diff --git a/paddle/fluid/inference/analysis/subgraph_splitter.cc b/paddle/fluid/inference/analysis/subgraph_splitter.cc
index b879067d2f2f6294c50e0adb21f9399a7c36698a..efc144394128d542a17f19160cb4fd84a9e028d4 100644
--- a/paddle/fluid/inference/analysis/subgraph_splitter.cc
+++ b/paddle/fluid/inference/analysis/subgraph_splitter.cc
@@ -309,6 +309,7 @@ void SubGraphFuse::operator()() { ReplaceNodesWithSubGraphs(); }
 void SubGraphFuse::ReplaceNodesWithSubGraphs() {
   auto subgraphs = SubGraphSplitter(graph_, node_inside_subgraph_teller_)();
   for (auto &subgraph : subgraphs) {
+    if (subgraph.size() <= 3) continue;
     std::unordered_set<Node *> subgraph_uniq(subgraph.begin(), subgraph.end());
     // replace this sub-graph with the first node. Two steps: 1. Create a Block
     // Node that contains this subgraph 2. Mark the nodes inside the sub-graph
diff --git a/paddle/fluid/inference/tests/api/CMakeLists.txt b/paddle/fluid/inference/tests/api/CMakeLists.txt
index 508ef1ce40aa0882a0f39a85f97511fd9ea2a8a5..9c057affca9c4b6b3fcd4574a587af1b78145b5c 100644
--- a/paddle/fluid/inference/tests/api/CMakeLists.txt
+++ b/paddle/fluid/inference/tests/api/CMakeLists.txt
@@ -85,3 +85,11 @@ if (WITH_ANAKIN AND WITH_MKL) # only needed in CI
                DEPS inference_anakin_api_shared dynload_cuda SERIAL)
    endif()
 endif()
+
+if(WITH_GPU AND TENSORRT_FOUND)
+   set(TRT_MODEL_INSTALL_DIR "${INFERENCE_DEMO_INSTALL_DIR}/trt")
+   inference_download_and_uncompress(${TRT_MODEL_INSTALL_DIR} ${INFERENCE_URL}/tensorrt_test "trt_test_models.tar.gz")
+   cc_test(test_trt_models SRCS trt_models_tester.cc  
+     ARGS --dirname=${TRT_MODEL_INSTALL_DIR}/trt_test_models
+     DEPS paddle_inference_tensorrt_subgraph_engine)
+endif()
diff --git a/paddle/fluid/inference/tests/api/trt_models_tester.cc b/paddle/fluid/inference/tests/api/trt_models_tester.cc
new file mode 100644
index 0000000000000000000000000000000000000000..79ee9b23a94da6b6b038c89b595b169e24da564c
--- /dev/null
+++ b/paddle/fluid/inference/tests/api/trt_models_tester.cc
@@ -0,0 +1,105 @@
+// Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserved.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//     http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include <gflags/gflags.h>
+#include <glog/logging.h>
+#include <gtest/gtest.h>
+#include "paddle/fluid/inference/analysis/analyzer.h"
+#include "paddle/fluid/inference/api/paddle_inference_api.h"
+
+namespace paddle {
+
+DEFINE_string(dirname, "", "Directory of the inference model.");
+
+NativeConfig GetConfigNative() {
+  NativeConfig config;
+  config.model_dir = FLAGS_dirname;
+  // LOG(INFO) << "dirname  " << config.model_dir;
+  config.fraction_of_gpu_memory = 0.7;
+  config.use_gpu = true;
+  config.device = 0;
+  return config;
+}
+
+TensorRTConfig GetConfigTRT() {
+  TensorRTConfig config;
+  config.model_dir = FLAGS_dirname;
+  config.use_gpu = true;
+  config.fraction_of_gpu_memory = 0.1;
+  config.device = 0;
+  config.max_batch_size = 3;
+  return config;
+}
+
+void CompareTensorRTWithFluid(int batch_size, std::string model_dirname) {
+  NativeConfig config0 = GetConfigNative();
+  config0.model_dir = model_dirname;
+
+  TensorRTConfig config1 = GetConfigTRT();
+  config1.model_dir = model_dirname;
+  config1.max_batch_size = batch_size;
+
+  auto predictor0 =
+      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config0);
+  auto predictor1 =
+      CreatePaddlePredictor<TensorRTConfig,
+                            PaddleEngineKind::kAutoMixedTensorRT>(config1);
+  // Prepare inputs
+  int height = 224;
+  int width = 224;
+  float *data = new float[batch_size * 3 * height * width];
+  memset(data, 0, sizeof(float) * (batch_size * 3 * height * width));
+  data[0] = 1.0f;
+
+  // Prepare inputs
+  PaddleTensor tensor;
+  tensor.name = "input_0";
+  tensor.shape = std::vector<int>({batch_size, 3, height, width});
+  tensor.data = PaddleBuf(static_cast<void *>(data),
+                          sizeof(float) * (batch_size * 3 * height * width));
+  tensor.dtype = PaddleDType::FLOAT32;
+  std::vector<PaddleTensor> paddle_tensor_feeds(1, tensor);
+
+  // Prepare outputs
+  std::vector<PaddleTensor> outputs0;
+  std::vector<PaddleTensor> outputs1;
+  CHECK(predictor0->Run(paddle_tensor_feeds, &outputs0));
+
+  CHECK(predictor1->Run(paddle_tensor_feeds, &outputs1, batch_size));
+
+  // Get output.
+  ASSERT_EQ(outputs0.size(), 1UL);
+  ASSERT_EQ(outputs1.size(), 1UL);
+
+  const size_t num_elements = outputs0.front().data.length() / sizeof(float);
+  const size_t num_elements1 = outputs1.front().data.length() / sizeof(float);
+  EXPECT_EQ(num_elements, num_elements1);
+
+  auto *data0 = static_cast<float *>(outputs0.front().data.data());
+  auto *data1 = static_cast<float *>(outputs1.front().data.data());
+
+  ASSERT_GT(num_elements, 0UL);
+  for (size_t i = 0; i < std::min(num_elements, num_elements1); i++) {
+    EXPECT_NEAR(data0[i], data1[i], 1e-3);
+  }
+}
+
+TEST(trt_models_test, main) {
+  std::vector<std::string> infer_models = {"mobilenet", "resnet50",
+                                           "resnext50"};
+  for (auto &model_dir : infer_models) {
+    CompareTensorRTWithFluid(1, FLAGS_dirname + "/" + model_dir);
+  }
+}
+}  // namespace paddle