diff --git a/mindspore/lite/src/lite_session.cc b/mindspore/lite/src/lite_session.cc
index 82ec0d8147..446e259f23 100644
--- a/mindspore/lite/src/lite_session.cc
+++ b/mindspore/lite/src/lite_session.cc
@@ -380,6 +380,7 @@ int LiteSession::Init(const Context *context) {
     auto opencl_runtime = ocl_runtime_wrap_.GetInstance();
     opencl_runtime->SetFp16Enable(gpu_device_info.enable_float16_);
     if (opencl_runtime->Init() != RET_OK) {
+      this->context_->device_list_ = {{DT_CPU, {gpu_device_info.enable_float16_, MID_CPU}}};
       MS_LOG(WARNING) << "Init OpenCL runtime failed, change to CPU mode.";
     } else {
       MS_LOG(INFO) << "Init OpenCL runtime success.";
diff --git a/mindspore/lite/src/runtime/kernel/opencl/subgraph_opencl_kernel.cc b/mindspore/lite/src/runtime/kernel/opencl/subgraph_opencl_kernel.cc
index 0f23b982fa..c77acbc360 100644
--- a/mindspore/lite/src/runtime/kernel/opencl/subgraph_opencl_kernel.cc
+++ b/mindspore/lite/src/runtime/kernel/opencl/subgraph_opencl_kernel.cc
@@ -101,6 +101,10 @@ int SubGraphOpenCLKernel::GenToFormatOp(const std::vector<lite::Tensor *> &in_te
   out_convert_ops->clear();
   MS_ASSERT(in_tensors.size() == to_kernels.size());
   MS_ASSERT(in_tensors.size() == from_kernels.size());
+  std::vector<std::vector<kernel::LiteKernel *>> loop_kernels;
+  if (mem_type == OpenCLMemType::BUF) {
+    GetKernelFromToTensor(in_tensors, nodes_, &loop_kernels, true);
+  }
 
   ReplaceOutTensorAndKernelToNull(in_tensors, in_kernels, mem_type);
 
@@ -163,8 +167,6 @@ int SubGraphOpenCLKernel::GenToFormatOp(const std::vector<lite::Tensor *> &in_te
 
     // replace in_tensor of inner kernel which use out tensor
     if (mem_type == OpenCLMemType::BUF) {
-      std::vector<std::vector<kernel::LiteKernel *>> loop_kernels;
-      GetKernelFromToTensor(in_tensors, nodes_, &loop_kernels, true);
       for (auto &iv : loop_kernels[i]) {
         auto tensors = iv->in_tensors();
         auto jv = std::find(tensors.begin(), tensors.end(), in_tensors.at(i));