diff --git a/paddle/fluid/inference/api/demo_ci/run.sh b/paddle/fluid/inference/api/demo_ci/run.sh
index 76238070cda725463820e0e0834f2594382c8e57..65c95f0834a9356fc14faed8342f5d1e474edf8f 100755
--- a/paddle/fluid/inference/api/demo_ci/run.sh
+++ b/paddle/fluid/inference/api/demo_ci/run.sh
@@ -100,19 +100,17 @@ for WITH_STATIC_LIB in ON OFF; do
     rm -rf *
     cmake .. -DPADDLE_LIB=${PADDLE_ROOT}/build/fluid_install_dir/ \
       -DWITH_MKL=$TURN_ON_MKL \
-      -DDEMO_NAME=vis_demo \
+      -DDEMO_NAME=trt_mobilenet_demo \
       -DWITH_GPU=$TEST_GPU_CPU \
       -DWITH_STATIC_LIB=$WITH_STATIC_LIB \
       -DUSE_TENSORRT=$USE_TENSORRT \
       -DTENSORRT_INCLUDE_DIR=$TENSORRT_INCLUDE_DIR \
       -DTENSORRT_LIB_DIR=$TENSORRT_LIB_DIR
     make -j 
-    ./vis_demo \
+    ./trt_mobilenet_demo \
       --modeldir=$DATA_DIR/mobilenet/model \
       --data=$DATA_DIR/mobilenet/data.txt \
-      --refer=$DATA_DIR/mobilenet/result.txt \
-      --use_gpu=true \
-      --use_trt=true
+      --refer=$DATA_DIR/mobilenet/result.txt 
   fi
 done
 set +x
diff --git a/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc
new file mode 100644
index 0000000000000000000000000000000000000000..43776278598681a54e22f8443f728c75a7d43e61
--- /dev/null
+++ b/paddle/fluid/inference/api/demo_ci/trt_mobilenet_demo.cc
@@ -0,0 +1,88 @@
+/* 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. */
+
+/*
+ * This file contains demo of mobilenet for tensorrt.
+ */
+
+#include <gflags/gflags.h>
+#include <glog/logging.h>  // use glog instead of CHECK to avoid importing other paddle header files.
+#include <fstream>
+#include <iostream>
+
+// #include "paddle/fluid/platform/enforce.h"
+#include "paddle/fluid/inference/demo_ci/utils.h"
+
+#ifdef PADDLE_WITH_CUDA
+DECLARE_double(fraction_of_gpu_memory_to_use);
+#endif
+DEFINE_string(modeldir, "", "Directory of the inference model.");
+DEFINE_string(refer, "", "path to reference result for comparison.");
+DEFINE_string(
+    data, "",
+    "path of data; each line is a record, format is "
+    "'<space splitted floats as data>\t<space splitted ints as shape'");
+
+namespace paddle {
+namespace demo {
+
+/*
+ * Use the native fluid engine to inference the demo.
+ */
+void Main() {
+  std::unique_ptr<PaddlePredictor> predictor;
+  paddle::contrib::MixedRTConfig config;
+  config.param_file = FLAGS_modeldir + "/__params__";
+  config.prog_file = FLAGS_modeldir + "/__model__";
+  config.use_gpu = true;
+  config.device = 0;
+  config.max_batch_size = 1;
+  config.fraction_of_gpu_memory = 0.1;  // set by yourself
+  predictor = CreatePaddlePredictor<paddle::contrib::MixedRTConfig>(config);
+
+  VLOG(3) << "begin to process data";
+  // Just a single batch of data.
+  std::string line;
+  std::ifstream file(FLAGS_data);
+  std::getline(file, line);
+  auto record = ProcessALine(line);
+  file.close();
+
+  // Inference.
+  PaddleTensor input;
+  input.shape = record.shape;
+  input.data =
+      PaddleBuf(record.data.data(), record.data.size() * sizeof(float));
+  input.dtype = PaddleDType::FLOAT32;
+
+  VLOG(3) << "run executor";
+  std::vector<PaddleTensor> output;
+  predictor->Run({input}, &output, 1);
+
+  VLOG(3) << "output.size " << output.size();
+  auto& tensor = output.front();
+  VLOG(3) << "output: " << SummaryTensor(tensor);
+
+  // compare with reference result
+  CheckOutput(FLAGS_refer, tensor);
+}
+
+}  // namespace demo
+}  // namespace paddle
+
+int main(int argc, char** argv) {
+  google::ParseCommandLineFlags(&argc, &argv, true);
+  paddle::demo::Main();
+  return 0;
+}
diff --git a/paddle/fluid/inference/api/demo_ci/utils.h b/paddle/fluid/inference/api/demo_ci/utils.h
index cb8990671162dff47228736e69617229528cc093..4792c97fe7d0a3f9c904774ad4a8e580cefcf237 100644
--- a/paddle/fluid/inference/api/demo_ci/utils.h
+++ b/paddle/fluid/inference/api/demo_ci/utils.h
@@ -14,6 +14,8 @@
 
 #pragma once
 #include <algorithm>
+#include <fstream>
+#include <iostream>
 #include <string>
 #include <vector>
 #include "paddle/fluid/inference/paddle_inference_api.h"
@@ -21,6 +23,11 @@
 namespace paddle {
 namespace demo {
 
+struct Record {
+  std::vector<float> data;
+  std::vector<int32_t> shape;
+};
+
 static void split(const std::string& str, char sep,
                   std::vector<std::string>* pieces) {
   pieces->clear();
@@ -39,6 +46,58 @@ static void split(const std::string& str, char sep,
   }
 }
 
+Record ProcessALine(const std::string& line) {
+  VLOG(3) << "process a line";
+  std::vector<std::string> columns;
+  split(line, '\t', &columns);
+  CHECK_EQ(columns.size(), 2UL)
+      << "data format error, should be <data>\t<shape>";
+
+  Record record;
+  std::vector<std::string> data_strs;
+  split(columns[0], ' ', &data_strs);
+  for (auto& d : data_strs) {
+    record.data.push_back(std::stof(d));
+  }
+
+  std::vector<std::string> shape_strs;
+  split(columns[1], ' ', &shape_strs);
+  for (auto& s : shape_strs) {
+    record.shape.push_back(std::stoi(s));
+  }
+  VLOG(3) << "data size " << record.data.size();
+  VLOG(3) << "data shape size " << record.shape.size();
+  return record;
+}
+
+void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
+  std::string line;
+  std::ifstream file(referfile);
+  std::getline(file, line);
+  auto refer = ProcessALine(line);
+  file.close();
+
+  size_t numel = output.data.length() / PaddleDtypeSize(output.dtype);
+  VLOG(3) << "predictor output numel " << numel;
+  VLOG(3) << "reference output numel " << refer.data.size();
+  CHECK_EQ(numel, refer.data.size());
+  switch (output.dtype) {
+    case PaddleDType::INT64: {
+      for (size_t i = 0; i < numel; ++i) {
+        CHECK_EQ(static_cast<int64_t*>(output.data.data())[i], refer.data[i]);
+      }
+      break;
+    }
+    case PaddleDType::FLOAT32:
+      for (size_t i = 0; i < numel; ++i) {
+        CHECK_LT(
+            fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
+            1e-5);
+      }
+      break;
+  }
+}
+
 /*
  * Get a summary of a PaddleTensor content.
  */
diff --git a/paddle/fluid/inference/api/demo_ci/vis_demo.cc b/paddle/fluid/inference/api/demo_ci/vis_demo.cc
index b9d627b4a510df7901b776de01fc00420c6eec53..db61786e2fefda29256d84b5357028ec9c39b014 100644
--- a/paddle/fluid/inference/api/demo_ci/vis_demo.cc
+++ b/paddle/fluid/inference/api/demo_ci/vis_demo.cc
@@ -18,10 +18,6 @@ limitations under the License. */
 
 #include <gflags/gflags.h>
 #include <glog/logging.h>  // use glog instead of CHECK to avoid importing other paddle header files.
-#include <fstream>
-#include <iostream>
-
-// #include "paddle/fluid/platform/enforce.h"
 #include "paddle/fluid/inference/demo_ci/utils.h"
 
 #ifdef PADDLE_WITH_CUDA
@@ -34,99 +30,28 @@ DEFINE_string(
     "path of data; each line is a record, format is "
     "'<space splitted floats as data>\t<space splitted ints as shape'");
 DEFINE_bool(use_gpu, false, "Whether use gpu.");
-DEFINE_bool(use_trt, false, "Whether use trt.");
 
 namespace paddle {
 namespace demo {
 
-struct Record {
-  std::vector<float> data;
-  std::vector<int32_t> shape;
-};
-
-void split(const std::string& str, char sep, std::vector<std::string>* pieces);
-
-Record ProcessALine(const std::string& line) {
-  VLOG(3) << "process a line";
-  std::vector<std::string> columns;
-  split(line, '\t', &columns);
-  CHECK_EQ(columns.size(), 2UL)
-      << "data format error, should be <data>\t<shape>";
-
-  Record record;
-  std::vector<std::string> data_strs;
-  split(columns[0], ' ', &data_strs);
-  for (auto& d : data_strs) {
-    record.data.push_back(std::stof(d));
-  }
-
-  std::vector<std::string> shape_strs;
-  split(columns[1], ' ', &shape_strs);
-  for (auto& s : shape_strs) {
-    record.shape.push_back(std::stoi(s));
-  }
-  VLOG(3) << "data size " << record.data.size();
-  VLOG(3) << "data shape size " << record.shape.size();
-  return record;
-}
-
-void CheckOutput(const std::string& referfile, const PaddleTensor& output) {
-  std::string line;
-  std::ifstream file(referfile);
-  std::getline(file, line);
-  auto refer = ProcessALine(line);
-  file.close();
-
-  size_t numel = output.data.length() / PaddleDtypeSize(output.dtype);
-  VLOG(3) << "predictor output numel " << numel;
-  VLOG(3) << "reference output numel " << refer.data.size();
-  CHECK_EQ(numel, refer.data.size());
-  switch (output.dtype) {
-    case PaddleDType::INT64: {
-      for (size_t i = 0; i < numel; ++i) {
-        CHECK_EQ(static_cast<int64_t*>(output.data.data())[i], refer.data[i]);
-      }
-      break;
-    }
-    case PaddleDType::FLOAT32:
-      for (size_t i = 0; i < numel; ++i) {
-        CHECK_LT(
-            fabs(static_cast<float*>(output.data.data())[i] - refer.data[i]),
-            1e-5);
-      }
-      break;
-  }
-}
-
 /*
  * Use the native fluid engine to inference the demo.
  */
-void Main(bool use_gpu, bool use_trt) {
+void Main(bool use_gpu) {
   std::unique_ptr<PaddlePredictor> predictor;
-  if (!use_trt) {
-    NativeConfig config;
-    config.param_file = FLAGS_modeldir + "/__params__";
-    config.prog_file = FLAGS_modeldir + "/__model__";
-    config.use_gpu = use_gpu;
-    config.device = 0;
-    if (FLAGS_use_gpu) {
-      config.fraction_of_gpu_memory = 0.1;  // set by yourself
-    }
-
-    VLOG(3) << "init predictor";
-    predictor =
-        CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
-  } else {
-    paddle::contrib::MixedRTConfig config;
-    config.param_file = FLAGS_modeldir + "/__params__";
-    config.prog_file = FLAGS_modeldir + "/__model__";
-    config.use_gpu = true;
-    config.device = 0;
-    config.max_batch_size = 1;
+  NativeConfig config;
+  config.param_file = FLAGS_modeldir + "/__params__";
+  config.prog_file = FLAGS_modeldir + "/__model__";
+  config.use_gpu = use_gpu;
+  config.device = 0;
+  if (FLAGS_use_gpu) {
     config.fraction_of_gpu_memory = 0.1;  // set by yourself
-    predictor = CreatePaddlePredictor<paddle::contrib::MixedRTConfig>(config);
   }
 
+  VLOG(3) << "init predictor";
+  predictor =
+      CreatePaddlePredictor<NativeConfig, PaddleEngineKind::kNative>(config);
+
   VLOG(3) << "begin to process data";
   // Just a single batch of data.
   std::string line;
@@ -159,12 +84,10 @@ void Main(bool use_gpu, bool use_trt) {
 
 int main(int argc, char** argv) {
   google::ParseCommandLineFlags(&argc, &argv, true);
-  if (FLAGS_use_gpu && FLAGS_use_trt) {
-    paddle::demo::Main(true /*use_gpu*/, true);
-  } else if (FLAGS_use_gpu) {
-    paddle::demo::Main(true /*use_gpu*/, false);
+  if (FLAGS_use_gpu) {
+    paddle::demo::Main(true /*use_gpu*/);
   } else {
-    paddle::demo::Main(false /*use_gpu*/, false /*use_tensorrt*/);
+    paddle::demo::Main(false /*use_gpu*/);
   }
   return 0;
 }