diff --git a/demo-client/src/data_pre.h b/demo-client/src/data_pre.h
index ecd1168dae85275923dccf263b9b25e30d21295a..0d410c250dc2a8207dd5c5f479e97a611cfbf1cd 100644
--- a/demo-client/src/data_pre.h
+++ b/demo-client/src/data_pre.h
@@ -14,6 +14,7 @@
 #ifndef SERVING_DEMO_CLIENT_SRC_DATA_PRE_H_
 #define SERVING_DEMO_CLIENT_SRC_DATA_PRE_H_
 #include <sys/stat.h>
+#include <iostream>
 #include <map>
 #include <string>
 #include <vector>
diff --git a/demo-serving/conf/model_toolkit.prototxt b/demo-serving/conf/model_toolkit.prototxt
new file mode 100644
index 0000000000000000000000000000000000000000..88efaefba514f4763952d1b6d21d880d6d68942f
--- /dev/null
+++ b/demo-serving/conf/model_toolkit.prototxt
@@ -0,0 +1,10 @@
+engines {
+  name: "bert"
+  type: "FLUID_GPU_ANALYSIS_DIR"
+  reloadable_meta: "./data/model/paddle/fluid_time_file"
+  reloadable_type: "timestamp_ne"
+  model_data_path: "./data/model/paddle/fluid/bert"
+  runtime_thread_num: 0
+  batch_infer_size: 0
+  enable_batch_align: 0
+}
diff --git a/demo-serving/conf/service.prototxt b/demo-serving/conf/service.prototxt
new file mode 100644
index 0000000000000000000000000000000000000000..e630669e38cbc1c804611ef10e744b6e0b7fe76c
--- /dev/null
+++ b/demo-serving/conf/service.prototxt
@@ -0,0 +1,4 @@
+services {
+  name: "BertService"
+  workflows: "workflow9"
+}
diff --git a/demo-serving/op/bert_service_op.cpp b/demo-serving/op/bert_service_op.cpp
index 58ba0470f1627622f8591f7d62229b58b56b8609..d393e04ab0215dce70b6b190a76b97ca09806be6 100644
--- a/demo-serving/op/bert_service_op.cpp
+++ b/demo-serving/op/bert_service_op.cpp
@@ -137,7 +137,7 @@ int BertServiceOp::inference() {
         << " seq_len : " << out->at(0).shape[1]
         << " emb_size : " << out->at(0).shape[2];
 
-    float *out_data = reinterpret_cast<float *>out->at(0).data.data();
+    float *out_data = reinterpret_cast<float *>(out->at(0).data.data());
     for (uint32_t bi = 0; bi < batch_size; bi++) {
       BertResInstance *res_instance = res->add_instances();
       for (uint32_t si = 0; si < MAX_SEQ_LEN; si++) {
@@ -151,7 +151,7 @@ int BertServiceOp::inference() {
 #else
   LOG(INFO) << "batch_size : " << out->at(0).shape[0]
             << " emb_size : " << out->at(0).shape[1];
-  float *out_data = reinterpret_cast<float *> out->at(0).data.data();
+  float *out_data = reinterpret_cast<float *>(out->at(0).data.data());
   for (uint32_t bi = 0; bi < batch_size; bi++) {
     BertResInstance *res_instance = res->add_instances();
     for (uint32_t si = 0; si < 1; si++) {