPtr> &input_map,
+                                               std::set<DeviceAddressPtr> *bound_addresses,
+                                               std::vector<tensor::TensorPtr> *need_sync_outputs) {
+  auto &input_node = kernel_with_index.first;
+  auto index = kernel_with_index.second;
   MS_EXCEPTION_IF_NULL(input_node);
   if (input_node->isa<CNode>() && AnfAlgo::GetCNodeName(input_node) == prim::kPrimMakeTuple->name()) {
     auto cnode = input_node->cast<CNodePtr>();
@@ -148,7 +153,7 @@ BaseRef CPUKernelRuntime::CreatTensorForOutput(const AnfNodePtr &input_node, siz
     VectorRef ret;
     for (size_t i = 1; i < cnode->inputs().size(); i++) {
       auto item_with_index = AnfAlgo::VisitKernelWithReturnType(cnode->input(i), 0);
-      auto out = CreatTensorForOutput(item_with_index.first, item_with_index.second, input_map);
+      auto out = CreatTensorForOutput(item_with_index, input_map, bound_addresses, need_sync_outputs);
       ret.push_back(out);
     }
     return ret;
@@ -169,11 +174,13 @@ BaseRef CPUKernelRuntime::CreatTensorForOutput(const AnfNodePtr &input_node, siz
     type_id = GetCPUSupportOutputTypeId(type_id);
     tensor::TensorPtr tensor = std::make_shared<tensor::Tensor>(type_id, temp_shape);
     MS_EXCEPTION_IF_NULL(tensor);
-    if (address->ref_count_ > 0 && address->ptr_ != nullptr) {
+    if (bound_addresses->find(address) != bound_addresses->end()) {
       tensor->set_device_address(address);
+      need_sync_outputs->emplace_back(tensor);
     } else {
       address->ptr_ = tensor->data_c(true);
       address->ref_count_ = INIT_NODE_REF;
+      (void)bound_addresses->insert(address);
     }
     tensor->set_dirty(false);
     return tensor;
@@ -187,7 +194,8 @@ BaseRef CPUKernelRuntime::CreatTensorForOutput(const AnfNodePtr &input_node, siz
 }
 
 void CPUKernelRuntime::BindInputOutput(const session::KernelGraph *kernel_graph,
-                                       const std::vector<tensor::TensorPtr> &inputs, VectorRef *outputs) {
+                                       const std::vector<tensor::TensorPtr> &inputs, VectorRef *outputs,
+                                       std::vector<tensor::TensorPtr> *need_sync_outputs) {
   MS_EXCEPTION_IF_NULL(kernel_graph);
   MS_EXCEPTION_IF_NULL(outputs);
   // bind input ptr
@@ -195,10 +203,8 @@ void CPUKernelRuntime::BindInputOutput(const session::KernelGraph *kernel_graph,
   if (input_nodes.size() != inputs.size()) {
     MS_LOG(EXCEPTION) << "Input size not equal to input node size!";
   }
-
   std::unordered_map<AnfNode *, tensor::TensorPtr> input_map;
   size_t input_idx = 0;
-  size_t type_size = sizeof(float);
   for (auto &item : input_nodes) {
     MS_EXCEPTION_IF_NULL(item);
     input_map[item.get()] = inputs[input_idx];
@@ -212,7 +218,8 @@ void CPUKernelRuntime::BindInputOutput(const session::KernelGraph *kernel_graph,
         (void)tensor->data_sync();
       }
       std::vector<int> data_shape = tensor->shape();
-      size_t tensor_size = std::accumulate(data_shape.begin(), data_shape.end(), type_size, std::multiplies<size_t>());
+      size_t tensor_size =
+        std::accumulate(data_shape.begin(), data_shape.end(), sizeof(float), std::multiplies<size_t>());
       if (tensor->data_type() == kNumberTypeFloat32 || tensor->data_type() == kNumberTypeInt32) {
         address->ptr_ = tensor->data_c(false);
       } else {
@@ -223,18 +230,17 @@ void CPUKernelRuntime::BindInputOutput(const session::KernelGraph *kernel_graph,
         }
         tensor->set_dirty(true);
       }
-
       address->ref_count_ = INIT_NODE_REF;
       tensor->set_device_address(address);
     }
     input_idx++;
   }
-
   // new output and bind ptr
+  std::set<DeviceAddressPtr> bound_addresses;
   auto output_nodes = kernel_graph->outputs();
   for (const auto &item : output_nodes) {
     auto item_with_index = AnfAlgo::VisitKernelWithReturnType(item, 0, true);
-    auto out = CreatTensorForOutput(item_with_index.first, item_with_index.second, input_map);
+    auto out = CreatTensorForOutput(item_with_index, input_map, &bound_addresses, need_sync_outputs);
     outputs->push_back(std::move(out));
   }
 }
diff --git a/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.h b/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.h
index ac63f55d3ee..27dcefdba91 100644
--- a/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.h
+++ b/mindspore/ccsrc/device/cpu/cpu_kernel_runtime.h
@@ -20,10 +20,12 @@
 #include <vector>
 #include <string>
 #include <unordered_map>
+#include <set>
 #include "device/kernel_runtime.h"
 #include "session/kernel_graph.h"
 #include "session/session_basic.h"
 #include "device/cpu/cpu_resource_manager.h"
+#include "session/anf_runtime_algorithm.h"
 #include "utils/any.h"
 namespace mindspore {
 namespace device {
@@ -37,7 +39,7 @@ class CPUKernelRuntime : public KernelRuntime {
   bool Run(session::KernelGraph *graph) override;
   void AssignKernelAddress(session::KernelGraph *kernel_graph);
   void BindInputOutput(const session::KernelGraph *kernel_graph, const std::vector<tensor::TensorPtr> &inputs,
-                       VectorRef *outputs);
+                       VectorRef *outputs, std::vector<tensor::TensorPtr> *need_sync_outputs);
   void IncreaseSummaryRefCount(const session::NamedSummaryOutputs &summary_outputs);
   void DecreaseSummaryRefCount(const session::NamedSummaryOutputs &summary_outputs);
 
@@ -47,8 +49,10 @@ class CPUKernelRuntime : public KernelRuntime {
                                        TypeId type_id) override;
 
  private:
-  BaseRef CreatTensorForOutput(const AnfNodePtr &input_node, size_t index,
-                               const std::unordered_map<AnfNode *, tensor::TensorPtr> &input_map);
+  BaseRef CreatTensorForOutput(const session::KernelWithIndex &kernel_with_index,
+                               const std::unordered_map<AnfNode *, tensor::TensorPtr> &input_map,
+                               std::set<DeviceAddressPtr> *bound_addresses,
+                               std::vector<tensor::TensorPtr> *need_sync_outputs);
   void AssignValueNodeAddress(session::KernelGraph *kernel_graph);
   void AssignInputNodeAddress(const session::KernelGraph *kernel_graph);
   void AssignKernelOutputAddress(const session::KernelGraph *kernel_graph);
diff --git a/mindspore/ccsrc/kernel/cpu/gather_cpu_kernel.cc b/mindspore/ccsrc/kernel/cpu/gather_cpu_kernel.cc
index cb311043ac9..9117a533c8b 100644
--- a/mindspore/ccsrc/kernel/cpu/gather_cpu_kernel.cc
+++ b/mindspore/ccsrc/kernel/cpu/gather_cpu_kernel.cc
@@ -74,8 +74,8 @@ void GatherV2CPUKernel::CopyDataToOutput(const std::vector<kernel::AddressPtr> &
                                          size_t dim2, float **output_addr, size_t *buff_size) {
   auto input_addr = reinterpret_cast<float *>(inputs[0]->addr);
   auto indices_addr = reinterpret_cast<int *>(inputs[1]->addr);
-
-  for (size_t i = 0; i < output_shape_[axis_]; ++i) {
+  size_t elem_num = inputs[1]->size / 4;
+  for (size_t i = 0; i < elem_num; ++i) {
     size_t index = IntToSize(indices_addr[i]);
     size_t pos = 0;
     if (axis_ == 3) {
diff --git a/mindspore/ccsrc/session/cpu_session.cc b/mindspore/ccsrc/session/cpu_session.cc
index 8d6bc0f2b9e..e70e5510227 100644
--- a/mindspore/ccsrc/session/cpu_session.cc
+++ b/mindspore/ccsrc/session/cpu_session.cc
@@ -63,7 +63,8 @@ void CPUSession::RunGraph(const GraphId &graph_id, const std::vector<tensor::Ten
   auto &kernel_graph = graphs_[graph_id];
   MS_EXCEPTION_IF_NULL(kernel_graph);
   MS_LOG(INFO) << "Bind input output address";
-  runtime_.BindInputOutput(kernel_graph.get(), inputs, outputs);
+  std::vector<tensor::TensorPtr> need_sync_outputs;
+  runtime_.BindInputOutput(kernel_graph.get(), inputs, outputs, &need_sync_outputs);
   MS_LOG(INFO) << "Run graph start";
   predictmodel::StepConvertWeight(inputs);
   auto execution_order = kernel_graph->execution_order();
@@ -82,6 +83,9 @@ void CPUSession::RunGraph(const GraphId &graph_id, const std::vector<tensor::Ten
   if (!ret) {
     MS_LOG(EXCEPTION) << "Run graph failed";
   }
+  for (auto output : need_sync_outputs) {
+    (void)output->data_sync();
+  }
 
   if (enable_summary) {
     Summary(kernel_graph.get());