diff --git a/CMakeLists.txt b/CMakeLists.txt
index 7c497e3e048c4dd8d5c1291286de2ab9d218b914..59d6fcb07d27e1f3ab259e69d36708b775c1852a 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -54,6 +54,7 @@ option(SERVER	    "Compile Paddle Serving Server"		    OFF)
 option(APP          "Compile Paddle Serving App package"	    OFF)
 option(WITH_ELASTIC_CTR "Compile ELASITC-CTR solution"              OFF)
 option(PACK         "Compile for whl"                               OFF)
+option(WITH_TRT     "Compile Paddle Serving with TRT"       OFF)
 
 set(WITH_MKLML ${WITH_MKL})
 if (NOT DEFINED WITH_MKLDNN)
diff --git a/cmake/paddlepaddle.cmake b/cmake/paddlepaddle.cmake
index 54aae0bdc249c6eacbd4bf6b5cc42cbba9f08784..eee5369f775fae11a49ca92681a6ae84953617f0 100644
--- a/cmake/paddlepaddle.cmake
+++ b/cmake/paddlepaddle.cmake
@@ -34,7 +34,10 @@ message( "WITH_GPU = ${WITH_GPU}")
 SET(PADDLE_VERSION "1.8.1")
 
 if (WITH_GPU)
-    SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10-cudnn7-avx-mkl")
+    if (WITH_TRT)
+        SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10.1-cudnn7.6-avx-mkl-trt6")
+    else()
+        SET(PADDLE_LIB_VERSION "${PADDLE_VERSION}-gpu-cuda10-cudnn7-avx-mkl")
 else()
     if (WITH_AVX)
         if (WITH_MKLML)
diff --git a/core/configure/proto/server_configure.proto b/core/configure/proto/server_configure.proto
index 8956022685090c94be2037445c646e9fbffd1a5c..3dfc1db4412c95c9e82c7c5c2a21a29519b84267 100644
--- a/core/configure/proto/server_configure.proto
+++ b/core/configure/proto/server_configure.proto
@@ -44,6 +44,7 @@ message EngineDesc {
   optional bool static_optimization = 14;
   optional bool force_update_static_cache = 15;
   optional bool enable_ir_optimization = 16;
+  optional bool use_trt = 17;
 };
 
 // model_toolkit conf
diff --git a/paddle_inference/inferencer-fluid-gpu/include/fluid_gpu_engine.h b/paddle_inference/inferencer-fluid-gpu/include/fluid_gpu_engine.h
index 2a4da4b9b03e716b9e8148dbfd0200b887ee66e1..e9af5ecea1a170b49622abde85a9bea3af52cecf 100644
--- a/paddle_inference/inferencer-fluid-gpu/include/fluid_gpu_engine.h
+++ b/paddle_inference/inferencer-fluid-gpu/include/fluid_gpu_engine.h
@@ -178,7 +178,7 @@ class FluidGpuNativeCore : public FluidFamilyCore {
   }
 };
 
-class FluidGpuAnalysisDirCore : public FluidFamilyCore {
+class FluidTRTAnalysisDirCore : public FluidFamilyCore {
  public:
   int create(const predictor::InferEngineCreationParams& params) {
     std::string data_path = params.get_path();
@@ -198,13 +198,7 @@ class FluidGpuAnalysisDirCore : public FluidFamilyCore {
       analysis_config.EnableMemoryOptim();
     }
 
-    /*
-     if (params.enable_ir_optimization()) {
-       analysis_config.SwitchIrOptim(true);
-     } else {
-       analysis_config.SwitchIrOptim(false);
-     }
-    */
+#if 0
 
     int min_seq_len = 1;
     int max_seq_len = 512;
@@ -241,16 +235,18 @@ class FluidGpuAnalysisDirCore : public FluidFamilyCore {
         {input4_name, {batch, head_number, opt_seq_len, opt_seq_len}},
     };
 
-    analysis_config.EnableTensorRtEngine(
-        1 << 30,
-        batch,
-        5,
-        paddle::AnalysisConfig::Precision::kHalf,
-        true,
-        true);
     analysis_config.SetTRTDynamicShapeInfo(
         min_input_shape, max_input_shape, opt_input_shape);
-
+#endif
+    if (params.use_trt()) {
+      analysis_config.EnableTensorRtEngine(
+          1 << 30,
+          batch,
+          5,
+          paddle::AnalysisConfig::Precision::kFloat32,
+          true,
+          true);
+    }
     AutoLock lock(GlobalPaddleCreateMutex::instance());
     _core =
         paddle::CreatePaddlePredictor<paddle::AnalysisConfig>(analysis_config);
diff --git a/python/paddle_serving_server_gpu/__init__.py b/python/paddle_serving_server_gpu/__init__.py
index becfbb20090799aaf40d79973964e497cf599436..0d6936e94a74ab72c92ef7870fce02598a5e5306 100644
--- a/python/paddle_serving_server_gpu/__init__.py
+++ b/python/paddle_serving_server_gpu/__init__.py
@@ -73,6 +73,8 @@ def serve_args():
         default=False,
         action="store_true",
         help="Use Multi-language-service")
+    parser.add_argument(
+        "--use_trt", default=False, action="store_true", help="Use TensorRT")
     return parser.parse_args()
 
 
@@ -195,6 +197,7 @@ class Server(object):
         self.cur_path = os.getcwd()
         self.use_local_bin = False
         self.gpuid = 0
+        self.use_trt = False
         self.model_config_paths = None  # for multi-model in a workflow
 
     def set_max_concurrency(self, concurrency):
@@ -245,6 +248,9 @@ class Server(object):
     def set_gpuid(self, gpuid=0):
         self.gpuid = gpuid
 
+    def use_trt(self):
+        self.use_trt = True
+
     def _prepare_engine(self, model_config_paths, device):
         if self.model_toolkit_conf == None:
             self.model_toolkit_conf = server_sdk.ModelToolkitConf()
@@ -264,6 +270,7 @@ class Server(object):
             engine.enable_ir_optimization = self.ir_optimization
             engine.static_optimization = False
             engine.force_update_static_cache = False
+            engine.use_trt = self.use_trt
 
             if device == "cpu":
                 engine.type = "FLUID_CPU_ANALYSIS_DIR"
diff --git a/python/paddle_serving_server_gpu/serve.py b/python/paddle_serving_server_gpu/serve.py
index 3b0941a97560f11a52808fc7e152419e2cec0ba0..8f16e0c7a266c7554475d2a3c07a762cad88b91d 100644
--- a/python/paddle_serving_server_gpu/serve.py
+++ b/python/paddle_serving_server_gpu/serve.py
@@ -64,6 +64,8 @@ def start_gpu_card_model(index, gpuid, args):  # pylint: disable=doc-string-miss
     server.set_memory_optimize(mem_optim)
     server.set_ir_optimize(ir_optim)
     server.set_max_body_size(max_body_size)
+    if args.use_trt:
+        server.use_trt()
 
     server.load_model_config(model)
     server.prepare_server(workdir=workdir, port=port, device=device)