diff --git a/mindspore/lite/src/inner_kernel.h b/mindspore/lite/src/inner_kernel.h
index d91218f584b..d7be21a01b3 100644
--- a/mindspore/lite/src/inner_kernel.h
+++ b/mindspore/lite/src/inner_kernel.h
@@ -40,13 +40,6 @@ class InnerKernel : public Kernel {
               const lite::Context *ctx)
       : op_parameter_(parameter), in_tensors_(std::move(in_tensors)), out_tensors_(std::move(out_tensors)) {
     context_ = ctx;
-    if (parameter != nullptr && parameter->thread_num_ == 0) {
-      if (ctx != nullptr) {
-        op_parameter_->thread_num_ = ctx->thread_num_;
-      } else {
-        op_parameter_->thread_num_ = 1;
-      }
-    }
   }
 
   virtual ~InnerKernel() {
diff --git a/mindspore/lite/src/lite_mindrt.cc b/mindspore/lite/src/lite_mindrt.cc
index 05db2fbc29d..9eaf08c410f 100644
--- a/mindspore/lite/src/lite_mindrt.cc
+++ b/mindspore/lite/src/lite_mindrt.cc
@@ -28,24 +28,12 @@ namespace mindspore::lite {
 void LiteOpActor::RunOpData(OpData<lite::Tensor> *inputs, OpContext<lite::Tensor> *context) {
   auto op_uuid = context->sequential_num_;
   input_op_datas_[op_uuid].push_back(inputs);
-
   inputs_data_[inputs->index_] = inputs->data_;
-  /* in-case infershape done in runtime */
-  kernel_->in_tensors()[inputs->index_]->set_shape(inputs->data_->shape());
-  kernel_->in_tensors()[inputs->index_]->set_format(inputs->data_->format());
-
   if (input_op_datas_[op_uuid].size() < kernel_->in_tensors().size()) {
     return;
   }
 
-  auto ret = CheckInputData();
-  if (ret != RET_OK) {
-    input_op_datas_.erase(op_uuid);
-    context->SetFailed(ret);
-    return;
-  }
-
-  ret = SetInputData();
+  auto ret = SetInputData();
   if (ret != RET_OK) {
     input_op_datas_.erase(op_uuid);
     context->SetFailed(ret);
@@ -87,7 +75,7 @@ void LiteOpActor::IsolateInputData(std::vector<std::shared_ptr<LiteOpActor>> *ac
     for (QuantArg quant : old_tensor->quant_params()) {
       new_tensor->AddQuantParam(quant);
     }
-    isolate_input_map_.insert(std::make_pair(old_tensor, new_tensor));
+    isolate_input_map_.insert(std::make_pair(new_tensor, old_tensor));
 
     int ref_count = 0;
     /* set op input for calculate */
@@ -126,8 +114,8 @@ int LiteOpActor::LiteActorInit(std::vector<std::shared_ptr<LiteOpActor>> *actors
 int LiteOpActor::ResizeGraphInput(const std::vector<mindspore::tensor::MSTensor *> &inputs,
                                   const std::vector<std::vector<int>> &dims) {
   for (auto map : isolate_input_map_) {
-    auto src_tensor = map.first;
-    auto isolate_tensor = map.second;
+    auto isolate_tensor = map.first;
+    auto src_tensor = map.second;
     for (size_t i = 0; i < inputs.size(); i++) {
       if (src_tensor == inputs[i]) {
         isolate_tensor->set_shape(dims[i]);
@@ -225,23 +213,6 @@ int LiteOpActor::CompileArrow() {
   return ret;
 }
 
-int LiteOpActor::CheckInputData() {
-  if (kernel_->in_tensors().size() != inputs_data_.size()) {
-    MS_LOG(ERROR) << "kernel:" << kernel_->name() << " inputs_data_.size(): " << inputs_data_.size()
-                  << " vs kernel_->in_tensors().size(): " << kernel_->in_tensors().size() << " are not equal.";
-    return RET_PARAM_INVALID;
-  }
-
-  for (size_t i = 0; i < inputs_data_.size(); ++i) {
-    if (kernel_->in_tensors()[i]->shape() != inputs_data_[i]->shape()) {
-      MS_LOG(ERROR) << "inputs_data_[" << i << "].shape: " << inputs_data_[i]->shape() << " vs kernel_->in_tensors()["
-                    << i << "].shape: " << kernel_->in_tensors()[i]->shape() << " are not equal.";
-      return RET_PARAM_INVALID;
-    }
-  }
-  return RET_OK;
-}
-
 void LiteOpActor::MoveInputData(Tensor *dst_tensor, Tensor *src_tensor) {
   MS_ASSERT(src_tensor != dst_tensor);
 
@@ -302,6 +273,11 @@ int LiteOpActor::SetInputData() {
   for (size_t i = 0; i < inputs_data_.size(); ++i) {
     auto dst_tensor = kernel_->in_tensors()[i];
     auto src_tensor = inputs_data_[i];
+
+    /* infershape done in runtime */
+    dst_tensor->set_shape(src_tensor->shape());
+    dst_tensor->set_format(src_tensor->format());
+
     if (src_tensor->data_type() != dst_tensor->data_type()) {
       CopyInputData(dst_tensor, src_tensor);
     } else {
@@ -567,10 +543,7 @@ void LiteSwitchOpActor::AsyncFalseBranchOutput(OpContext<Tensor> *context) {
   }
 }
 
-int MindrtInit(bool enable_parallel) {
-  int thread_count = enable_parallel ? 2 : 1;
-  return mindspore::Initialize("tcp://127.0.0.1:8080", "", "", "", thread_count);
-}
+int MindrtInit() { return mindspore::Initialize("tcp://127.0.0.1:8080", "", "", ""); }
 
 void MindrtTerminate(const std::vector<std::shared_ptr<LiteOpActor>> &actor_list) {
   for (const auto &actor : actor_list) {
diff --git a/mindspore/lite/src/lite_mindrt.h b/mindspore/lite/src/lite_mindrt.h
index 9de49171450..e3d70630b11 100644
--- a/mindspore/lite/src/lite_mindrt.h
+++ b/mindspore/lite/src/lite_mindrt.h
@@ -42,7 +42,7 @@ class LiteOpActor : public OpActor<lite::Tensor> {
   }
   ~LiteOpActor() override {
     for (auto map : isolate_input_map_) {
-      auto isolate_input_tensor = map.second;
+      auto isolate_input_tensor = map.first;
       isolate_input_tensor->set_data(nullptr);
       delete isolate_input_tensor;
     }
@@ -67,7 +67,6 @@ class LiteOpActor : public OpActor<lite::Tensor> {
   void SetPartialMap(const std::unordered_map<size_t, AID> &partial_map) { subgraph_index_to_actor = partial_map; }
 
  protected:
-  int CheckInputData();
   int SetInputData();
   void SetOutputData(OpContext<Tensor> *context);
   void AsyncOutput(OpContext<Tensor> *context);
@@ -89,7 +88,7 @@ class LiteOpActor : public OpActor<lite::Tensor> {
  private:
   kernel::LiteKernel *partial_node_ = nullptr;
   kernel::LiteKernel *call_node_ = nullptr;
-  std::unordered_map<Tensor *, Tensor *> isolate_input_map_;
+  std::unordered_map<Tensor *, Tensor *> isolate_input_map_; /* <calculate-tensor,  src-input-tensor> */
 };
 
 class LiteSwitchOpActor : public LiteOpActor {
@@ -104,14 +103,7 @@ class LiteSwitchOpActor : public LiteOpActor {
       return;
     }
 
-    int ret = CheckInputData();
-    if (ret != RET_OK) {
-      input_op_datas_.erase(op_uuid);
-      context->SetFailed(ret);
-      return;
-    }
-
-    ret = SetInputData();
+    auto ret = SetInputData();
     if (ret != RET_OK) {
       input_op_datas_.erase(op_uuid);
       context->SetFailed(ret);
@@ -182,7 +174,7 @@ class LiteSwitchOpActor : public LiteOpActor {
   std::vector<OpDataPtr<Tensor>> false_branch_outputs_data_;
 };
 
-int MindrtInit(bool subgraph_split = false);
+int MindrtInit();
 void MindrtTerminate(const std::vector<std::shared_ptr<LiteOpActor>> &);
 
 std::vector<std::shared_ptr<LiteOpActor>> CreateOpActor(const std::vector<kernel::LiteKernel *> &kernels,
diff --git a/mindspore/lite/src/mindrt_executor.cc b/mindspore/lite/src/mindrt_executor.cc
index 7efb2d7f75a..1636c42678b 100644
--- a/mindspore/lite/src/mindrt_executor.cc
+++ b/mindspore/lite/src/mindrt_executor.cc
@@ -77,9 +77,7 @@ int MindrtExecutor::Resize(const std::vector<mindspore::tensor::MSTensor *> &inp
 
 int MindrtExecutor::Prepare(const std::vector<kernel::LiteKernel *> &kernels, const std::vector<Tensor *> &inputs,
                             const std::vector<Tensor *> &outputs, const lite::InnerContext *ctx) {
-  MS_ASSERT(kernels.size() != 0);
-
-  auto ret = MindrtInit(ctx->enable_parallel_);
+  auto ret = MindrtInit();
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "MindrtInit failed";
     return ret;
diff --git a/mindspore/lite/src/runtime/agent/npu/optimizer/npu_pass_utils.cc b/mindspore/lite/src/runtime/agent/npu/optimizer/npu_pass_utils.cc
index cad7a82fa36..dfc7274f7c1 100644
--- a/mindspore/lite/src/runtime/agent/npu/optimizer/npu_pass_utils.cc
+++ b/mindspore/lite/src/runtime/agent/npu/optimizer/npu_pass_utils.cc
@@ -40,6 +40,7 @@ kernel::LiteKernel *NPUPassUtils::CreateNchw2NhwcKernel(const std::vector<Tensor
   }
   memset(transpose_param, 0, sizeof(TransposeParameter));
   transpose_param->op_parameter_.type_ = schema::PrimitiveType_Transpose;
+  transpose_param->op_parameter_.thread_num_ = ctx->thread_num_;
   transpose_param->perm_[0] = 0;
   transpose_param->perm_[1] = 2;
   transpose_param->perm_[2] = 3;
@@ -76,6 +77,7 @@ kernel::LiteKernel *NPUPassUtils::CreateNhwc2NchwKernel(const std::vector<Tensor
   }
   memset(transpose_param, 0, sizeof(TransposeParameter));
   transpose_param->op_parameter_.type_ = schema::PrimitiveType_Transpose;
+  transpose_param->op_parameter_.thread_num_ = ctx->thread_num_;
   transpose_param->perm_[0] = 0;
   transpose_param->perm_[1] = 3;
   transpose_param->perm_[2] = 1;
diff --git a/mindspore/lite/src/scheduler.cc b/mindspore/lite/src/scheduler.cc
index 9d9bc2ad8da..70ec73a1147 100644
--- a/mindspore/lite/src/scheduler.cc
+++ b/mindspore/lite/src/scheduler.cc
@@ -88,13 +88,7 @@ int Scheduler::Schedule(std::vector<kernel::LiteKernel *> *dst_kernels) {
   if (context_->enable_parallel_) {
     auto search_sub_graph =
       SearchSubGraph(context_, src_model_, src_tensors_, &op_parameters_, &graph_output_node_indexes_);
-
-    bool offline_parallel_enable = src_model_->all_nodes_.front()->device_type_ != kDefaultDeviceType;
-    if (offline_parallel_enable) {
-      search_sub_graph.SubGraphSplitByOffLineParallel();
-    } else {
-      search_sub_graph.SubGraphSplitByOutput();
-    }
+    search_sub_graph.SubGraphSplit();
   }
 
   ret = ScheduleSubGraphToKernels(kMainSubGraphIndex, dst_kernels, nullptr, nullptr);
@@ -111,14 +105,7 @@ int Scheduler::Schedule(std::vector<kernel::LiteKernel *> *dst_kernels) {
     }
   }
   FindAllInoutKernels(*dst_kernels);
-  // origin kernel init
-  for (size_t i = 0; i < dst_kernels->size(); i++) {
-    ret = (*dst_kernels)[i]->Init();
-    if (ret != RET_OK) {
-      MS_LOG(ERROR) << "Kernel " << (*dst_kernels)[i]->name() << " Init failed.";
-      return ret;
-    }
-  }
+
   ret = RunPass(dst_kernels);
   if (ret != RET_OK) {
     MS_LOG(ERROR) << "Schedule run pass failed.";
@@ -679,13 +666,6 @@ kernel::LiteKernel *Scheduler::SchedulePartialToKernel(const lite::Model::Node *
     MS_LOG(ERROR) << "Schedule partial failed, name: " << src_node->name_;
     return nullptr;
   }
-  for (auto kernel : sub_kernels) {
-    ret = kernel->Init();
-    if (ret != RET_OK) {
-      MS_LOG(ERROR) << "Schedule partial kernel init failed, name: " << kernel->name();
-      return nullptr;
-    }
-  }
 
   FindAllInoutKernels(sub_kernels);
   ret = RunPass(&sub_kernels);
@@ -725,6 +705,7 @@ int Scheduler::ScheduleSubGraphToKernels(size_t subgraph_index, std::vector<kern
   MS_ASSERT(dst_kernels->empty());
   auto subgraph = src_model_->sub_graphs_.at(subgraph_index);
   for (auto node_index : subgraph->node_indices_) {
+    auto ret = RET_OK;
     auto node = src_model_->all_nodes_[node_index];
     MS_ASSERT(node != nullptr);
     auto *primitive = node->primitive_;
@@ -735,8 +716,11 @@ int Scheduler::ScheduleSubGraphToKernels(size_t subgraph_index, std::vector<kern
       kernel = SchedulePartialToKernel(node);
     } else {  // kernel
       kernel = ScheduleNodeToKernel(node, prefer_data_type);
+      if (kernel != nullptr) {
+        ret = kernel->Init();
+      }
     }
-    if (kernel == nullptr) {
+    if (kernel == nullptr || ret != RET_OK) {
       MS_LOG(ERROR) << "FindBackendKernel return nullptr, name: " << node->name_
                     << ", type: " << PrimitiveTypeName(prim_type);
       return RET_ERROR;
diff --git a/mindspore/lite/src/sub_graph_split.cc b/mindspore/lite/src/sub_graph_split.cc
index 850ffbec3cd..f83e786b20d 100644
--- a/mindspore/lite/src/sub_graph_split.cc
+++ b/mindspore/lite/src/sub_graph_split.cc
@@ -559,4 +559,13 @@ void SearchSubGraph::SubGraphSplitByOffLineParallel() {
   MS_LOG(DEBUG) << "end to split offline parallel subgraph";
 }
 
+void SearchSubGraph::SubGraphSplit() {
+  bool offline_parallel_enable = model_->all_nodes_.front()->device_type_ != kDefaultDeviceType;
+  if (offline_parallel_enable) {
+    SubGraphSplitByOffLineParallel();
+  } else {
+    SubGraphSplitByOutput();
+  }
+  return;
+}
 }  // namespace mindspore::lite
diff --git a/mindspore/lite/src/sub_graph_split.h b/mindspore/lite/src/sub_graph_split.h
index 3fd45a8374e..3bda86ba4b3 100644
--- a/mindspore/lite/src/sub_graph_split.h
+++ b/mindspore/lite/src/sub_graph_split.h
@@ -73,14 +73,19 @@ class SearchSubGraph {
   ~SearchSubGraph() = default;
 
  public:
-  void SubGraphSplitByOutput();
-  void SubGraphSplitByMiddle();
-  void SubGraphSplitByOffLineParallel();
+  void SubGraphSplit();
 
  private:
+  void SubGraphSplitByOutput();
   void InitSearchSubGraphByOutput();
+
+ private:
+  void SubGraphSplitByMiddle();
   void InitSearchSubGraphByMiddle();
 
+ private:
+  void SubGraphSplitByOffLineParallel();
+
  private:
   void InitSearchTensor();
   void InitSearchParallelSubGraph();
diff --git a/mindspore/lite/test/config/models_mindrt_parallel.cfg b/mindspore/lite/test/config/models_mindrt_parallel.cfg
index 86ec61c6c6c..7f2c9e7b909 100644
--- a/mindspore/lite/test/config/models_mindrt_parallel.cfg
+++ b/mindspore/lite/test/config/models_mindrt_parallel.cfg
@@ -1,4 +1,5 @@
 # mindrt enable parallel model
 # model run both CPU-CPU & GPU-CPU
 # model_file  ###  accuracy_limit  ###  enable_fp16(true or false)
-mtk_model_normalize_object_scene_ps_20200519_f32.tflite 0.5 false
\ No newline at end of file
+mtk_model_normalize_object_scene_ps_20200519_f32.tflite;0.5;false
+# end
\ No newline at end of file
diff --git a/mindspore/lite/test/st/scripts/run_benchmark_gpu.sh b/mindspore/lite/test/st/scripts/run_benchmark_gpu.sh
index 5011498e2f7..7c57f941e09 100644
--- a/mindspore/lite/test/st/scripts/run_benchmark_gpu.sh
+++ b/mindspore/lite/test/st/scripts/run_benchmark_gpu.sh
@@ -277,17 +277,17 @@ function Run_mindrt_parallel() {
 
         data_path="/data/local/tmp/input_output/"
         output=${data_path}'output/'${model_name}'.ms.out'
-        input=${model_name}.ms.bin
+        input=${data_path}'input/'${model_name}'.ms.bin'
         model=${model_name}'.ms'
         echo ${model_name} >> "${run_parallel_log_file}"
         echo "run mindrt parallel test : ${model_name}"
 
-        ########## RUN CPU-GPU parallel
+        ########## RUN CPU-CPU parallel
         echo 'cd /data/local/tmp/benchmark_test' > adb_run_cmd.txt
-        echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/local/tmp/benchmark_test' > adb_run_cmd.txt
+        echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/local/tmp/benchmark_test' >> adb_run_cmd.txt
 
-        echo './benchmark --enableParallel=true --device=GPU --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> adb_run_cmd.txt
-        echo './benchmark --enableParallel=true --device=GPU --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> "${run_parallel_log_file}"
+        echo './benchmark --enableParallel=true --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> adb_run_cmd.txt
+        echo './benchmark --enableParallel=true --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> "${run_parallel_log_file}"
 
         adb -s ${device_id} shell < adb_run_cmd.txt >> "${run_parallel_log_file}"
         if [ $? = 0 ]; then
@@ -296,12 +296,12 @@ function Run_mindrt_parallel() {
             run_result='mindrt_parallel_CPU_GPU: '${model_name}' failed'; echo ${run_result} >> ${run_parallel_result_file}; return 1
         fi
 
-        ########## RUN CPU-CPU parallel
+        ########## RUN CPU-GPU parallel
         echo 'cd /data/local/tmp/benchmark_test' > adb_run_cmd.txt
-        echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/local/tmp/benchmark_test' > adb_run_cmd.txt
+        echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:/data/local/tmp/benchmark_test' >> adb_run_cmd.txt
 
-        echo './benchmark --enableParallel=true --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> adb_run_cmd.txt
-        echo './benchmark --enableParallel=true --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> "${run_parallel_log_file}"
+        echo './benchmark --enableParallel=true --device=GPU --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> adb_run_cmd.txt
+        echo './benchmark --enableParallel=true --device=GPU --enableFp16='${fp16}' --accuracyThreshold='${limit}' --modelFile='${model}' --inDataFile='${input}' --benchmarkDataFile='${output} >> "${run_parallel_log_file}"
 
         adb -s ${device_id} shell < adb_run_cmd.txt >> "${run_parallel_log_file}"
         if [ $? = 0 ]; then
@@ -486,7 +486,7 @@ if [[ $backend == "all" || $backend == "gpu" ]]; then
 
     if [[ ${Run_mindrt_parallel_status} != 0 ]];then
         echo "Run_mindrt_parallel failed"
-        cat ${run_gpu_log_file}
+        cat ${run_parallel_log_file}
     fi
 
     echo "Run_parallel is ended"