diff --git a/ge/graph/load/model_manager/davinci_model.cc b/ge/graph/load/model_manager/davinci_model.cc
index 231ef755..9954ceaf 100755
--- a/ge/graph/load/model_manager/davinci_model.cc
+++ b/ge/graph/load/model_manager/davinci_model.cc
@@ -147,7 +147,6 @@ DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener
       runtime_model_id_(0),
       version_(0),
       ge_model_(nullptr),
-      thread_id_(),
       listener_(listener),
       run_flg_(false),
       priority_(priority),
@@ -168,6 +167,8 @@ DavinciModel::DavinciModel(int32_t priority, const std::shared_ptr<ModelListener
 
 DavinciModel::~DavinciModel() {
   try {
+    GE_CHK_STATUS(ModelRunStop());
+
     Status ret = data_dumper_.UnloadDumpInfo();
     if (ret != SUCCESS) {
       GELOGW("UnloadDumpInfo failed, ret: %u.", ret);
@@ -175,8 +176,6 @@ DavinciModel::~DavinciModel() {
 
     ClearTaskAddrs();
 
-    GE_CHK_STATUS(ModelRunStop());
-
     op_list_.clear();
     tensor_name_to_fixed_addr_size_.clear();
     tensor_name_to_peer_output_index_.clear();
@@ -730,9 +729,9 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
 
   SetProfileTime(MODEL_LOAD_END);
   // collect profiling for ge
-  GE_CHK_STATUS_RET(InitModelProfile(), "Init model profile failed");
   auto &profiling_manager = ProfilingManager::Instance();
   if (profiling_manager.ProfilingModelLoadOn()) {
+    GE_CHK_STATUS_RET(InitModelProfile(), "Init model profile failed");
     Status p_ret = ReportProfilingData();
     if (p_ret != SUCCESS) {
       GELOGE(p_ret, "Report profiling data failed.");
@@ -740,7 +739,7 @@ Status DavinciModel::Init(void *dev_ptr, size_t mem_size, void *weight_ptr, size
     }
   }
 
-  Shrink();
+  CREATE_STD_THREAD(shrink_id_, &DavinciModel::Shrink, this);
   return SUCCESS;
 }
 
@@ -2764,16 +2763,20 @@ void *DavinciModel::Run(DavinciModel *model) {
 /// @author
 ///
 Status DavinciModel::DestroyThread() {
-  GE_CHK_BOOL_RET_STATUS(data_inputer_ != nullptr, INTERNAL_ERROR, "data_inputer_ is nullptr.");
-
   run_flg_ = false;
 
-  data_inputer_->Stop();
+  if (data_inputer_ != nullptr) {
+    data_inputer_->Stop();
+  }
 
   if (thread_id_.joinable()) {
     thread_id_.join();
   }
 
+  if (shrink_id_.joinable()) {
+    shrink_id_.join();
+  }
+
   return SUCCESS;
 }
 
@@ -2820,8 +2823,6 @@ Status DavinciModel::ModelRunStop() {
   LockRunFlg();
   GE_MAKE_GUARD(tmp_lock, [&] { UnlockRunFlg(); });
 
-  GE_IF_BOOL_EXEC(!run_flg_, return SUCCESS);
-
   GE_CHK_STATUS_RET(DestroyThread(), "DestoyThead failed.");
 
   return SUCCESS;
diff --git a/ge/graph/load/model_manager/davinci_model.h b/ge/graph/load/model_manager/davinci_model.h
index 8ed82912..31cc9939 100755
--- a/ge/graph/load/model_manager/davinci_model.h
+++ b/ge/graph/load/model_manager/davinci_model.h
@@ -907,6 +907,7 @@ class DavinciModel {
   vector<int64_t> output_memory_size_list_;
 
   thread thread_id_;
+  thread shrink_id_;
 
   shared_ptr<ModelListener> listener_;
 
diff --git a/tests/ut/ge/graph/load/davinci_model_unittest.cc b/tests/ut/ge/graph/load/davinci_model_unittest.cc
index 5f08f425..c90a4997 100644
--- a/tests/ut/ge/graph/load/davinci_model_unittest.cc
+++ b/tests/ut/ge/graph/load/davinci_model_unittest.cc
@@ -284,7 +284,9 @@ TEST_F(UtestDavinciModel, init_unknown) {
   memcpy_async->set_op_index(2);
 
   EXPECT_EQ(model.Assign(ge_model), SUCCESS);
+  ProfilingManager::Instance().is_load_profiling_ = true;
   EXPECT_EQ(model.Init(), SUCCESS);
+  ProfilingManager::Instance().is_load_profiling_ = false;
 
   EXPECT_EQ(model.input_addrs_list_.size(), 1);
   EXPECT_EQ(model.output_addrs_list_.size(), 1);