diff --git a/ge/graph/load/model_manager/davinci_model.cc b/ge/graph/load/model_manager/davinci_model.cc
index 52642086..08d60b59 100755
--- a/ge/graph/load/model_manager/davinci_model.cc
+++ b/ge/graph/load/model_manager/davinci_model.cc
@@ -111,6 +111,9 @@ const char *const kWorkSpaceSize = "workspace_size";
 const char *const kTotalSize = "total_size";
 const char *const kTaskCount = "task_count";
 const char *const kTaskId = "task_id";
+const char *const kIndexId = "index_id";
+const char *const kTimeStamp = "time_stamp";
+const char *const kTagId = "tag_id";
 const char* const kRequestId = "request_id";
 const char* const kThreadId = "thread_id";
 const char* const kInputBeginTime = "input_begin_time";
@@ -2305,6 +2308,61 @@ Status DavinciModel::SinkTimeProfile(const InputData &current_data) {
   return SUCCESS;
 }
 
+Status DavinciModel::ProfileStepInfo(DavinciModel *model, uint16_t tag_id) {
+  if (ProfilingManager::Instance().ProfilingModelExecuteOn()) {
+    uint64_t index_id = model->iterator_count_ + 1;
+    uint64_t model_id = static_cast<uint64_t>(model->Id());
+    rtError_t rt_ret = RT_ERROR_NONE;
+#ifndef ONLY_COMPILE_OPEN_SRC
+    rtStream_t stream = model->rt_model_stream_;
+    GELOGD("Profiling Step Info TraceTask execute async start, index_id = %lu, model_id = %lu, tag_id = %u",
+           index_id, model_id, tag_id);
+    rt_ret = rtProfilerTraceEx(index_id, model_id, tag_id, stream);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "[Call][rtProfilerTraceEx] failed, ret: 0x%X", rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+    GELOGD("Profiling Step Info TraceTask execute async success, index_id = %lu, model_id = %lu, tag_id = %u",
+           index_id, model_id, tag_id);
+#endif
+
+    mmTimespec timespec = mmGetTickCount();
+    // 1000 ^ 3 converts second to nanosecond
+    int64_t time = timespec.tv_sec * 1000 * 1000 * 1000 + timespec.tv_nsec;
+    uint32_t task_id = 0;
+    uint32_t stream_id = 0;
+    rt_ret = rtGetTaskIdAndStreamID(&task_id, &stream_id);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "[Get][RtsInfo] task_id and stream_id failed, ret: 0x%X.", rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+    GELOGD("Get profiling args, task_id[%u], stream_id[%u]", task_id, stream_id);
+    int32_t device_id = static_cast<int32_t>(model->GetDeviceId());
+
+    Json step_info;
+    step_info[kIndexId] = index_id;
+    step_info[kModeleId] = model_id;
+    step_info[kTimeStamp] = time;
+    step_info[kTagId] = tag_id;
+    step_info[kTaskId] = task_id;
+    step_info[kStreamId] = stream_id;
+    step_info[kThreadId] = mmGetTid();
+
+    std::string reported_data;
+    try {
+      reported_data = step_info.dump(kInteval, ' ', false, Json::error_handler_t::ignore);
+    } catch (std::exception &e) {
+      GELOGE(FAILED, "Failed to convert JSON to string, reason: %s.", e.what());
+    } catch (...) {
+      GELOGE(FAILED, "Failed to convert JSON to string.");
+    }
+    reported_data.append(",")
+                 .append("\n");
+    ProfilingManager::Instance().ReportData(device_id, reported_data, "step_info");
+  }
+  return SUCCESS;
+}
+
 void DavinciModel::SetProfileTime(ModelProcStage stage, int64_t endTime) {
   int64_t time = endTime;
 
@@ -2599,6 +2657,8 @@ void *DavinciModel::Run(DavinciModel *model) {
       GELOGI("data_wrapper is null!");
       continue;
     }
+    // tag_id 0 means step begin, 1 meas step end.
+    (void)DavinciModel::ProfileStepInfo(model, 0);
     GELOGI("Getting the input data, model_id:%u", model_id);
     GE_IF_BOOL_EXEC(!model->RunFlag(), break);
 
@@ -2678,6 +2738,8 @@ void *DavinciModel::Run(DavinciModel *model) {
     GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(),
                     model->SetProfileTime(MODEL_AFTER_PROC_END));
     GE_IF_BOOL_EXEC(ProfilingManager::Instance().ProfilingModelExecuteOn(), (void)model->SinkTimeProfile(current_data));
+    // tag_id 0 means step begin, 1 meas step end.
+    (void)DavinciModel::ProfileStepInfo(model, 1);
 
     model->iterator_count_++;
     model->is_first_execute_ = false;
@@ -3693,12 +3755,16 @@ Status DavinciModel::NnExecute(rtStream_t stream, bool async_mode, const InputDa
   GE_IF_BOOL_EXEC(profiling_model_execute_on, SetProfileTime(MODEL_PRE_PROC_END));
 
   if (!task_list_.empty()) {
+    // tag_id 0 means step begin, 1 meas step end.
+    GE_CHK_STATUS_RET_NOLOG(ProfileStepInfo(this, 0));
     GELOGD("rtModelExecute do");
     GE_IF_BOOL_EXEC(profiling_model_execute_on, SetProfileTime(MODEL_INFER_START));
     rtError_t rt_ret = rtModelExecute(rt_model_handle_, rt_model_stream_, 0);
     GE_CHK_RT_EXEC(rt_ret, return RT_ERROR_TO_GE_STATUS(rt_ret));
     GE_IF_BOOL_EXEC(profiling_model_execute_on, SetProfileTime(MODEL_INFER_END));
     GELOGD("rtModelExecute end");
+    GE_CHK_STATUS_RET_NOLOG(ProfileStepInfo(this, 1));
+    iterator_count_++;
   }
 
   if (!is_async_mode_) {
diff --git a/ge/graph/load/model_manager/davinci_model.h b/ge/graph/load/model_manager/davinci_model.h
index 58478b0f..1adeb1cc 100755
--- a/ge/graph/load/model_manager/davinci_model.h
+++ b/ge/graph/load/model_manager/davinci_model.h
@@ -878,6 +878,8 @@ class DavinciModel {
   Status GetRealOutputSizeOfCase(const ComputeGraphPtr &graph, size_t input_index, const NodePtr &case_node);
   Status GetGearAndRealOutShapeInfo(const ComputeGraphPtr &graph, const NodePtr &node);
 
+  static Status ProfileStepInfo(DavinciModel *model, uint16_t tag_id);
+
   bool is_weight_mem_has_inited_;
   bool is_feature_map_mem_has_inited_;
 
diff --git a/ge/hybrid/executor/hybrid_model_executor.cc b/ge/hybrid/executor/hybrid_model_executor.cc
index 4b589a03..3e65b5c1 100755
--- a/ge/hybrid/executor/hybrid_model_executor.cc
+++ b/ge/hybrid/executor/hybrid_model_executor.cc
@@ -18,12 +18,22 @@
 #include "graph/ge_context.h"
 #include "graph/runtime_inference_context.h"
 #include "common/dump/dump_manager.h"
+#include "common/profiling/profiling_manager.h"
+#include "mmpa/mmpa_api.h"
 
 namespace ge {
 namespace hybrid {
 namespace {
 const int kIntBase = 10;
 const char *const kEnvProfilingLevel = "HYBRID_PROFILING_LEVEL";
+const char *const kIndexId = "index_id";
+const char *const kModeleId = "model_id";
+const char *const kTimeStamp = "time_stamp";
+const char *const kStreamId = "stream_id";
+const char *const kTaskId = "task_id";
+const char *const kTagId = "tag_id";
+const char *const kThreadId = "thread_id";
+const uint32_t kInteval = 2;
 } // namespace
 HybridModelExecutor::HybridModelExecutor(HybridModel *model, uint32_t device_id, rtStream_t stream)
     : model_(model), device_id_(device_id), stream_(stream) {
@@ -71,15 +81,72 @@ Status HybridModelExecutor::Execute(HybridModelExecutor::ExecuteArgs &args) {
   return SUCCESS;
 }
 
+Status HybridModelExecutor::ProfileStepInfo(uint16_t tag_id) {
+  auto &prof_mgr = ProfilingManager::Instance();
+  if (prof_mgr.ProfilingModelExecuteOn()) {
+    uint64_t index_id = context_.iteration + 1;
+    uint64_t model_id = static_cast<uint64_t>(model_->GetModelId());
+    rtError_t rt_ret = RT_ERROR_NONE;
+#ifndef ONLY_COMPILE_OPEN_SRC
+    GELOGD("Profiling Step Info TraceTask execute async start, index_id = %lu, model_id = %lu, tag_id = %u",
+           index_id, model_id, tag_id);
+    rt_ret = rtProfilerTraceEx(index_id,  model_id, tag_id, stream_);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "[Call][rtProfilerTraceEx] failed, ret: 0x%X", rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+    GELOGD("Profiling Step Info TraceTask execute async success, index_id = %lu, model_id = %lu, tag_id = %u",
+           index_id, model_id, tag_id);
+#endif
+
+    mmTimespec timespec = mmGetTickCount();
+    // 1000 ^ 3 converts second to nanosecond
+    int64_t time = timespec.tv_sec * 1000 * 1000 * 1000 + timespec.tv_nsec;
+    uint32_t task_id = 0;
+    uint32_t stream_id = 0;
+    rt_ret = rtGetTaskIdAndStreamID(&task_id, &stream_id);
+    if (rt_ret != RT_ERROR_NONE) {
+      GELOGE(RT_FAILED, "[Get][RtsInfo] task_id and stream_id failed, ret: 0x%X.", rt_ret);
+      return RT_ERROR_TO_GE_STATUS(rt_ret);
+    }
+    GELOGD("Get profiling args, task_id[%u], stream_id[%u]", task_id, stream_id);
+
+    Json step_info;
+    step_info[kIndexId] = index_id;
+    step_info[kModeleId] = model_id;
+    step_info[kTimeStamp] = time;
+    step_info[kTagId] = tag_id;
+    step_info[kTaskId] = task_id;
+    step_info[kStreamId] = stream_id;
+    step_info[kThreadId] = mmGetTid();
+
+    std::string reported_data;
+    try {
+      reported_data = step_info.dump(kInteval, ' ', false, Json::error_handler_t::ignore);
+    } catch (std::exception &e) {
+      GELOGE(FAILED, "Failed to convert JSON to string, reason: %s.", e.what());
+    } catch (...) {
+      GELOGE(FAILED, "Failed to convert JSON to string.");
+    }
+    reported_data.append(",")
+                 .append("\n");
+    prof_mgr.ReportData(device_id_, reported_data, "step_info");
+  }
+  return SUCCESS;
+}
+
 Status HybridModelExecutor::ExecuteGraphInternal(SubgraphExecutor &executor,
                                                  HybridModelExecutor::ExecuteArgs &args) {
   RECORD_MODEL_EXECUTION_EVENT(&context_, "[InitContext] Start");
   GE_CHK_STATUS_RET_NOLOG(ResetExecutionContext(context_));
   RECORD_MODEL_EXECUTION_EVENT(&context_, "[InitContext] End");
 
+  // tag_id 0 means step begin, 1 meas step end.
+  GE_CHK_STATUS_RET_NOLOG(ProfileStepInfo(0));
   HYBRID_CHK_STATUS_RET(executor.ExecuteAsync(args.inputs, args.input_desc, args.outputs),
                         "Failed to execute partitioned call.");
   RECORD_MODEL_EXECUTION_EVENT(&context_, "[ExecuteAsync] End");
+  GE_CHK_STATUS_RET_NOLOG(ProfileStepInfo(1));
 
   HYBRID_CHK_STATUS_RET(executor.Synchronize(), "Failed to sync root graph.");
   RECORD_MODEL_EXECUTION_EVENT(&context_, "[Synchronize] End");
diff --git a/ge/hybrid/executor/hybrid_model_executor.h b/ge/hybrid/executor/hybrid_model_executor.h
index 0b2cd1ed..af0bd536 100644
--- a/ge/hybrid/executor/hybrid_model_executor.h
+++ b/ge/hybrid/executor/hybrid_model_executor.h
@@ -51,6 +51,8 @@ class HybridModelExecutor {
   Status ExecuteGraphInternal(SubgraphExecutor &executor, ExecuteArgs &args);
   Status Cleanup();
   Status InitExecutionContext();
+  Status ProfileStepInfo(uint16_t tag_id);
+
   static Status ResetExecutionContext(GraphExecutionContext &context);
 
   HybridModel *model_;
diff --git a/tests/depends/runtime/src/runtime_stub.cc b/tests/depends/runtime/src/runtime_stub.cc
index f8eedcbc..b062ec80 100644
--- a/tests/depends/runtime/src/runtime_stub.cc
+++ b/tests/depends/runtime/src/runtime_stub.cc
@@ -313,6 +313,8 @@ rtError_t rtFlushCache(uint64_t base, uint32_t len) { return RT_ERROR_NONE; }
 
 rtError_t rtProfilerTrace(uint64_t id, bool notify, uint32_t flags, rtStream_t stream_) { return RT_ERROR_NONE; }
 
+rtError_t rtProfilerTraceEx(uint64_t id, uint64_t modelId, uint16_t tagId, rtStream_t stream) { return RT_ERROR_NONE; }
+
 rtError_t rtMemSetRC(const void *dev_ptr, uint64_t size, uint32_t read_count) { return RT_ERROR_NONE; }
 
 rtError_t rtStreamSwitch(void *ptr, rtCondition_t condition, int64_t value, rtStream_t true_stream, rtStream_t stream) {
diff --git a/tests/ut/ge/CMakeLists.txt b/tests/ut/ge/CMakeLists.txt
index 3e6f10e2..4cfeddf2 100755
--- a/tests/ut/ge/CMakeLists.txt
+++ b/tests/ut/ge/CMakeLists.txt
@@ -850,6 +850,7 @@ target_link_libraries(ge_ut_common PRIVATE
     ascend_protobuf
     json
     ge_ut_graph
+    runtime_stub
 )
 
 # build common format
diff --git a/tests/ut/ge/graph/load/davinci_model_unittest.cc b/tests/ut/ge/graph/load/davinci_model_unittest.cc
index 3487f8ed..ae6afcc5 100644
--- a/tests/ut/ge/graph/load/davinci_model_unittest.cc
+++ b/tests/ut/ge/graph/load/davinci_model_unittest.cc
@@ -898,6 +898,13 @@ TEST_F(UtestDavinciModel, Sink_time_profile) {
   model.SinkTimeProfile(current_data);
 }
 
+TEST_F(UtestDavinciModel, Step_info_profile) {
+  ProfilingManager::Instance().prof_cb_.msprofReporterCallback = MsprofReport;
+  ProfilingManager::Instance().device_id_.emplace_back(0);
+  DavinciModel model(0, nullptr);
+  EXPECT_EQ(DavinciModel::ProfileStepInfo(&model, 0), SUCCESS);
+}
+
 class ClassTest {
 public:
     virtual ~ClassTest() {}
diff --git a/tests/ut/ge/hybrid/ge_hybrid_unittest.cc b/tests/ut/ge/hybrid/ge_hybrid_unittest.cc
index 3b5d19e6..e377ca66 100644
--- a/tests/ut/ge/hybrid/ge_hybrid_unittest.cc
+++ b/tests/ut/ge/hybrid/ge_hybrid_unittest.cc
@@ -37,6 +37,7 @@
 #include "hybrid/common/npu_memory_allocator.h"
 #include "graph/types.h"
 #include "graph/utils/tensor_utils.h"
+#include "common/profiling/profiling_manager.h"
 
 #undef private
 #undef protected
@@ -46,6 +47,11 @@ using namespace testing;
 using namespace ge;
 using namespace hybrid;
 
+namespace {
+int32_t MsprofReport(uint32_t moduleId, uint32_t type, void *data, uint32_t len) {
+  return 0;
+}
+}  // namespace
 
 class UtestGeHybrid : public testing::Test {
  protected:
@@ -245,7 +251,7 @@ TEST_F(UtestGeHybrid, init_weight_success) {
   ASSERT_EQ(ret,PARAM_INVALID);
 }
 
-  TEST_F(UtestGeHybrid, hybrid_model_executor) {
+TEST_F(UtestGeHybrid, hybrid_model_executor) {
   ComputeGraphPtr compute_graph = MakeShared<ComputeGraph>("abc");
   GeRootModelPtr root_model = MakeShared<ge::GeRootModel>(compute_graph);
   HybridModel model(root_model);
@@ -256,3 +262,13 @@ TEST_F(UtestGeHybrid, init_weight_success) {
   HybridModelExecutor executor(model_ptr, device_id, stream);
   executor.Init();
 }
+
+TEST_F(UtestGeHybrid, Step_info_profile) {
+  ProfilingManager::Instance().prof_cb_.msprofReporterCallback = MsprofReport;
+  ProfilingManager::Instance().device_id_.emplace_back(0);
+  auto graph = make_shared<ComputeGraph>("graph");
+  auto root_model = make_shared<GeRootModel>(graph);
+  HybridModel model(root_model);
+  HybridModelExecutor executor(&model, 0, nullptr);
+  EXPECT_EQ(executor.ProfileStepInfo(0), SUCCESS);
+}