diff --git a/CMakeLists.txt b/CMakeLists.txt
index adc8bc655eebd6767a9077334921684a2690c02c..1df5591109c38782865c71452e2d0e167ed7f3c2 100644
--- a/CMakeLists.txt
+++ b/CMakeLists.txt
@@ -43,8 +43,10 @@ set(THIRD_PARTY_PATH "${CMAKE_BINARY_DIR}/third_party" CACHE STRING
 
 set(THIRD_PARTY_BUILD_TYPE Release)
 
-option(WITH_AVX         "Compile PaddlePaddle with AVX intrinsics"      ${AVX_FOUND})
-option(WITH_MKL         "Compile PaddlePaddle with MKL support."        ${AVX_FOUND})
+option(WITH_AVX         "Compile Paddle Serving with AVX intrinsics"      ${AVX_FOUND})
+option(WITH_MKL         "Compile Paddle Serving with MKL support."        ${AVX_FOUND})
+option(CLIENT_ONLY      "Compile client libraries and demos only"
+        FALSE)
 
 set(WITH_MKLML ${WITH_MKL})
 if (NOT DEFINED WITH_MKLDNN)
@@ -56,19 +58,22 @@ if (NOT DEFINED WITH_MKLDNN)
     endif()
 endif()
 
-include(external/mklml)
+include(external/leveldb)
 include(external/zlib)
+include(external/boost)
+include(external/protobuf)
+include(external/brpc)
 include(external/gflags)
 include(external/glog)
-include(external/leveldb)
-include(external/protobuf)
 include(external/snappy)
-include(external/brpc)
-include(external/boost)
-include(flags)
 include(generic)
+include(flags)
+
+if (NOT CLIENT_ONLY)
+include(external/mklml)
 include(paddlepaddle)
 include(external/opencv)
+endif()
 
 include_directories(${PADDLE_SERVING_SOURCE_DIR})
 include_directories(${PADDLE_SERVING_BINARY_DIR})
@@ -81,6 +86,7 @@ set(EXTERNAL_LIBS
     brpc
 )
 
+
 if(WITH_MKLML)
     list(APPEND EXTERNAL_LIBS ${MKLML_IOMP_LIB})
 endif()
@@ -90,9 +96,17 @@ if(WITH_MKLDNN)
     list(APPEND EXTERNAL_LIBS ${MKLDNN_LIB})
 endif()
 
+if (NOT CLIENT_ONLY)
+    list(APPEND EXTERNAL_LIBS paddlepaddle)
+    list(APPEND EXTERNAL_LIBS opencv)
+endif()
+
 add_subdirectory(configure)
 add_subdirectory(pdcodegen)
+add_subdirectory(sdk-cpp)
+
+if (NOT CLIENT_ONLY)
 add_subdirectory(predictor)
 add_subdirectory(inferencer-fluid-cpu)
 add_subdirectory(serving)
-add_subdirectory(sdk-cpp)
+endif()
diff --git a/doc/INSTALL.md b/doc/INSTALL.md
index 1285c905eee9826f80d5204b33fda0ccd23b5ab9..64e482d6438f2d6e5da8ea6ce98e0933b18a0f7b 100644
--- a/doc/INSTALL.md
+++ b/doc/INSTALL.md
@@ -60,7 +60,8 @@ $ make install
 
 因Paddle Serving依托于PaddlePaddle项目进行构建，以下编译选项其实是传递给PaddlePaddle的编译选项：
 
-| 编译选项 | 说明|
-|----------|-----|
+| 编译选项 | 说明 |
+|----------|------|
 | WITH_AVX | Compile PaddlePaddle with AVX intrinsics |
 | WITH_MKL | Compile PaddlePaddle with MKLML library |
+| CLINET_ONLY | Compile client libraries and demos only |
diff --git a/sdk-cpp/demo/dense_format.cpp b/sdk-cpp/demo/dense_format.cpp
index 43d73630c6ffad4aa400e3f9b6c4e80fc8778d2d..beb13877fffa1cc48c73a8b831ffda302987314a 100644
--- a/sdk-cpp/demo/dense_format.cpp
+++ b/sdk-cpp/demo/dense_format.cpp
@@ -17,7 +17,7 @@
 #include <unistd.h>
 
 #include <fstream>
-#include "predictor/builtin_format.pb.h"
+#include "sdk-cpp/builtin_format.pb.h"
 #include "sdk-cpp/dense_service.pb.h"
 #include "sdk-cpp/include/common.h"
 #include "sdk-cpp/include/predictor_sdk.h"
diff --git a/sdk-cpp/demo/echo.cpp b/sdk-cpp/demo/echo.cpp
index b296cd281cb891a7c6f147e89e7d510e1c3cc0c6..ca68d2cbafd177864bb3e97164bb57959adf7430 100644
--- a/sdk-cpp/demo/echo.cpp
+++ b/sdk-cpp/demo/echo.cpp
@@ -17,7 +17,7 @@
 #include <unistd.h>
 
 #include <fstream>
-#include "predictor/builtin_format.pb.h"
+#include "sdk-cpp/builtin_format.pb.h"
 #include "sdk-cpp/echo_service.pb.h"
 #include "sdk-cpp/include/common.h"
 #include "sdk-cpp/include/predictor_sdk.h"
diff --git a/sdk-cpp/demo/int64tensor_format.cpp b/sdk-cpp/demo/int64tensor_format.cpp
index 15db68a2779bb93d8ceb7e9ea3573cd7b780a950..97160d82969d5932cc89416c8aecdbad117705de 100644
--- a/sdk-cpp/demo/int64tensor_format.cpp
+++ b/sdk-cpp/demo/int64tensor_format.cpp
@@ -17,7 +17,7 @@
 #include <unistd.h>
 
 #include <fstream>
-#include "predictor/builtin_format.pb.h"
+#include "sdk-cpp/builtin_format.pb.h"
 #include "sdk-cpp/include/common.h"
 #include "sdk-cpp/include/predictor_sdk.h"
 #include "sdk-cpp/int64tensor_service.pb.h"
diff --git a/sdk-cpp/demo/sparse_format.cpp b/sdk-cpp/demo/sparse_format.cpp
index 78c237d76827d8d2fa591a026e2ac299b3dca1ce..c428881dc909237754a66f5c116fa2b4e266afeb 100644
--- a/sdk-cpp/demo/sparse_format.cpp
+++ b/sdk-cpp/demo/sparse_format.cpp
@@ -17,7 +17,7 @@
 #include <unistd.h>
 
 #include <fstream>
-#include "predictor/builtin_format.pb.h"
+#include "sdk-cpp/builtin_format.pb.h"
 #include "sdk-cpp/include/common.h"
 #include "sdk-cpp/include/predictor_sdk.h"
 #include "sdk-cpp/sparse_service.pb.h"
diff --git a/sdk-cpp/demo/text_classification.cpp b/sdk-cpp/demo/text_classification.cpp
index 493a24b9a977ea458d6073e8419ace6aa8a6c6f0..6e9555d06a7c3d7c63645aa1f38bec2c62399318 100644
--- a/sdk-cpp/demo/text_classification.cpp
+++ b/sdk-cpp/demo/text_classification.cpp
@@ -17,7 +17,7 @@
 #include <unistd.h>
 
 #include <fstream>
-#include "predictor/builtin_format.pb.h"
+#include "sdk-cpp/builtin_format.pb.h"
 #include "sdk-cpp/include/common.h"
 #include "sdk-cpp/include/predictor_sdk.h"
 #include "sdk-cpp/text_classification.pb.h"
diff --git a/sdk-cpp/demo/text_classification_press.cpp b/sdk-cpp/demo/text_classification_press.cpp
index a51afa3fa7e2dd8aa6f0e6f90d935d1f9415ebd1..04a9d2f693fd391676e8655dfff9cbfdb3333734 100644
--- a/sdk-cpp/demo/text_classification_press.cpp
+++ b/sdk-cpp/demo/text_classification_press.cpp
@@ -19,7 +19,7 @@
 #include <atomic>
 #include <fstream>
 #include <thread>  // NOLINT
-#include "predictor/builtin_format.pb.h"
+#include "sdk-cpp/builtin_format.pb.h"
 #include "sdk-cpp/include/common.h"
 #include "sdk-cpp/include/predictor_sdk.h"
 #include "sdk-cpp/text_classification.pb.h"
@@ -328,18 +328,26 @@ int main(int argc, char **argv) {
   LOG(INFO) << "Total requests: " << round_times.size();
   LOG(INFO) << "Max concurrency: " << FLAGS_concurrency;
   LOG(INFO) << "Elapse ms (wall-time): " << elapse_ms;
-  double qps = (static_cast<double>(count) / elapse_ms) * 1000;
+
+  double qps = 0.0;
+  if (elapse_ms != 0) {
+    qps = (static_cast<double>(count) / elapse_ms) * 1000;
+  }
 
   LOG(INFO) << "QPS: " << qps / FLAGS_batch_size << "/s";
   LOG(INFO) << "Accuracy " << static_cast<double>(correct) / count;
 
   LOG(INFO) << "Latency statistics: ";
-  LOG(INFO) << "Average ms: " << total_ms / round_times.size();
-  LOG(INFO) << "50 percent ms: " << round_times[percent_pos_50];
-  LOG(INFO) << "80 percent ms: " << round_times[percent_pos_80];
-  LOG(INFO) << "90 percent ms: " << round_times[percent_pos_90];
-  LOG(INFO) << "99 percent ms: " << round_times[percent_pos_99];
-  LOG(INFO) << "99.9 percent ms: " << round_times[percent_pos_999];
+  if (round_times.size() != 0) {
+    LOG(INFO) << "Average ms: " << total_ms / round_times.size();
+    LOG(INFO) << "50 percent ms: " << round_times[percent_pos_50];
+    LOG(INFO) << "80 percent ms: " << round_times[percent_pos_80];
+    LOG(INFO) << "90 percent ms: " << round_times[percent_pos_90];
+    LOG(INFO) << "99 percent ms: " << round_times[percent_pos_99];
+    LOG(INFO) << "99.9 percent ms: " << round_times[percent_pos_999];
+  } else {
+    LOG(INFO) << "N/A";
+  }
 
   return 0;
 }
diff --git a/sdk-cpp/demo/ximage.cpp b/sdk-cpp/demo/ximage.cpp
index 226f6420da3903bf5ac6e8afafeef8bddecb95d7..426a138e6f186acbaa99acb2215cbff00100ebab 100644
--- a/sdk-cpp/demo/ximage.cpp
+++ b/sdk-cpp/demo/ximage.cpp
@@ -17,7 +17,7 @@
 #include <unistd.h>
 
 #include <fstream>
-#include "predictor/builtin_format.pb.h"
+#include "sdk-cpp/builtin_format.pb.h"
 #include "sdk-cpp/image_class.pb.h"
 #include "sdk-cpp/include/common.h"
 #include "sdk-cpp/include/predictor_sdk.h"