diff --git a/lite/backends/bm/target_wrapper.cc b/lite/backends/bm/target_wrapper.cc
index 6dab2a574d9c270573c00688768ad45a767abeae..83aa4dc8c1a6462bfd38a1c59f438e4836a3da00 100644
--- a/lite/backends/bm/target_wrapper.cc
+++ b/lite/backends/bm/target_wrapper.cc
@@ -23,7 +23,7 @@ int TargetWrapperBM::device_id_ = 0;
 std::map<int, void*> TargetWrapperBM::bm_hds_;
 
 size_t TargetWrapperBM::num_devices() {
-  int count = 0;
+  int count = 1;
   bm_status_t ret = bm_dev_getcount(&count);
   CHECK_EQ(ret, BM_SUCCESS) << "Failed with error code: "
                             << static_cast<int>(ret);
diff --git a/lite/kernels/bm/subgraph_compute.cc b/lite/kernels/bm/subgraph_compute.cc
index efbb848313a008484187202c876c90ddab80e5b3..eeb81ba9daa020083d519d5806af65ff74ad8a86 100644
--- a/lite/kernels/bm/subgraph_compute.cc
+++ b/lite/kernels/bm/subgraph_compute.cc
@@ -66,9 +66,9 @@ bool SubgraphEngine::BuildDeviceProgram() {
       graph.GetCompilerHandle(), const_cast<char*>(unique_net_name.c_str()), 1);
   void* bmodel_data = nullptr;
   unsigned int data_size = 0;
-  bm_hd_ = static_cast<bm_handle_t>(ctx.GetHandle());
   finish_bmcompiler_data(graph.GetCompilerHandle(), &bmodel_data, &data_size);
   graph.UnlockCompilerMutex();
+  bm_hd_ = static_cast<bm_handle_t>(ctx.GetHandle());
   bmrt_hd_ = bmrt_create(bm_hd_);
   if (false == bmrt_load_bmodel_data(bmrt_hd_, bmodel_data, data_size)) {
     return false;
@@ -79,15 +79,15 @@ bool SubgraphEngine::BuildDeviceProgram() {
   // input
   device_inputs_.resize(input_names_.size());
   for (size_t i = 0; i < input_names_.size(); i++) {
-    origin_itensors_[i] =
+    auto origin_itensor =
         exec_scope_->FindMutableTensor(net_info_->input_names[i]);
-    CHECK(origin_itensors_[i]);
+    CHECK(origin_itensor);
     bm_device_mem_t* p_mem =
         static_cast<bm_device_mem_t*>(malloc(sizeof(bm_device_mem_t)));
     CHECK(p_mem != nullptr);
-    CHECK_EQ(bm_malloc_device_byte(
-                 bm_hd_, p_mem, origin_itensors_[i]->memory_size()),
-             BM_SUCCESS);
+    CHECK_EQ(
+        bm_malloc_device_byte(bm_hd_, p_mem, origin_itensor->memory_size()),
+        BM_SUCCESS);
     bmrt_tensor_with_device(&device_inputs_[i],
                             *p_mem,
                             net_info_->input_dtypes[i],
@@ -124,9 +124,11 @@ bool SubgraphEngine::BuildDeviceProgram() {
 
 bool SubgraphEngine::LaunchDeviceProgram() {
   for (size_t i = 0; i < device_inputs_.size(); i++) {
+    auto origin_itensor =
+        exec_scope_->FindMutableTensor(net_info_->input_names[i]);
     bm_memcpy_s2d(bm_hd_,
                   device_inputs_[i].device_mem,
-                  const_cast<void*>(origin_itensors_[i]->raw_data()));
+                  const_cast<void*>(origin_itensor->raw_data()));
   }
   bmrt_launch_tensor_ex(bmrt_hd_,
                         net_names_[0],