diff --git a/paddle/contrib/inference/demo/simple_on_word2vec.cc b/paddle/contrib/inference/demo/simple_on_word2vec.cc
index a4ef3b71c533d7248ab6c8a222e93fd85c1e103a..9c36aa44ec75f83a7d26f17b13f1d3aad27eecd0 100644
--- a/paddle/contrib/inference/demo/simple_on_word2vec.cc
+++ b/paddle/contrib/inference/demo/simple_on_word2vec.cc
@@ -66,12 +66,12 @@ void Main(bool use_gpu) {
   }
 }
 
-void MainThreads(int num_threads) {
+void MainThreads(int num_threads, bool use_gpu) {
   // Multi-threads only support on CPU
   // 0. Create PaddlePredictor with a config.
   NativeConfig config;
   config.model_dir = FLAGS_dirname + "word2vec.inference.model";
-  config.use_gpu = false;
+  config.use_gpu = use_gpu;
   auto main_predictor =
       CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
 
@@ -113,11 +113,13 @@ void MainThreads(int num_threads) {
 }
 
 TEST(demo, word2vec_cpu) { Main(false /*use_gpu*/); }
-TEST(demo_multi_threads, word2vec_cpu_1) { MainThreads(1); }
-TEST(demo_multi_threads, word2vec_cpu_4) { MainThreads(4); }
+TEST(demo_multi_threads, word2vec_cpu_1) { MainThreads(1, false /*use_gpu*/); }
+TEST(demo_multi_threads, word2vec_cpu_4) { MainThreads(4, false /*use_gpu*/); }
 
 #ifdef PADDLE_WITH_CUDA
 TEST(demo, word2vec_gpu) { Main(true /*use_gpu*/); }
+TEST(demo_multi_threads, word2vec_gpu_1) { MainThreads(1, true /*use_gpu*/); }
+TEST(demo_multi_threads, word2vec_gpu_4) { MainThreads(4, true /*use_gpu*/); }
 #endif
 
 }  // namespace demo
diff --git a/paddle/contrib/inference/test_paddle_inference_api_impl.cc b/paddle/contrib/inference/test_paddle_inference_api_impl.cc
index 8ffe102cb913747dbd61d7b223d76c309ca975df..4b6cb7b051d1ad2c63e895017c7faf1245c22612 100644
--- a/paddle/contrib/inference/test_paddle_inference_api_impl.cc
+++ b/paddle/contrib/inference/test_paddle_inference_api_impl.cc
@@ -56,9 +56,10 @@ NativeConfig GetConfig() {
   return config;
 }
 
-TEST(paddle_inference_api_impl, word2vec) {
+void MainWord2Vec(bool use_gpu) {
   NativeConfig config = GetConfig();
   auto predictor = CreatePaddlePredictor<NativeConfig>(config);
+  config.use_gpu = use_gpu;
 
   framework::LoDTensor first_word, second_word, third_word, fourth_word;
   framework::LoD lod{{0, 1}};
@@ -106,11 +107,12 @@ TEST(paddle_inference_api_impl, word2vec) {
   free(outputs[0].data.data);
 }
 
-TEST(paddle_inference_api_impl, image_classification) {
+void MainImageClassification(bool use_gpu) {
   int batch_size = 2;
   bool use_mkldnn = false;
   bool repeat = false;
   NativeConfig config = GetConfig();
+  config.use_gpu = use_gpu;
   config.model_dir =
       FLAGS_dirname + "image_classification_resnet.inference.model";
 
@@ -155,9 +157,9 @@ TEST(paddle_inference_api_impl, image_classification) {
   free(data);
 }
 
-TEST(paddle_inference_api_native_multithreads, word2vec) {
+void MainThreadsWord2Vec(bool use_gpu) {
   NativeConfig config = GetConfig();
-  config.use_gpu = false;
+  config.use_gpu = use_gpu;
   auto main_predictor = CreatePaddlePredictor<NativeConfig>(config);
 
   // prepare inputs data and reference results
@@ -216,11 +218,11 @@ TEST(paddle_inference_api_native_multithreads, word2vec) {
   }
 }
 
-TEST(paddle_inference_api_native_multithreads, image_classification) {
+void MainThreadsImageClassification(bool use_gpu) {
   constexpr int num_jobs = 4;  // each job run 1 batch
   constexpr int batch_size = 1;
   NativeConfig config = GetConfig();
-  config.use_gpu = false;
+  config.use_gpu = use_gpu;
   config.model_dir =
       FLAGS_dirname + "image_classification_resnet.inference.model";
 
@@ -269,4 +271,29 @@ TEST(paddle_inference_api_native_multithreads, image_classification) {
   }
 }
 
+TEST(inference_api_native, word2vec_cpu) { MainWord2Vec(false /*use_gpu*/); }
+TEST(inference_api_native, word2vec_cpu_threads) {
+  MainThreadsWord2Vec(false /*use_gpu*/);
+}
+TEST(inference_api_native, image_classification_cpu) {
+  MainThreadsImageClassification(false /*use_gpu*/);
+}
+TEST(inference_api_native, image_classification_cpu_threads) {
+  MainThreadsImageClassification(false /*use_gpu*/);
+}
+
+#ifdef PADDLE_WITH_CUDA
+TEST(inference_api_native, word2vec_gpu) { MainWord2Vec(true /*use_gpu*/); }
+TEST(inference_api_native, word2vec_gpu_threads) {
+  MainThreadsWord2Vec(true /*use_gpu*/);
+}
+TEST(inference_api_native, image_classification_gpu) {
+  MainThreadsImageClassification(true /*use_gpu*/);
+}
+TEST(inference_api_native, image_classification_gpu_threads) {
+  MainThreadsImageClassification(true /*use_gpu*/);
+}
+
+#endif
+
 }  // namespace paddle