diff --git a/mindspore/ccsrc/debug/data_dump/e2e_dump_util.cc b/mindspore/ccsrc/debug/data_dump/e2e_dump_util.cc
index fc3682b670..09f305b431 100644
--- a/mindspore/ccsrc/debug/data_dump/e2e_dump_util.cc
+++ b/mindspore/ccsrc/debug/data_dump/e2e_dump_util.cc
@@ -121,8 +121,8 @@ void E2eDumpUtil::DumpOutput(const session::KernelGraph *graph, const std::strin
       auto type = AnfAlgo::GetOutputInferDataType(node, j);
       std::string file_path = dump_path + '/' + kernel_name + '_' + "output_" + std::to_string(j);
       if (IsDeviceTargetGPU()) {
-        DumpGPUMemToFile(file_path, node->fullname_with_scope() + "_output", NOT_NULL(addr), trans_flag, int_shapes,
-                         type, j, debugger);
+        DumpGPUMemToFile(file_path, node->fullname_with_scope(), NOT_NULL(addr), trans_flag, int_shapes, type, j,
+                         debugger);
       } else {
         DumpMemToFile(file_path, NOT_NULL(addr), trans_flag, int_shapes, type);
       }
@@ -155,13 +155,24 @@ void E2eDumpUtil::DumpInput(const session::KernelGraph *graph, const std::string
       auto index = kernel_with_index.second;
       auto addr = AnfAlgo::GetOutputAddr(input, index);
 
+      std::string tensor_name;
+      size_t slot;
+      if (IsDeviceTargetGPU()) {
+        auto input_kernel = node->input(j + 1);
+        std::string input_kernel_name = input_kernel->fullname_with_scope();
+        tensor_name = input_kernel_name;
+        slot = 0;
+      } else {
+        tensor_name = node->fullname_with_scope();
+        slot = j;
+      }
+
       ShapeVector int_shapes;
       GetDumpIntShape(input, index, trans_flag, NOT_NULL(&int_shapes));
       auto type = AnfAlgo::GetOutputInferDataType(input, index);
       std::string file_path = dump_path + '/' + kernel_name + '_' + "input_" + std::to_string(j);
       if (IsDeviceTargetGPU()) {
-        DumpGPUMemToFile(file_path, node->fullname_with_scope() + "_input", NOT_NULL(addr), trans_flag, int_shapes,
-                         type, j, debugger);
+        DumpGPUMemToFile(file_path, tensor_name, NOT_NULL(addr), trans_flag, int_shapes, type, slot, debugger);
       } else {
         DumpMemToFile(file_path, NOT_NULL(addr), trans_flag, int_shapes, type);
       }
@@ -249,7 +260,7 @@ void E2eDumpUtil::DumpSingleAnfnode(const AnfNodePtr &anf_node, const size_t out
 
   std::string file_path = dump_path + '/' + dump_name + '_' + "output_0";
   if (IsDeviceTargetGPU()) {
-    DumpGPUMemToFile(file_path, node_name + "_output", NOT_NULL(addr), trans_flag, int_shapes, type, 0, debugger);
+    DumpGPUMemToFile(file_path, node_name, NOT_NULL(addr), trans_flag, int_shapes, type, 0, debugger);
   } else {
     DumpMemToFile(file_path, NOT_NULL(addr), trans_flag, int_shapes, type);
   }
diff --git a/mindspore/ccsrc/debug/debugger/debugger.cc b/mindspore/ccsrc/debug/debugger/debugger.cc
index 2d67d62fc5..72ef147496 100644
--- a/mindspore/ccsrc/debug/debugger/debugger.cc
+++ b/mindspore/ccsrc/debug/debugger/debugger.cc
@@ -909,7 +909,7 @@ void Debugger::LoadSingleAnfnode(const AnfNodePtr &anf_node, const size_t output
   MS_EXCEPTION_IF_NULL(addr);
   auto type = AnfAlgo::GetOutputInferDataType(anf_node, output_index);
   auto format = kOpFormat_DEFAULT;
-  string tensor_name = node_name + "_output:" + "0";
+  string tensor_name = node_name + ':' + "0";
   ShapeVector int_shapes;
   auto shape = AnfAlgo::GetOutputDeviceShape(anf_node, output_index);
   (void)std::transform(shape.begin(), shape.end(), std::back_inserter(int_shapes),
diff --git a/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc b/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc
index 4cfeb00a75..216ecb0eee 100644
--- a/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc
+++ b/mindspore/ccsrc/runtime/device/gpu/gpu_kernel_runtime.cc
@@ -119,27 +119,22 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
     return;
   }
 
-  if (dump_json_parser.InputNeedDump()) {
+  if (debugger->debugger_enabled() || dump_json_parser.InputNeedDump()) {
     // get inputs
     auto input_size = AnfAlgo::GetInputTensorNum(kernel);
     for (size_t j = 0; j < input_size; ++j) {
-      auto kernel_with_index = AnfAlgo::GetPrevNodeOutput(kernel, j);
-      auto input_kernel = kernel_with_index.first;
-      auto index = kernel_with_index.second;
-
+      auto input_kernel = kernel->input(j + 1);
       std::string input_kernel_name = input_kernel->fullname_with_scope();
       auto addr = kernel_inputs[j];
       auto type = AnfAlgo::GetOutputInferDataType(input_kernel, PARAMETER_OUTPUT_INDEX);
       auto format = kOpFormat_DEFAULT;
       auto gpu_addr = std::make_unique<GPUDeviceAddress>(addr->addr, addr->size, format, type);
-      string input_tensor_name = input_kernel_name + "_output:" + std::to_string(index);
+      string input_tensor_name = input_kernel_name + ':' + "0";
       ShapeVector int_shapes;
       auto shape = AnfAlgo::GetOutputDeviceShape(input_kernel, PARAMETER_OUTPUT_INDEX);
       (void)std::transform(shape.begin(), shape.end(), std::back_inserter(int_shapes),
                            [](size_t inner_item) { return SizeToInt(inner_item); });
-      auto cur_node_name = kernel->fullname_with_scope();
-      auto ret = gpu_addr->LoadMemToHost(cur_node_name + "_input:" + std::to_string(j), exec_order, format, int_shapes,
-                                         type, 0, true);
+      auto ret = gpu_addr->LoadMemToHost(input_tensor_name, exec_order, format, int_shapes, type, 0, true);
       if (!ret) {
         MS_LOG(ERROR) << "LoadMemToHost:"
                       << ", tensor_name:" << input_tensor_name << ", host_format:" << format << ".!";
@@ -147,7 +142,7 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
     }
   }
 
-  if (dump_json_parser.OutputNeedDump()) {
+  if (debugger->debugger_enabled() || dump_json_parser.OutputNeedDump()) {
     // get outputs
     auto output_size = AnfAlgo::GetOutputTensorNum(kernel);
     auto node_name = AnfAlgo::GetCNodeName(kernel);
@@ -161,7 +156,7 @@ void LoadKernelData(Debugger *debugger, const CNodePtr &kernel,
       auto type = AnfAlgo::GetOutputInferDataType(kernel, j);
       auto format = kOpFormat_DEFAULT;
       auto gpu_addr = std::make_unique<GPUDeviceAddress>(addr->addr, addr->size, format, type);
-      string tensor_name = kernel_name + "_output:" + "0";
+      string tensor_name = kernel_name + ':' + std::to_string(j);
       ShapeVector int_shapes;
       auto shape = AnfAlgo::GetOutputDeviceShape(kernel, j);
       (void)std::transform(shape.begin(), shape.end(), std::back_inserter(int_shapes),