diff --git a/cmake/external/lite.cmake b/cmake/external/lite.cmake
index a39bb3b69955782f11d892639b4e1de6ea3799f8..274511e3d39df860b9012fe498277e3d74eec494 100644
--- a/cmake/external/lite.cmake
+++ b/cmake/external/lite.cmake
@@ -132,7 +132,11 @@ if (NOT LITE_SOURCE_DIR OR NOT LITE_BINARY_DIR)
 endif()
 
 if (WITH_ARM)
-  set(LITE_OUTPUT_BIN_DIR inference_lite_lib.armlinux.armv8)
+  if(LITE_WITH_XPU)
+    set(LITE_OUTPUT_BIN_DIR inference_lite_lib.armlinux.armv8.xpu)
+  else()
+    set(LITE_OUTPUT_BIN_DIR inference_lite_lib.armlinux.armv8)
+  endif()
 else()
   set(LITE_OUTPUT_BIN_DIR inference_lite_lib)
 endif()
diff --git a/paddle/fluid/inference/tests/api/lite_resnet50_test.cc b/paddle/fluid/inference/tests/api/lite_resnet50_test.cc
index b88f09ae6a6a8619cbad4e97c630b453c1a8aad1..da56a7978a2e4815e7fb9534dab8a9a41d91000c 100644
--- a/paddle/fluid/inference/tests/api/lite_resnet50_test.cc
+++ b/paddle/fluid/inference/tests/api/lite_resnet50_test.cc
@@ -26,11 +26,7 @@ namespace inference {
 TEST(AnalysisPredictor, use_gpu) {
   std::string model_dir = FLAGS_infer_model + "/" + "model";
   AnalysisConfig config;
-#if defined(PADDLE_WITH_CUDA)
   config.EnableUseGpu(100, 0);
-#elif defined(LITE_SUBGRAPH_WITH_XPU)
-  config.EnableXpu(100);
-#endif
   config.SetModel(model_dir + "/model", model_dir + "/params");
   config.EnableLiteEngine(paddle::AnalysisConfig::Precision::kFloat32, true);
 
@@ -73,6 +69,54 @@ TEST(AnalysisPredictor, use_gpu) {
   }
 }
 
+#ifdef LITE_SUBGRAPH_WITH_XPU
+TEST(AnalysisPredictor, use_xpu) {
+  std::string model_dir = FLAGS_infer_model + "/" + "model";
+  AnalysisConfig config;
+  config.EnableLiteEngine(paddle::AnalysisConfig::Precision::kFloat32, true);
+  config.EnableXpu(100);
+  config.SetModel(model_dir + "/model", model_dir + "/params");
+
+  std::vector<PaddleTensor> inputs;
+  auto predictor = CreatePaddlePredictor(config);
+  const int batch = 1;
+  const int channel = 3;
+  const int height = 318;
+  const int width = 318;
+  const int input_num = batch * channel * height * width;
+  std::vector<float> input(input_num, 1);
+
+  PaddleTensor in;
+  in.shape = {batch, channel, height, width};
+  in.data =
+      PaddleBuf(static_cast<void*>(input.data()), input_num * sizeof(float));
+  in.dtype = PaddleDType::FLOAT32;
+  inputs.emplace_back(in);
+
+  std::vector<PaddleTensor> outputs;
+  ASSERT_TRUE(predictor->Run(inputs, &outputs));
+
+  const std::vector<float> truth_values = {
+      127.84,   738.088,  1013.22,  -438.055, 366.451,  927.585,  736.341,
+      -633.776, -329.904, -430.149, -633.082, -146.597, -1324.19, -1349.29,
+      -242.68,  117.541,  -801.704, -391.428, -404.756, 453.995,  515.373,
+      -133.003, 69.3941,  590.056,  -1434.66, -1070.81, 307.093,  400.463,
+      -316.094, -587.089, -161.033, 800.357,  -96.4212, 748.706,  868.226,
+      -447.936, 112.782,  1127.24,  47.4587,  677.698,  593.126,  -336.462,
+      551.328,  397.816,  78.3572,  -715.269, 406.002,  404.149,  246.067,
+      -8.4649,  131.345,  -647.951,
+  };
+
+  const size_t expected_size = 1;
+  EXPECT_EQ(outputs.size(), expected_size);
+  float* data_o = static_cast<float*>(outputs[0].data.data());
+  for (size_t j = 0; j < outputs[0].data.length() / sizeof(float); j += 10) {
+    EXPECT_NEAR((data_o[j] - truth_values[j / 10]) / truth_values[j / 10], 0.,
+                10e-5);
+  }
+}
+#endif
+
 }  // namespace inference
 }  // namespace paddle
 
diff --git a/paddle/scripts/paddle_build.sh b/paddle/scripts/paddle_build.sh
index 21eedc6066b49ca42e5a12444f446a93401f1632..e555832ba09365b61f5f721523bd4b8d862930ae 100755
--- a/paddle/scripts/paddle_build.sh
+++ b/paddle/scripts/paddle_build.sh
@@ -288,6 +288,7 @@ EOF
         -DWITH_GLOO=${gloo_flag} \
         -DLITE_GIT_TAG=develop \
         -DWITH_XPU=${WITH_XPU:-OFF} \
+        -DXPU_SDK_ROOT=${XPU_SDK_ROOT:-""} \
         -DWITH_LITE=${WITH_LITE:-OFF} \
         -DWITH_UNITY_BUILD=${WITH_UNITY_BUILD:-OFF};build_error=$?
     if [ "$build_error" != 0 ];then