diff --git a/mindspore/ccsrc/backend/session/ascend_session.cc b/mindspore/ccsrc/backend/session/ascend_session.cc
index ee2d07e7241..e3638264a2c 100644
--- a/mindspore/ccsrc/backend/session/ascend_session.cc
+++ b/mindspore/ccsrc/backend/session/ascend_session.cc
@@ -644,7 +644,7 @@ void AscendSession::PostExecuteGraph(const std::shared_ptr<KernelGraph> &kernel_
   }
 #endif
 #ifndef ENABLE_SECURITY
-  DumpSetup(kernel_graph);
+  E2eDump::UpdateIterOldRTDump(kernel_graph.get());
 #endif
 }
 
@@ -1401,13 +1401,6 @@ void AscendSession::Execute(const std::shared_ptr<KernelGraph> &kernel_graph, bo
 }
 
 #ifndef ENABLE_SECURITY
-void AscendSession::DumpSetup(const std::shared_ptr<KernelGraph> &kernel_graph) const {
-  MS_LOG(DEBUG) << "Start!";
-  MS_EXCEPTION_IF_NULL(kernel_graph);
-  E2eDump::DumpSetup(kernel_graph.get());
-  MS_LOG(DEBUG) << "Finish!";
-}
-
 void AscendSession::Dump(const std::shared_ptr<KernelGraph> &kernel_graph) const {
   MS_LOG(DEBUG) << "Start!";
   MS_EXCEPTION_IF_NULL(kernel_graph);
diff --git a/mindspore/ccsrc/backend/session/ascend_session.h b/mindspore/ccsrc/backend/session/ascend_session.h
index cd18129d99d..19c37380f07 100644
--- a/mindspore/ccsrc/backend/session/ascend_session.h
+++ b/mindspore/ccsrc/backend/session/ascend_session.h
@@ -114,7 +114,6 @@ class AscendSession : public SessionBasic {
   void Execute(const std::shared_ptr<KernelGraph> &kernel_graph, bool is_task) const;
 #ifndef ENABLE_SECURITY
   void Dump(const std::shared_ptr<KernelGraph> &kernel_graph) const;
-  void DumpSetup(const std::shared_ptr<KernelGraph> &kernel_graph) const;
 #endif
   void LoadTensor(const std::shared_ptr<KernelGraph> &kernel_graph) const;
   // below functions are used for run op
diff --git a/mindspore/ccsrc/backend/session/gpu_session.cc b/mindspore/ccsrc/backend/session/gpu_session.cc
index 2880c4134be..05e93742dbe 100644
--- a/mindspore/ccsrc/backend/session/gpu_session.cc
+++ b/mindspore/ccsrc/backend/session/gpu_session.cc
@@ -518,7 +518,7 @@ void GPUSession::PreExecuteGraph(const std::shared_ptr<KernelGraph> &kernel_grap
     debugger_->PreExecute(kernel_graph);
   }
 
-  DumpSetup(kernel_graph);
+  E2eDump::UpdateIterOldRTDump(kernel_graph.get());
 #endif
 
 #if ENABLE_CPU && ENABLE_GPU
@@ -725,12 +725,6 @@ void GPUSession::RunOpImpl(const GraphInfo &graph_info, OpRunInfo *op_run_info,
 }
 
 #ifdef ENABLE_DEBUGGER
-void GPUSession::DumpSetup(const std::shared_ptr<KernelGraph> &kernel_graph) const {
-  MS_LOG(INFO) << "Start!";
-  MS_EXCEPTION_IF_NULL(kernel_graph);
-  E2eDump::DumpSetup(kernel_graph.get());
-  MS_LOG(INFO) << "Finish!";
-}
 
 void GPUSession::Dump(const std::shared_ptr<KernelGraph> &kernel_graph) const {
   // Dump graph and graph history file if e2e_dump is enabled and update cur_dump_iter for GPU old runtime.
diff --git a/mindspore/ccsrc/backend/session/gpu_session.h b/mindspore/ccsrc/backend/session/gpu_session.h
index 4b722fc8444..acbf3ebcce4 100644
--- a/mindspore/ccsrc/backend/session/gpu_session.h
+++ b/mindspore/ccsrc/backend/session/gpu_session.h
@@ -94,8 +94,6 @@ class GPUSession : public SessionBasic {
 #ifdef ENABLE_DEBUGGER
   void Dump(const std::shared_ptr<KernelGraph> &kernel_graph) const;
 
-  void DumpSetup(const std::shared_ptr<KernelGraph> &kernel_graph) const;
-
   bool DumpDataEnabledIteration() const;
 #endif
 
diff --git a/mindspore/ccsrc/backend/session/kernel_graph.cc b/mindspore/ccsrc/backend/session/kernel_graph.cc
index 2464cbb1d5e..5b3f4a1abf6 100644
--- a/mindspore/ccsrc/backend/session/kernel_graph.cc
+++ b/mindspore/ccsrc/backend/session/kernel_graph.cc
@@ -1422,6 +1422,10 @@ void KernelGraph::SetOptimizerFlag() {
 bool KernelGraph::IsDatasetGraph() const {
   // check if there is InitDataSetQueue node
   const auto &nodes = execution_order_;
+  // The size of execution_order for the dataset graph is equal to 1.
+  if (execution_order_.size() > 1) {
+    return false;
+  }
   for (const auto &node : nodes) {
     auto node_name = AnfAlgo::GetCNodeName(node);
     if (node_name == prim::kPrimInitDataSetQueue->name()) {
diff --git a/mindspore/ccsrc/debug/data_dump/dump_utils.cc b/mindspore/ccsrc/debug/data_dump/dump_utils.cc
index 1ceec4ef639..520b6d21ec6 100644
--- a/mindspore/ccsrc/debug/data_dump/dump_utils.cc
+++ b/mindspore/ccsrc/debug/data_dump/dump_utils.cc
@@ -26,6 +26,9 @@
 #include "runtime/device/kernel_runtime_manager.h"
 #include "utils/utils.h"
 #include "debug/common.h"
+#include "runtime/framework/device_tensor_store.h"
+
+using mindspore::runtime::DeviceTensorStore;
 
 namespace mindspore {
 uint32_t ConvertPhysicalDeviceId(uint32_t device_id) {
@@ -90,6 +93,24 @@ void GetDumpIntShape(const AnfNodePtr &node, size_t index, NotNull<ShapeVector *
   }
 }
 
+const DeviceTensorPtr GetParameterInfo(const AnfNodePtr &node, NotNull<ShapeVector *> int_shapes,
+                                       NotNull<TypeId *> host_type, NotNull<TypeId *> device_type) {
+  const auto &device_tensors = DeviceTensorStore::GetInstance().Fetch(node.get());
+  if (device_tensors.size() < 1) {
+    return nullptr;
+  }
+  auto device_addr = device_tensors[0];
+  MS_EXCEPTION_IF_NULL(device_addr);
+  auto &dump_json_parser = DumpJsonParser::GetInstance();
+  bool trans_flag = dump_json_parser.trans_flag();
+  auto ref_node = device_addr->GetNodeIndex().first;
+  MS_EXCEPTION_IF_NULL(ref_node);
+  GetDumpIntShape(ref_node, PARAMETER_OUTPUT_INDEX, int_shapes, trans_flag);
+  *host_type = AnfAlgo::GetOutputInferDataType(ref_node, PARAMETER_OUTPUT_INDEX);
+  *device_type = AnfAlgo::GetOutputDeviceDataType(ref_node, PARAMETER_OUTPUT_INDEX);
+  return device_addr;
+}
+
 /*
  * Feature group: Dump.
  * Target device group: Ascend, CPU.
diff --git a/mindspore/ccsrc/debug/data_dump/dump_utils.h b/mindspore/ccsrc/debug/data_dump/dump_utils.h
index c5ac74d028f..0b27748d8a3 100644
--- a/mindspore/ccsrc/debug/data_dump/dump_utils.h
+++ b/mindspore/ccsrc/debug/data_dump/dump_utils.h
@@ -19,10 +19,14 @@
 
 #include <map>
 #include <string>
+#include <memory>
 
 #include "backend/session/kernel_graph.h"
 #include "runtime/device/device_address.h"
 
+using DeviceTensor = mindspore::device::DeviceAddress;
+using DeviceTensorPtr = std::shared_ptr<DeviceTensor>;
+
 namespace mindspore {
 static const size_t PARAMETER_OUTPUT_INDEX = 0;
 static const size_t VALUE_NODE_OUTPUT_INDEX = 0;
@@ -33,6 +37,9 @@ void GetFileKernelName(NotNull<std::string *> kernel_name);
 
 void GetDumpIntShape(const AnfNodePtr &node, size_t index, NotNull<ShapeVector *> int_shapes, bool trans_flag = false);
 
+const DeviceTensorPtr GetParameterInfo(const AnfNodePtr &node, NotNull<ShapeVector *> int_shapes,
+                                       NotNull<TypeId *> host_type, NotNull<TypeId *> device_type);
+
 void DumpMemToFile(const std::string &file_path, const device::DeviceAddress &addr, const ShapeVector &int_shapes,
                    const TypeId &type, bool trans_flag = false);
 // Get time stamp since epoch in microseconds
diff --git a/mindspore/ccsrc/debug/data_dump/e2e_dump.cc b/mindspore/ccsrc/debug/data_dump/e2e_dump.cc
index d0be2e47fea..39bf6b2d874 100644
--- a/mindspore/ccsrc/debug/data_dump/e2e_dump.cc
+++ b/mindspore/ccsrc/debug/data_dump/e2e_dump.cc
@@ -305,7 +305,6 @@ void E2eDump::DumpSingleAnfNode(const AnfNodePtr &anf_node, const size_t output_
     dump_name = node_name.substr(cst_prefix.length());
     trans_flag = false;
   }
-
   // check if output address exists, if not, return;
   if (!AnfAlgo::OutputAddrExist(anf_node, output_index)) {
     return;
@@ -334,6 +333,49 @@ void E2eDump::DumpSingleAnfNode(const AnfNodePtr &anf_node, const size_t output_
   }
 }
 
+/*
+ * Feature group: Dump.
+ * Target device group: Ascend, GPU.
+ * Runtime category: MindRT.
+ * Description: This function is similar to DumpSingleAnfNode function but it is only for dumping parameters in mindRT.
+ * This function uses GetParameterInfo to get dump info for the parameter node.
+ */
+void E2eDump::DumpSingleParameterNode(const AnfNodePtr &anf_node, const std::string &dump_path, bool trans_flag,
+                                      const Debugger *debugger) {
+  MS_EXCEPTION_IF_NULL(anf_node);
+  auto &dump_json_parser = DumpJsonParser::GetInstance();
+  std::string node_name = GetKernelNodeName(anf_node);
+  if (!anf_node->isa<Parameter>() || !dump_json_parser.NeedDump(node_name)) {
+    return;
+  }
+  DumpJsonParser::GetInstance().MatchKernel(node_name);
+  GetFileKernelName(NOT_NULL(&node_name));
+  ShapeVector int_shapes;
+  TypeId type;
+  TypeId device_type;
+  auto addr = GetParameterInfo(anf_node, NOT_NULL(&int_shapes), NOT_NULL(&type), NOT_NULL(&device_type));
+  if (addr == nullptr) {
+    MS_LOG(DEBUG) << "Skip node: " << node_name << ". Parameter data is not available for mindRT.";
+    return;
+  }
+  uint64_t timestamp = GetTimeStamp();
+  uint32_t task_id = 0;
+  uint32_t stream_id = 0;
+  std::string file_path = dump_path + "/Parameter." + node_name + '.' + std::to_string(task_id) + '.' +
+                          std::to_string(stream_id) + '.' + std::to_string(timestamp) + ".output.0";
+  if (IsDeviceTargetGPU()) {
+    if (dump_json_parser.IsStatisticDump()) {
+      TensorStatDump stat_dump("Parameter", node_name, task_id, stream_id, timestamp, false, 0, 0);
+      stat_dump.DumpTensorStatsToFile(node_name, dump_path, debugger);
+    }
+    if (dump_json_parser.IsTensorDump()) {
+      DumpGPUMemToFile(file_path, node_name, *addr, int_shapes, type, device_type, trans_flag, 0, debugger);
+    }
+  } else {
+    DumpMemToFile(file_path, *addr, int_shapes, type, trans_flag);
+  }
+}
+
 void E2eDump::DumpParameters(const session::KernelGraph *graph, const std::string &dump_path,
                              const Debugger *debugger) {
   MS_EXCEPTION_IF_NULL(graph);
@@ -380,9 +422,16 @@ void E2eDump::DumpConstantData(const session::KernelGraph *graph, const std::str
   }
 }
 
-void E2eDump::UpdateIterDumpSetup(const session::KernelGraph *graph, bool sink_mode) {
-  uint32_t graph_id = graph->graph_id();
+/*
+ * Feature group: Dump.
+ * Target device group: Ascend, GPU.
+ * Runtime category: Old runtime.
+ * Description: This function is for updating dump iteration for GPU and ascend old runtime.
+ */
+void E2eDump::UpdateIterOldRTDump(const session::KernelGraph *graph) {
+  MS_EXCEPTION_IF_NULL(graph);
   auto &dump_json_parser = DumpJsonParser::GetInstance();
+  uint32_t graph_id = graph->graph_id();
   if (IsDeviceTargetGPU()) {
     if (starting_graph_id == INT32_MAX) {
       starting_graph_id = graph_id;
@@ -394,7 +443,7 @@ void E2eDump::UpdateIterDumpSetup(const session::KernelGraph *graph, bool sink_m
     return;
   }
   // If device target is Ascend
-  if (sink_mode && graph->IsDatasetGraph()) {
+  if (graph->IsDatasetGraph()) {
     MS_LOG(INFO) << "No need to update iteration for dataset graph.";
     return;
   }
@@ -403,29 +452,23 @@ void E2eDump::UpdateIterDumpSetup(const session::KernelGraph *graph, bool sink_m
   dump_json_parser.UpdateDumpIter();
 }
 
-/*
- * Feature group: Dump.
- * Target device group: Ascend, GPU.
- * Runtime category: Old runtime, MindRT.
- * Description: This function is for updating dump iteration for GPU and ascend old runtime and ascend super
- * kernel MindRT.
- */
-void E2eDump::DumpSetup(const session::KernelGraph *graph) {
-  auto &dump_json_parser = DumpJsonParser::GetInstance();
-  bool sink_mode = (ConfigManager::GetInstance().dataset_mode() || E2eDump::isDatasetGraph(graph));
-
-  if (dump_json_parser.async_dump_enabled() || dump_json_parser.e2e_dump_enabled()) {
-    UpdateIterDumpSetup(graph, sink_mode);
-  }
-}
-
 /*
  * Feature group: Dump.
  * Target device group: Ascend, GPU.
  * Runtime category: MindRT.
- * Description: This function is for updating dump iteration for GPU and kernel by kernel ascend MindRT dump.
+ * Description: This function is for updating dump iteration for GPU and ascend MindRT dump. Please note that dump with
+ * dataset_sink_mode = True is not supported for GPU.
  */
 void E2eDump::UpdateIterMindRTDump() {
+  auto debugger = Debugger::GetInstance();
+  // Dataset graph is always the first graph in the list when dataset_sink_mode is true.
+  auto graph = (debugger->GetStepGraphPtrList())[0];
+  auto context = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(context);
+  if (context->get_param<std::string>(MS_CTX_DEVICE_TARGET) == kAscendDevice && graph->IsDatasetGraph()) {
+    MS_LOG(INFO) << "No need to update iteration for dataset graph.";
+    return;
+  }
   // update dump iter for GPU and kernel by kernel ascend dump.
   DumpJsonParser::GetInstance().UpdateDumpIter();
 }
@@ -464,7 +507,7 @@ void E2eDump::DumpRunIter(const KernelGraphPtr &graph, uint32_t rank_id) {
     MS_LOG(WARNING) << "Open file for saving graph global execution order failed.";
     return;
   }
-  if (sink_mode && json_parser.async_dump_enabled()) {
+  if (sink_mode && json_parser.async_dump_enabled() && !Debugger::GetInstance()->GetAscendKernelByKernelFlag()) {
     // for async dump when sink_mode = true, cur_dump_iter() = current_epoch
     // dump history for all iterations in the epoch
     Debugger::GetInstance()->UpdateGraphIterMap(graph->graph_id(), iter_num);
@@ -501,16 +544,16 @@ void E2eDump::DumpData(const session::KernelGraph *graph, uint32_t rank_id, cons
     MS_LOG(INFO) << "Start e2e dump. Current iteration is " << dump_json_parser.cur_dump_iter();
     MS_LOG(INFO) << "Current graph id is " << graph_id;
     std::string dump_path = GenerateDumpPath(graph_id, rank_id);
-    std::string cst_path = GenerateDumpPath(graph_id, rank_id, true);
-
     if (dump_json_parser.IsStatisticDump()) {
       TensorStatDump::OpenStatisticsFile(dump_path);
     }
     DumpInput(graph, dump_path, debugger);
     DumpOutput(graph, dump_path, debugger);
-    DumpParameters(graph, dump_path, debugger);
-    if (IsDeviceTargetGPU() && dump_json_parser.e2e_dump_enabled()) {
-      DumpConstantData(graph, cst_path, debugger);
+    if (!MsContext::GetInstance()->get_param<bool>(MS_CTX_ENABLE_MINDRT)) {
+      // Dump parameters for old runtime. For mindRT it is done in PostExecuteGraphDebugger.
+      DumpParameters(graph, dump_path, debugger);
+      // DumpConstantData for GPU old runtime.
+      DumpConstantData(graph, rank_id, debugger);
     }
     if (dump_json_parser.IsStatisticDump()) {
       CsvWriter::GetInstance().CloseFile();
@@ -543,29 +586,29 @@ bool E2eDump::DumpSingleNodeData(const CNodePtr &node, uint32_t graph_id, uint32
   return success;
 }
 
-bool E2eDump::DumpParametersData(const session::KernelGraph *graph, uint32_t rank_id, const Debugger *debugger) {
-  bool success = false;
-  uint32_t graph_id = graph->graph_id();
+/*
+ * Feature group: Dump.
+ * Target device group: Ascend, GPU.
+ * Runtime category: MindRT.
+ * Description: This function is for dumping all the parameters in the current root graph for GPU, Ascend superkernel
+ * (e2e dump) and Ascend kernel-by-kernel (e2e and async dump).
+ */
+void E2eDump::DumpParametersData(uint32_t rank_id, const Debugger *debugger) {
+  uint32_t root_graph_id = debugger->GetCurrentRootGraphId();
   auto &dump_json_parser = DumpJsonParser::GetInstance();
+  if (dump_json_parser.async_dump_enabled() && !debugger->GetAscendKernelByKernelFlag()) {
+    // Dump parameters for mindRT in async dump only for kernel by kernel mode.
+    return;
+  }
   if (dump_json_parser.DumpEnabledForIter()) {
     MS_LOG(INFO) << "DumpParameters. Current iteration is " << dump_json_parser.cur_dump_iter();
-    MS_LOG(INFO) << "Current graph id is " << graph_id;
-    std::string dump_path = GenerateDumpPath(graph_id, rank_id);
-    DumpParameters(graph, dump_path, debugger);
-    success = true;
-  }
-  return success;
-}
-bool E2eDump::isDatasetGraph(const session::KernelGraph *graph) {
-  // check if there is GetNext or InitDataSetQueue node
-  const auto &nodes = graph->execution_order();
-  for (const auto &node : nodes) {
-    auto node_name = AnfAlgo::GetCNodeName(node);
-    if (node_name == prim::kPrimGetNext->name() || node_name == prim::kPrimInitDataSetQueue->name()) {
-      return true;
+    MS_LOG(INFO) << "Current root graph id is " << root_graph_id;
+    std::string dump_path = GenerateDumpPath(root_graph_id, rank_id);
+    bool trans_flag = dump_json_parser.trans_flag();
+    for (auto &item : debugger->GetParametersMindRT()) {
+      DumpSingleParameterNode(item, dump_path, trans_flag, debugger);
     }
   }
-  return false;
 }
 
 #ifdef ENABLE_D
diff --git a/mindspore/ccsrc/debug/data_dump/e2e_dump.h b/mindspore/ccsrc/debug/data_dump/e2e_dump.h
index 1bef8a2f7f2..ae475f8356b 100644
--- a/mindspore/ccsrc/debug/data_dump/e2e_dump.h
+++ b/mindspore/ccsrc/debug/data_dump/e2e_dump.h
@@ -38,10 +38,10 @@ class E2eDump {
  public:
   E2eDump() = default;
   ~E2eDump() = default;
-  static void DumpSetup(const session::KernelGraph *graph);
-
   static void UpdateIterMindRTDump();
 
+  static void UpdateIterOldRTDump(const session::KernelGraph *graph);
+
   static void DumpRunIter(const KernelGraphPtr &graph_ptr, uint32_t rank_id = 0);
 
   static void DumpData(const session::KernelGraph *graph, uint32_t rank_id, const Debugger *debugger = nullptr);
@@ -51,13 +51,11 @@ class E2eDump {
 
   static void DumpConstantData(const session::KernelGraph *graph, uint32_t rank_id, const Debugger *debugger = nullptr);
 
-  static bool DumpParametersData(const session::KernelGraph *graph, uint32_t rank_id, const Debugger *debugger);
+  static void DumpParametersData(uint32_t rank_id, const Debugger *debugger);
 
   static bool DumpSingleNodeData(const CNodePtr &node, uint32_t graph_id, uint32_t rank_id,
                                  const Debugger *debugger = nullptr);
 
-  static bool isDatasetGraph(const session::KernelGraph *graph);
-
   // Dump data when task error.
   static void DumpInputImpl(const CNodePtr &node, bool trans_flag, const std::string &dump_path,
                             std::string *kernel_name, const Debugger *debugger);
@@ -91,7 +89,8 @@ class E2eDump {
   static void DumpSingleAnfNode(const AnfNodePtr &anf_node, const size_t output_index, const std::string &dump_path,
                                 bool trans_flag, const Debugger *debugger);
 
-  static void UpdateIterDumpSetup(const session::KernelGraph *graph, bool sink_mode);
+  static void DumpSingleParameterNode(const AnfNodePtr &anf_node, const std::string &dump_path, bool trans_flag,
+                                      const Debugger *debugger);
 
 #ifdef ENABLE_D
   static nlohmann::json ParseOverflowInfo(char *data_ptr);
diff --git a/mindspore/ccsrc/debug/debugger/debugger.cc b/mindspore/ccsrc/debug/debugger/debugger.cc
index 74fb277b0e6..8cdcdbc99dd 100644
--- a/mindspore/ccsrc/debug/debugger/debugger.cc
+++ b/mindspore/ccsrc/debug/debugger/debugger.cc
@@ -38,6 +38,7 @@
 #include "runtime/hardware/device_context_manager.h"
 #include "debug/anf_ir_dump.h"
 #include "debug/anf_ir_utils.h"
+#include "runtime/framework/device_tensor_store.h"
 #ifdef ENABLE_DEBUGGER
 #include "debug/debugger/proto_exporter.h"
 #else
@@ -56,6 +57,7 @@ using debugger::WatchCondition_Condition_nan;
 using debugger::WatchCondition_Parameter;
 using debugger::WatchNode;
 using debugger::WatchpointHit;
+using mindspore::runtime::DeviceTensorStore;
 
 namespace mindspore {
 
@@ -287,6 +289,8 @@ void Debugger::Reset() {
   graph_proto_list_.clear();
   graph_ptr_list_.clear();
   graph_ptr_step_vec_.clear();
+  parameters_mindRT_.clear();
+  visited_root_graph_ids_.clear();
   MS_LOG(INFO) << "Release Debugger resource.";
 }
 
@@ -297,13 +301,15 @@ void Debugger::Reset() {
  * Description: Sets root_graph_id for all the graphs in the compiled graph list. Sets cur_root_graph_id_ and
  * prev_root_graph_id_ and calls PreExecute function for all the graphs.
  */
-void Debugger::PreExecuteGraphDebugger(const std::vector<KernelGraphPtr> &graphs) {
+void Debugger::PreExecuteGraphDebugger(const std::vector<KernelGraphPtr> &graphs,
+                                       const std::vector<AnfNodePtr> &origin_parameters_order) {
   // MindRTBackend for GPU and Ascend
   if (device_target_ == kCPUDevice) {
     return;
   }
   // Store graphs that are run in one step.
   graph_ptr_step_vec_ = graphs;
+  parameters_mindRT_ = origin_parameters_order;
   prev_root_graph_id_ = cur_root_graph_id_;
   // set first run graph as the root graph
   cur_root_graph_id_ = graph_ptr_step_vec_[0]->graph_id();
@@ -474,21 +480,31 @@ uint32_t Debugger::GetRankID() {
  * Feature group: Dump.
  * Target device group: Ascend, GPU.
  * Runtime category: MindRT.
- * Description: Dumps graph history and parameters for GPU and Ascend kernel-by-kernel MindRT. DumpConstantData for GPU.
+ * Description: When dump is enabled, this function: 1) Dumps parameters for the current root_graph_id to the
+ * root_graph's directory. 2) Dumps constant data once for each graph. 3) Dumps graph run history for each graph.
  */
-void Debugger::Dump(const KernelGraphPtr &kernel_graph) const {
-  if (!(ascend_kernel_by_kernel_ || device_target_ == kGPUDevice)) {
+void Debugger::DumpParamsAndConstAndHistory() {
+  if (!CheckDebuggerDumpEnabled()) {
     return;
   }
-  uint32_t rank_id = GetRankID();
-  E2eDump::DumpRunIter(kernel_graph, rank_id);
-  if (debugger_ && debugger_->DebuggerBackendEnabled()) {
-    MS_EXCEPTION_IF_NULL(kernel_graph);
-    (void)E2eDump::DumpParametersData(kernel_graph.get(), rank_id, debugger_.get());
-    // Dump constant data for GPU mindRT.
-    E2eDump::DumpConstantData(kernel_graph.get(), rank_id, debugger_.get());
-  } else {
-    DumpJsonParser::GetInstance().UpdateDumpIter();
+  LoadParametersAllGraphs();
+  (void)E2eDump::DumpParametersData(GetRankID(), debugger_.get());
+  // Whether constant data was already dumped for the current root graph.
+  bool cur_root_graph_checked = std::find(visited_root_graph_ids_.begin(), visited_root_graph_ids_.end(),
+                                          cur_root_graph_id_) != visited_root_graph_ids_.end();
+  for (auto graph : graph_ptr_step_vec_) {
+    if (!cur_root_graph_checked) {
+      LoadConstsForGraph(graph);
+      // Dump constant data for GPU.
+      E2eDump::DumpConstantData(graph.get(), GetRankID(), debugger_.get());
+      // Dump constant data for Ascend.
+      DumpConstantDataAscend(graph);
+    }
+    // Dump graph run hisotry for each graph.
+    E2eDump::DumpRunIter(graph, GetRankID());
+  }
+  if (!cur_root_graph_checked) {
+    visited_root_graph_ids_.push_back(cur_root_graph_id_);
   }
 }
 
@@ -556,29 +572,15 @@ void Debugger::PostExecuteGraphDebugger() {
     DumpJsonParser::GetInstance().UpdateDumpIter();
     return;
   }
-  // LoadParametersAndConst for all the graphs that have been run in the current step
-  if (debugger_ && device_target_ == kGPUDevice) {
-    for (auto graph : graph_ptr_step_vec_) {
-      debugger_->LoadParametersAndConst(graph);
-    }
-  }
+  DumpParamsAndConstAndHistory();
   // debug used for dump
-  if (debugger_ && debugger_->CheckDebuggerDumpEnabled()) {
-    // Dump Parameters and consts
-    for (auto graph : graph_ptr_step_vec_) {
-      debugger_->Dump(graph);
-      DumpConstantDataAscend(graph);
-      if (!debugger_->debugger_enabled()) {
-        debugger_->ClearCurrentData();
-      }
-    }
+  if (CheckDebuggerDumpEnabled() && !debugger_enabled()) {
+    ClearCurrentData();
   }
   if (debugger_) {
     debugger_->PostExecute();
   }
-  if (ascend_kernel_by_kernel_ || device_target_ == kGPUDevice) {
-    E2eDump::UpdateIterMindRTDump();
-  }
+  E2eDump::UpdateIterMindRTDump();
 }
 
 /*
@@ -1341,7 +1343,10 @@ bool Debugger::DumpTensorToFile(const std::string &tensor_name, bool trans_flag,
 }
 
 bool Debugger::LoadNewTensor(const std::shared_ptr<TensorData> &tensor, bool keep_prev) {
-  return debug_services_.get()->LoadNewTensor(tensor, keep_prev);
+  if (debug_services_ != nullptr) {
+    return debug_services_.get()->LoadNewTensor(tensor, keep_prev);
+  }
+  return false;
 }
 
 bool Debugger::debugger_enabled() const { return debugger_enabled_; }
@@ -1543,6 +1548,37 @@ void Debugger::LoadSingleAnfnode(const AnfNodePtr &anf_node, const size_t output
   }
 }
 
+void Debugger::LoadSingleParameterMindRT(const AnfNodePtr &node) {
+  MS_EXCEPTION_IF_NULL(node);
+  auto root_graph_id = cur_root_graph_id_;
+  // This function is only  for loading parameters mindRT.
+  std::string node_name = GetKernelNodeName(node);
+  GetFileKernelName(NOT_NULL(&node_name));
+  TypeId type;
+  TypeId device_type;
+  ShapeVector int_shapes;
+  auto device_addr = GetParameterInfo(node, NOT_NULL(&int_shapes), NOT_NULL(&type), NOT_NULL(&device_type));
+  if (device_addr == nullptr) {
+    MS_LOG(DEBUG) << "Skip node: " << node_name << ". Parameter data is not available for mindRT.";
+    return;
+  }
+  if (!IsTypeDebuggerSupported(type)) {
+    return;
+  }
+  auto format = kOpFormat_DEFAULT;
+  string tensor_name = node_name + ':' + "0";
+  if (debug_services_ != nullptr) {
+    debug_services_->MoveTensorCurrentToPrev(tensor_name);
+  }
+  // Keep_prev is True for parameters.
+  bool ret = device_addr->LoadMemToHost(tensor_name, 0, format, int_shapes, type, 0, true, root_graph_id);
+
+  if (!ret) {
+    MS_LOG(ERROR) << "LoadMemToHost:"
+                  << ", tensor_name:" << tensor_name << ", host_format:" << format << ".!";
+  }
+}
+
 /*
  * Feature group: Dump, Online debugger.
  * Target device group: Ascend, GPU.
@@ -1593,6 +1629,43 @@ void Debugger::LoadParametersAndConst(const KernelGraphPtr &graph) {
   }
 }
 
+/*
+ * Feature group: Dump.
+ * Target device group: GPU.
+ * Runtime category: MindRT.
+ * Description: This function is for loading parameters' data from device to host into tensor_list_map_ for GPU dump.
+ * Ascend does not use tensor_map_list_ for dump so it is not needed for ascend dump.
+ */
+void Debugger::LoadParametersAllGraphs() {
+  if (!(device_target_ == kGPUDevice && CheckDebuggerDumpEnabled())) {
+    return;
+  }
+  for (auto &node : parameters_mindRT_) {
+    LoadSingleParameterMindRT(node);
+  }
+}
+
+/*
+ * Feature group: Dump.
+ * Target device group: GPU.
+ * Runtime category: MindRT.
+ * Description: This function is for loading constant data from device to host into tensor_list_map_ for GPU dump.
+ * Ascend does not use tensor_map_list_ for dump so it is not needed for ascend dump.
+ */
+void Debugger::LoadConstsForGraph(const KernelGraphPtr &graph) {
+  if (!(device_target_ == kGPUDevice && CheckDebuggerDumpEnabled())) {
+    return;
+  }
+  // load value nodes
+  // get all constant values from the graph
+  MS_LOG(INFO) << "Start to load value nodes for graph " << graph->graph_id() << ".";
+  auto root_graph_id = graph->root_graph_id();
+  const auto value_nodes = graph->graph_value_nodes();
+  for (auto &item : value_nodes) {
+    LoadSingleAnfnode(item, VALUE_NODE_OUTPUT_INDEX, root_graph_id);
+  }
+}
+
 /*
  * Feature group: Online debugger.
  * Target device group: Ascend.
@@ -1683,7 +1756,10 @@ void Debugger::ClearCurrentData() {
 }
 
 bool Debugger::TensorExistsInCurrent(const std::string &tensor_name) {
-  return debug_services_->TensorExistsInCurrent(tensor_name);
+  if (debug_services_ != nullptr) {
+    return debug_services_->TensorExistsInCurrent(tensor_name);
+  }
+  return false;
 }
 
 #ifdef ENABLE_D
diff --git a/mindspore/ccsrc/debug/debugger/debugger.h b/mindspore/ccsrc/debug/debugger/debugger.h
index 52f368d8f23..51c631158a0 100644
--- a/mindspore/ccsrc/debug/debugger/debugger.h
+++ b/mindspore/ccsrc/debug/debugger/debugger.h
@@ -29,6 +29,7 @@
 #ifdef ENABLE_D
 #include "debug/dump_data_builder.h"
 #endif
+#include "runtime/device/device_address.h"
 
 using debugger::Chunk;
 using debugger::DataType;
@@ -41,6 +42,8 @@ using debugger::WatchCondition;
 using debugger::WatchCondition_Parameter;
 using debugger::WatchNode;
 using debugger::WatchpointHit;
+using DeviceTensor = mindspore::device::DeviceAddress;
+using DeviceTensorPtr = std::shared_ptr<DeviceTensor>;
 
 template <class T>
 using ProtoVector = google::protobuf::RepeatedPtrField<T>;
@@ -77,7 +80,8 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
   // reset debugger
   void Reset();
 
-  void PreExecuteGraphDebugger(const std::vector<KernelGraphPtr> &graphs);
+  void PreExecuteGraphDebugger(const std::vector<KernelGraphPtr> &graphs,
+                               const std::vector<AnfNodePtr> &origin_parameters_order);
   // enable debugger
   // send graph and wait for command
   // do nothing if graph is set already
@@ -87,6 +91,8 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
 
   void SetAscendKernelByKernelFlag(bool value) { ascend_kernel_by_kernel_ = value; }
 
+  bool GetAscendKernelByKernelFlag() const { return ascend_kernel_by_kernel_; }
+
   void StoreRunGraphIdList(uint32_t graph_id);
 
   // analyze tensors and wait for command
@@ -97,8 +103,6 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
 
   static uint32_t GetRankID();
 
-  void Dump(const KernelGraphPtr &kernel_graph) const;
-
   void DumpConstantDataAscend(const KernelGraphPtr &graph);
 
   void DumpSingleNode(const CNodePtr &node, uint32_t graph_id);
@@ -144,6 +148,12 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
 
   void LoadParametersAndConst(const KernelGraphPtr &graph);
 
+  void LoadParametersAllGraphs();
+
+  void LoadConstsForGraph(const KernelGraphPtr &graph);
+
+  void DumpParamsAndConstAndHistory();
+
   void UpdateStepNum(const session::KernelGraph *graph);
 
   void UpdateStepNumGPU();
@@ -162,6 +172,8 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
 
   uint32_t GetPrevRootGraphId() const { return prev_root_graph_id_; }
 
+  std::vector<KernelGraphPtr> GetStepGraphPtrList() const { return graph_ptr_step_vec_; }
+
   void SetGraphPtr(const KernelGraphPtr &graph_ptr) { graph_ptr_ = graph_ptr; }
 
   const KernelGraphPtr GetGraphPtr() const { return graph_ptr_; }
@@ -180,6 +192,8 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
 
   void UpdateGraphIterMap(uint32_t graph_id, int32_t iter_num);
 
+  std::vector<AnfNodePtr> GetParametersMindRT() const { return parameters_mindRT_; }
+
 #ifdef ENABLE_D
   std::shared_ptr<DumpDataBuilder> LoadDumpDataBuilder(const std::string &node_name);
 
@@ -271,6 +285,8 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
 
   void LoadSingleAnfnode(const AnfNodePtr &anf_node, const size_t output_index, uint32_t root_graph_id);
 
+  void LoadSingleParameterMindRT(const AnfNodePtr &anf_node);
+
   // class members
 
   std::unique_ptr<GrpcClient> grpc_client_;
@@ -301,6 +317,9 @@ class Debugger : public std::enable_shared_from_this<Debugger> {
   std::list<KernelGraphPtr> graph_ptr_list_;
   // The vector of graph pointers that have been run in the current step.
   std::vector<KernelGraphPtr> graph_ptr_step_vec_;
+  // The vector of all the parameters for the current step for mindRT.
+  std::vector<AnfNodePtr> parameters_mindRT_;
+  std::vector<uint32_t> visited_root_graph_ids_;
 
   // map to store iter num in each epoch when dataset_sink_mode is true
   std::map<uint32_t, int32_t> graph_iter_num_map_;
diff --git a/mindspore/ccsrc/debug/debugger/debugger_utils.cc b/mindspore/ccsrc/debug/debugger/debugger_utils.cc
index 82a7bc8c66e..a3d1897b209 100644
--- a/mindspore/ccsrc/debug/debugger/debugger_utils.cc
+++ b/mindspore/ccsrc/debug/debugger/debugger_utils.cc
@@ -159,12 +159,18 @@ bool CheckReadData(const CNodePtr &cnode) {
   return read_data;
 }
 
+bool IsDeviceTargetGPU() {
+  auto context = MsContext::GetInstance();
+  MS_EXCEPTION_IF_NULL(context);
+  return context->get_param<std::string>(MS_CTX_DEVICE_TARGET) == kGPUDevice;
+}
+
 /*
  * Feature group: Dump, Online debugger.
- * Target device group: GPU.
+ * Target device group: Ascend, GPU.
  * Runtime category: MindRT.
  * Description: Load inputs and outputs of the given node if needed and dump them if dump is enabled, then it performs
- * PostExecuteNode function on the given node.
+ * PostExecuteNode function on the given node for GPU.
  */
 void ReadDataAndDump(const CNodePtr &cnode, const KernelLaunchInfo *launch_info, uint32_t exec_order,
                      const DeviceContext *device_context) {
@@ -194,9 +200,11 @@ void ReadDataAndDump(const CNodePtr &cnode, const KernelLaunchInfo *launch_info,
       debugger->ClearCurrentData();
     }
   }
-  // check if the node is last kernel
-  bool last_kernel = !AnfAlgo::IsInplaceNode(cnode, "skip");
-  debugger->PostExecuteNode(cnode, last_kernel);
+  if (IsDeviceTargetGPU()) {
+    // check if the node is last kernel
+    bool last_kernel = !AnfAlgo::IsInplaceNode(cnode, "skip");
+    debugger->PostExecuteNode(cnode, last_kernel);
+  }
 }
 
 /*
@@ -210,10 +218,7 @@ std::string CheckDatasetSinkMode(const KernelGraphPtr &graph_ptr) {
   std::string error_info = "";
   bool sink_mode = ConfigManager::GetInstance().dataset_mode() || graph_ptr->IsDatasetGraph();
   auto debugger = Debugger::GetInstance();
-  auto context = MsContext::GetInstance();
-  MS_EXCEPTION_IF_NULL(context);
-  bool is_gpu = (context->get_param<std::string>(MS_CTX_DEVICE_TARGET) == kGPUDevice);
-  if (debugger->CheckDebuggerDumpEnabled() && sink_mode && is_gpu) {
+  if (debugger->CheckDebuggerDumpEnabled() && sink_mode && IsDeviceTargetGPU()) {
     error_info = "e2e_dump is not supported on GPU with dataset_sink_mode=True. Please set dataset_sink_mode=False";
   }
   if (debugger->CheckDebuggerEnabled() && sink_mode) {
@@ -250,17 +255,9 @@ void LoadDataForDebugger(const KernelGraphPtr &graph_ptr) {
 #endif
 }
 
-void DumpSetup(const KernelGraphPtr &graph) {
-  MS_LOG(DEBUG) << "Start!";
-  MS_EXCEPTION_IF_NULL(graph);
-  E2eDump::DumpSetup(graph.get());
-  MS_LOG(DEBUG) << "Finish!";
-}
-
 void Dump(const KernelGraphPtr &graph, uint32_t rank_id) {
   MS_LOG(DEBUG) << "Start!";
   MS_EXCEPTION_IF_NULL(graph);
-  E2eDump::DumpRunIter(graph, rank_id);
   E2eDump::DumpData(graph.get(), rank_id);
   MS_LOG(DEBUG) << "Finish!";
 }
@@ -280,7 +277,6 @@ uint32_t GetRankID() {
 void SuperKernelE2eDump(const KernelGraphPtr &graph) {
 #ifndef ENABLE_SECURITY
   Dump(graph, GetRankID());
-  DumpSetup(graph);
 #endif
 }
 
diff --git a/mindspore/ccsrc/runtime/framework/actor/data_prepare_actor.cc b/mindspore/ccsrc/runtime/framework/actor/data_prepare_actor.cc
index 6d7b9677f7a..1d469019ce9 100644
--- a/mindspore/ccsrc/runtime/framework/actor/data_prepare_actor.cc
+++ b/mindspore/ccsrc/runtime/framework/actor/data_prepare_actor.cc
@@ -294,7 +294,8 @@ void DataPrepareActor::PrepareData(const std::vector<std::vector<TensorPtr>> &in
 
 void DataPrepareActor::SendDebugReq(OpContext<DeviceTensor> *const context) {
   ActorDispatcher::Send(*debug_aid_, &DebugActor::DebugOnStepBegin, graph_compiler_info_->graphs_,
-                        graph_compiler_info_->device_contexts_, context, &GetAID());
+                        graph_compiler_info_->origin_parameters_order_, graph_compiler_info_->device_contexts_, context,
+                        &GetAID());
 }
 
 void DataPrepareActor::OnDebugFinish(OpContext<DeviceTensor> *const context) {
diff --git a/mindspore/ccsrc/runtime/framework/actor/debug_actor.cc b/mindspore/ccsrc/runtime/framework/actor/debug_actor.cc
index ce5bc0424d6..7998b2b461a 100644
--- a/mindspore/ccsrc/runtime/framework/actor/debug_actor.cc
+++ b/mindspore/ccsrc/runtime/framework/actor/debug_actor.cc
@@ -128,7 +128,9 @@ void DebugActor::DebugForGraph(const KernelGraphPtr &graph, const DeviceContext
  * Runtime category: MindRT.
  * Description: Checks dataset_sink_mode and generates the related error if any exist and calls PreExecuteGraphDebugger.
  */
-void DebugActor::DebugOnStepBegin(std::vector<KernelGraphPtr> graphs, std::vector<DeviceContext *> device_contexts,
+void DebugActor::DebugOnStepBegin(const std::vector<KernelGraphPtr> &graphs,
+                                  const std::vector<AnfNodePtr> &origin_parameters_order,
+                                  std::vector<DeviceContext *> device_contexts,
                                   OpContext<DeviceTensor> *const op_context, const AID *from_aid) {
   MS_EXCEPTION_IF_NULL(op_context);
   MS_EXCEPTION_IF_NULL(from_aid);
@@ -144,7 +146,7 @@ void DebugActor::DebugOnStepBegin(std::vector<KernelGraphPtr> graphs, std::vecto
   }
   auto debugger = Debugger::GetInstance();
   if (debugger != nullptr && debugger->DebuggerBackendEnabled()) {
-    debugger->PreExecuteGraphDebugger(graphs);
+    debugger->PreExecuteGraphDebugger(graphs, origin_parameters_order);
   }
 #endif
 
diff --git a/mindspore/ccsrc/runtime/framework/actor/debug_actor.h b/mindspore/ccsrc/runtime/framework/actor/debug_actor.h
index 824455c0314..5477d187967 100644
--- a/mindspore/ccsrc/runtime/framework/actor/debug_actor.h
+++ b/mindspore/ccsrc/runtime/framework/actor/debug_actor.h
@@ -43,8 +43,10 @@ class DebugActor : public ActorBase {
                      OpContext<DeviceTensor> *const op_context, const AID *from_aid);
 
   // The debug on step begin.
-  void DebugOnStepBegin(std::vector<KernelGraphPtr> graphs, std::vector<DeviceContext *> device_contexts,
-                        OpContext<DeviceTensor> *const op_context, const AID *from_aid);
+  void DebugOnStepBegin(const std::vector<KernelGraphPtr> &graphs,
+                        const std::vector<AnfNodePtr> &origin_parameters_order,
+                        std::vector<DeviceContext *> device_contexts, OpContext<DeviceTensor> *const op_context,
+                        const AID *from_aid);
 
   // The debug on step end.
   void DebugOnStepEnd(OpContext<DeviceTensor> *const op_context, const AID *from_aid);
diff --git a/mindspore/ccsrc/runtime/hardware/ascend/ascend_device_context.cc b/mindspore/ccsrc/runtime/hardware/ascend/ascend_device_context.cc
index 97de4555e8d..74308bb9c7f 100644
--- a/mindspore/ccsrc/runtime/hardware/ascend/ascend_device_context.cc
+++ b/mindspore/ccsrc/runtime/hardware/ascend/ascend_device_context.cc
@@ -278,6 +278,16 @@ bool AscendDeviceContext::IsGraphMode() {
 }
 
 void AscendDeviceContext::Destroy() {
+#ifdef ENABLE_DEBUGGER
+  auto debugger = Debugger::GetInstance();
+  if (debugger && debugger->debugger_enabled()) {
+    debugger->SetTrainingDone(true);
+    bool ret = debugger->SendMetadata(false);
+    if (!ret) {
+      MS_LOG(ERROR) << "Failed to SendMetadata when finalize";
+    }
+  }
+#endif
   MS_LOG(INFO) << "Status record: Enter Destroy...";
   if (!initialized_) {
     return;
diff --git a/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc b/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc
index d87376af1cb..2a11831daeb 100644
--- a/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc
+++ b/mindspore/ccsrc/runtime/hardware/gpu/gpu_device_context.cc
@@ -153,7 +153,10 @@ void GPUDeviceContext::Destroy() {
   auto debugger = Debugger::GetInstance();
   if (debugger && debugger->debugger_enabled()) {
     debugger->SetTrainingDone(true);
-    debugger->SendMetadata(false);
+    bool ret = debugger->SendMetadata(false);
+    if (!ret) {
+      MS_LOG(ERROR) << "Failed to SendMetadata when finalize";
+    }
   }
 #endif
 
diff --git a/tests/st/dump/test_data_dump.py b/tests/st/dump/test_data_dump.py
index 354727f69e8..73a2b1546ea 100644
--- a/tests/st/dump/test_data_dump.py
+++ b/tests/st/dump/test_data_dump.py
@@ -101,7 +101,7 @@ def run_e2e_dump():
         add = Net()
         add(Tensor(x), Tensor(y))
         if context.get_context("device_target") == "Ascend":
-            assert len(os.listdir(dump_file_path)) == 5
+            assert len(os.listdir(dump_file_path)) == 3
             output_name = "Add.Add-op*.0.0.*.output.0.DefaultFormat.npy"
         elif context.get_context("device_target") == "CPU":
             assert len(os.listdir(dump_file_path)) == 5
@@ -271,7 +271,7 @@ def test_dump_with_diagnostic_path():
             shutil.rmtree(diagnose_path)
         add = Net()
         add(Tensor(x), Tensor(y))
-        assert len(os.listdir(dump_file_path)) == 5
+        assert len(os.listdir(dump_file_path)) == 3
         del os.environ['MINDSPORE_DUMP_CONFIG']
         del os.environ['MS_DIAGNOSTIC_DATA_PATH']