Merge branch 'k8s-2' of https://github.com/bjjwwang/serving into k8s-2

CMakeLists.txt

@@ -60,6 +60,13 @@ option(WITH_TRT             "Compile Paddle Serving with TRT"
 option(PADDLE_ON_INFERENCE  "Compile for encryption"                             ON)
 option(WITH_OPENCV	    "Compile Paddle Serving with OPENCV"                    OFF)
 
+if(NOT DEFINED VERSION_TAG)
+  set(VERSION_TAG "0.0.0")
+endif()
+if (WITH_PYTHON)
+  message(STATUS "Compile Version Tag for wheel: ${VERSION_TAG}")
+endif()
+
 if (WITH_OPENCV)
     SET(OPENCV_DIR "" CACHE PATH "Location of libraries")
     if(NOT DEFINED OPENCV_DIR)

core/general-server/op/general_reader_op.cpp

@@ -47,17 +47,14 @@ int conf_check(const Request *req,
 
   for (int i = 0; i < var_num; ++i) {
     const Tensor &tensor = req->insts(0).tensor_array(i);
-    if (model_config->_feed_type[i] !=
-        tensor.elem_type()) {
+    if (model_config->_feed_type[i] != tensor.elem_type()) {
       LOG(ERROR) << "feed type not match.";
       return -1;
     }
-    if (model_config->_feed_shape[i].size() ==
-        tensor.shape_size()) {
+    if (model_config->_feed_shape[i].size() == tensor.shape_size()) {
       for (int j = 0; j < model_config->_feed_shape[i].size(); ++j) {
         tensor.shape(j);
-        if (model_config->_feed_shape[i][j] !=
-            tensor.shape(j)) {
+        if (model_config->_feed_shape[i][j] != tensor.shape(j)) {
           LOG(ERROR) << "feed shape not match.";
           return -1;
         }
@@ -73,6 +70,11 @@ int conf_check(const Request *req,
 int GeneralReaderOp::inference() {
   // read request from client
   const Request *req = dynamic_cast<const Request *>(get_request_message());
+  if (!req) {
+    LOG(ERROR) << "Failed get request message";
+    return -1;
+  }
+
   uint64_t log_id = req->log_id();
   int input_var_num = 0;
   std::vector<int64_t> elem_type;
@@ -80,14 +82,18 @@ int GeneralReaderOp::inference() {
   std::vector<int64_t> databuf_size;
 
   GeneralBlob *res = mutable_data<GeneralBlob>();
-  TensorVector *out = &(res->tensor_vector);
-  
-  res->SetLogId(log_id);
   if (!res) {
-    LOG(ERROR) << "(logid=" << log_id
-               << ") Failed get op tls reader object output";
+    LOG(ERROR) << "(logid=" << log_id << ") Failed get GeneralBlob";
+    return -1;
   }
 
+  TensorVector *out = &(res->tensor_vector);
+  if (!out) {
+    LOG(ERROR) << "(logid=" << log_id << ") Failed get tensor_vector of res";
+    return -1;
+  }
+
+  res->SetLogId(log_id);
   Timer timeline;
   int64_t start = timeline.TimeStampUS();
   int var_num = req->insts(0).tensor_array_size();
@@ -99,7 +105,7 @@ int GeneralReaderOp::inference() {
       baidu::paddle_serving::predictor::Resource::instance();
 
   VLOG(2) << "(logid=" << log_id << ") get resource pointer done.";
-  //get the first InferOP's model_config as ReaderOp's model_config by default.
+  // get the first InferOP's model_config as ReaderOp's model_config by default.
   std::shared_ptr<PaddleGeneralModelConfig> model_config =
       resource.get_general_model_config().front();
 
@@ -140,10 +146,11 @@ int GeneralReaderOp::inference() {
       lod_tensor.dtype = paddle::PaddleDType::INT32;
       data_len = tensor.int_data_size();
     } else if (elem_type[i] == P_STRING) {
-      //use paddle::PaddleDType::UINT8 as for String.
+      // use paddle::PaddleDType::UINT8 as for String.
       elem_size[i] = sizeof(uint8_t);
       lod_tensor.dtype = paddle::PaddleDType::UINT8;
-      //this is for vector<String>, cause the databuf_size != vector<String>.size()*sizeof(char);
+      // this is for vector<String>, cause the databuf_size !=
+      // vector<String>.size()*sizeof(char);
       for (int idx = 0; idx < tensor.data_size(); idx++) {
         data_len += tensor.data()[idx].length();
       }
@@ -157,34 +164,32 @@ int GeneralReaderOp::inference() {
       for (int k = 0; k < tensor.lod_size(); ++k) {
         lod_tensor.lod[0].push_back(tensor.lod(k));
       }
+      VLOG(2) << "(logid=" << log_id << ") var[" << i
+              << "] has lod_tensor and len=" << out->at(i).lod[0].back();
     }
 
     for (int k = 0; k < tensor.shape_size(); ++k) {
       int dim = tensor.shape(k);
-      VLOG(2) << "(logid=" << log_id << ") shape for var[" << i
-              << "]: " << dim;
+      VLOG(2) << "(logid=" << log_id << ") shape for var[" << i << "]: " << dim;
       lod_tensor.shape.push_back(dim);
     }
     lod_tensor.name = model_config->_feed_name[i];
     out->push_back(lod_tensor);
 
-    
     VLOG(2) << "(logid=" << log_id << ") tensor size for var[" << i
             << "]: " << data_len;
     databuf_size[i] = data_len * elem_size[i];
     out->at(i).data.Resize(data_len * elem_size[i]);
-    VLOG(2) << "(logid=" << log_id << ") var[" << i
-            << "] is lod_tensor and len=" << out->at(i).lod[0].back();
-    
+
     if (elem_type[i] == P_INT64) {
       int64_t *dst_ptr = static_cast<int64_t *>(out->at(i).data.data());
       VLOG(2) << "(logid=" << log_id << ") first element data in var[" << i
               << "] is " << tensor.int64_data(0);
       if (!dst_ptr) {
         LOG(ERROR) << "dst_ptr is nullptr";
-            return -1;
+        return -1;
       }
-      memcpy(dst_ptr, tensor.int64_data().data(),databuf_size[i]);
+      memcpy(dst_ptr, tensor.int64_data().data(), databuf_size[i]);
       /*
       int elem_num = tensor.int64_data_size();
       for (int k = 0; k < elem_num; ++k) {
@@ -197,9 +202,9 @@ int GeneralReaderOp::inference() {
               << "] is " << tensor.float_data(0);
       if (!dst_ptr) {
         LOG(ERROR) << "dst_ptr is nullptr";
-            return -1;
+        return -1;
       }
-      memcpy(dst_ptr, tensor.float_data().data(),databuf_size[i]);
+      memcpy(dst_ptr, tensor.float_data().data(), databuf_size[i]);
       /*int elem_num = tensor.float_data_size();
       for (int k = 0; k < elem_num; ++k) {
         dst_ptr[k] = tensor.float_data(k);
@@ -210,9 +215,9 @@ int GeneralReaderOp::inference() {
               << "] is " << tensor.int_data(0);
       if (!dst_ptr) {
         LOG(ERROR) << "dst_ptr is nullptr";
-            return -1;
+        return -1;
       }
-      memcpy(dst_ptr, tensor.int_data().data(),databuf_size[i]);
+      memcpy(dst_ptr, tensor.int_data().data(), databuf_size[i]);
       /*
       int elem_num = tensor.int_data_size();
       for (int k = 0; k < elem_num; ++k) {
@@ -225,7 +230,7 @@ int GeneralReaderOp::inference() {
               << "] is " << tensor.data(0);
       if (!dst_ptr) {
         LOG(ERROR) << "dst_ptr is nullptr";
-            return -1;
+        return -1;
       }
       int elem_num = tensor.data_size();
       for (int k = 0; k < elem_num; ++k) {

core/predictor/framework/infer.cpp

+// Copyright (c) 2019 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 "core/predictor/framework/infer.h"
+
+namespace baidu {
+namespace paddle_serving {
+namespace predictor {
+
+int ReloadableInferEngine::proc_initialize_impl(
+    const configure::EngineDesc& conf, bool version) {
+  _reload_tag_file = conf.reloadable_meta();
+  _reload_type = conf.reloadable_type();
+  _model_dir = conf.model_dir();
+  _infer_thread_num = conf.runtime_thread_num();
+  _infer_batch_size = conf.batch_infer_size();
+  _infer_batch_align = conf.enable_batch_align();
+
+  _conf = conf;
+
+  if (!check_need_reload() || load(conf) != 0) {
+    LOG(ERROR) << "Failed load model_data_path" << _model_dir;
+    return -1;
+  }
+
+  if (parse_version_info(conf, version) != 0) {
+    LOG(ERROR) << "Failed parse version info";
+    return -1;
+  }
+
+  LOG(WARNING) << "Succ load model:" << _model_dir;
+  return 0;
+}
+
+int ReloadableInferEngine::proc_initialize(const configure::EngineDesc& conf,
+                                           bool version) {
+  if (proc_initialize_impl(conf, version) != 0) {
+    LOG(ERROR) << "Failed proc initialize impl";
+    return -1;
+  }
+
+  // init bsf framework
+  if (_infer_thread_num <= 0) {
+    return 0;
+  }
+
+  // init bsf framework
+  im::bsf::TaskExecutor<TaskT>::instance()->set_thread_init_fn(
+      boost::bind(&InferEngine::thrd_initialize_impl, this));
+  im::bsf::TaskExecutor<TaskT>::instance()->set_thread_reset_fn(
+      boost::bind(&InferEngine::thrd_clear_impl, this));
+  im::bsf::TaskExecutor<TaskT>::instance()->set_thread_callback_fn(
+      boost::bind(&InferEngine::task_infer_impl, this, _1, _2));
+  im::bsf::TaskExecutor<TaskT>::instance()->set_batch_size(_infer_batch_size);
+  im::bsf::TaskExecutor<TaskT>::instance()->set_batch_align(_infer_batch_align);
+  if (im::bsf::TaskExecutor<TaskT>::instance()->start(_infer_thread_num) != 0) {
+    LOG(ERROR) << "Failed start bsf executor, threads:" << _infer_thread_num;
+    return -1;
+  }
+
+  LOG(WARNING) << "Enable batch schedule framework, thread_num:"
+               << _infer_thread_num << ", batch_size:" << _infer_batch_size
+               << ", enable_batch_align:" << _infer_batch_align;
+  return 0;
+}
+
+int ReloadableInferEngine::infer(const void* in,
+                                 void* out,
+                                 uint32_t batch_size) {
+  if (_infer_thread_num <= 0) {
+    return infer_impl(in, out, batch_size);
+  }
+
+  im::bsf::TaskManager<Tensor, Tensor> task_manager;
+  task_manager.schedule(*(reinterpret_cast<const BatchTensor*>(in)),
+                        *(reinterpret_cast<BatchTensor*>(out)));
+  task_manager.wait();
+  return 0;
+}
+
+int ReloadableInferEngine::thrd_initialize() {
+  if (_infer_thread_num > 0) {
+    return 0;
+  }
+  return thrd_initialize_impl();
+}
+
+int ReloadableInferEngine::thrd_clear() {
+  if (_infer_thread_num > 0) {
+    return 0;
+  }
+  return thrd_clear_impl();
+}
+
+int ReloadableInferEngine::proc_finalize() {
+  if (proc_finalize_impl() != 0) {
+    LOG(ERROR) << "Failed proc finalize impl";
+    return -1;
+  }
+
+  if (_infer_thread_num > 0) {
+    im::bsf::TaskExecutor<TaskT>::instance()->stop();
+  }
+  return 0;
+}
+
+int ReloadableInferEngine::reload() {
+  if (check_need_reload()) {
+    LOG(WARNING) << "begin reload model[" << _model_dir << "].";
+    return load(_conf);
+  }
+  return 0;
+}
+
+int ReloadableInferEngine::parse_version_info(
+    const configure::EngineDesc& config, bool version) {
+  _version = uint64_t(-1);
+  return 0;
+}
+
+bool ReloadableInferEngine::check_need_reload() {
+  if (_reload_type == "timestamp_ne") {
+    return check_timestamp_ne();
+  } else if (_reload_type == "timestamp_gt") {
+    return check_timestamp_gt();
+  } else if (_reload_type == "md5sum") {
+    return check_md5sum();
+  } else if (_reload_type == "revision") {
+    return check_revision();
+  } else if (_reload_type == "none") {
+    return false;
+  }
+
+  LOG(ERROR) << "Not support reload type: " << _reload_type;
+  return false;
+}
+
+bool ReloadableInferEngine::check_timestamp_ne() {
+  struct stat st;
+  if (stat(_reload_tag_file.c_str(), &st) != 0) {
+    LOG(ERROR) << "Failed stat config file:" << _reload_tag_file;
+    return false;
+  }
+
+  if ((st.st_mode & S_IFREG) && st.st_mtime != _last_record.timestamp) {
+    _last_record.timestamp = st.st_mtime;
+    return true;
+  }
+
+  return false;
+}
+
+bool ReloadableInferEngine::check_timestamp_gt() {
+  struct stat st;
+  if (stat(_reload_tag_file.c_str(), &st) != 0) {
+    LOG(ERROR) << "Failed stat config file:" << _reload_tag_file;
+    return false;
+  }
+
+  if ((st.st_mode & S_IFREG) && st.st_mtime > _last_record.timestamp) {
+    _last_record.timestamp = st.st_mtime;
+    return true;
+  }
+
+  return false;
+}
+
+int VersionedInferEngine::proc_initialize(const configure::EngineDesc& conf) {
+  if (proc_initialize(conf, false) != 0) {
+    LOG(ERROR) << "Failed proc intialize engine: " << conf.name().c_str();
+    return -1;
+  }
+
+  LOG(WARNING) << "Succ proc initialize engine: " << conf.name().c_str();
+  return 0;
+}
+
+int VersionedInferEngine::proc_initialize(const configure::EngineDesc& conf,
+                                          bool version) {
+  std::string engine_type = conf.type();
+  InferEngine* engine =
+      StaticInferFactory::instance().generate_object(engine_type);
+  if (!engine) {
+    LOG(ERROR) << "Failed generate engine with type:" << engine_type;
+    return -1;
+  }
+#ifndef BCLOUD
+  VLOG(2) << "FLAGS_logtostderr " << FLAGS_logtostderr;
+  int tmp = FLAGS_logtostderr;
+  if (engine->proc_initialize(conf, version) != 0) {
+    LOG(ERROR) << "Failed initialize engine, type:" << engine_type;
+    return -1;
+  }
+  VLOG(2) << "FLAGS_logtostderr " << FLAGS_logtostderr;
+  FLAGS_logtostderr = tmp;
+#else
+  if (engine->proc_initialize(conf, version) != 0) {
+    LOG(ERROR) << "Failed initialize engine, type:" << engine_type;
+    return -1;
+  }
+#endif
+  auto r = _versions.insert(std::make_pair(engine->version(), engine));
+  if (!r.second) {
+    LOG(ERROR) << "Failed insert item: " << engine->version()
+               << ", type: " << engine_type;
+    return -1;
+  }
+  LOG(WARNING) << "Succ proc initialize version engine: " << engine->version();
+  return 0;
+}
+
+int VersionedInferEngine::proc_finalize() {
+  for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
+    if (iter->second->proc_finalize() != 0) {
+      LOG(ERROR) << "Failed proc finalize version engine: " << iter->first;
+    }
+    LOG(WARNING) << "Succ proc finalize version engine: " << iter->first;
+  }
+  return 0;
+}
+
+int VersionedInferEngine::thrd_initialize() {
+  for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
+    if (iter->second->thrd_initialize() != 0) {
+      LOG(ERROR) << "Failed thrd initialize version engine: " << iter->first;
+      return -1;
+    }
+    LOG(WARNING) << "Succ thrd initialize version engine: " << iter->first;
+  }
+  return 0;
+}
+
+int VersionedInferEngine::thrd_clear() {
+  for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
+    if (iter->second->thrd_clear() != 0) {
+      LOG(ERROR) << "Failed thrd clear version engine: " << iter->first;
+      return -1;
+    }
+  }
+  return 0;
+}
+
+int VersionedInferEngine::thrd_finalize() {
+  for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
+    if (iter->second->thrd_finalize() != 0) {
+      LOG(ERROR) << "Failed thrd finalize version engine: " << iter->first;
+      return -1;
+    }
+    LOG(WARNING) << "Succ thrd finalize version engine: " << iter->first;
+  }
+  return 0;
+}
+
+int VersionedInferEngine::reload() {
+  for (auto iter = _versions.begin(); iter != _versions.end(); ++iter) {
+    if (iter->second->reload() != 0) {
+      LOG(ERROR) << "Failed reload version engine: " << iter->first;
+      return -1;
+    }
+    LOG(WARNING) << "Succ reload version engine: " << iter->first;
+  }
+  return 0;
+}
+
+uint64_t VersionedInferEngine::version() const {
+  InferEngine* engine = default_engine();
+  if (engine) {
+    return engine->version();
+  } else {
+    return uint64_t(-1);
+  }
+}
+
+// inference interface
+InferEngine* VersionedInferEngine::default_engine() const {
+  if (_versions.size() != 1) {
+    LOG(ERROR) << "Ambiguous default engine version:" << _versions.size();
+    return NULL;
+  }
+
+  return _versions.begin()->second;
+}
+
+int VersionedInferEngine::infer(const void* in,
+                                void* out,
+                                uint32_t batch_size) {
+  InferEngine* engine = default_engine();
+  if (!engine) {
+    LOG(WARNING) << "fail to get default engine";
+    return -1;
+  }
+  return engine->infer(in, out, batch_size);
+}
+
+// versioned inference interface
+int VersionedInferEngine::infer(const void* in,
+                                void* out,
+                                uint32_t batch_size,
+                                uint64_t version) {
+  auto iter = _versions.find(version);
+  if (iter == _versions.end()) {
+    LOG(ERROR) << "Not found version engine: " << version;
+    return -1;
+  }
+
+  return iter->second->infer(in, out, batch_size);
+}
+
+template <typename T>
+T* VersionedInferEngine::get_core() {
+  InferEngine* engine = default_engine();
+  if (!engine) {
+    LOG(WARNING) << "fail to get core";
+    return NULL;
+  }
+  auto db_engine = dynamic_cast<DBReloadableInferEngine<T>*>(engine);
+  if (db_engine) {
+    return db_engine->get_core();
+  }
+  LOG(WARNING) << "fail to get core";
+  return NULL;
+}
+
+template <typename T>
+T* VersionedInferEngine::get_core(uint64_t version) {
+  auto iter = _versions.find(version);
+  if (iter == _versions.end()) {
+    LOG(ERROR) << "Not found version engine: " << version;
+    return NULL;
+  }
+
+  auto db_engine = dynamic_cast<DBReloadableInferEngine<T>*>(iter->second);
+  if (db_engine) {
+    return db_engine->get_core();
+  }
+  LOG(WARNING) << "fail to get core for " << version;
+  return NULL;
+}
+
+int VersionedInferEngine::proc_initialize_impl(
+    const configure::EngineDesc& conf, bool) {
+  return -1;
+}
+
+int VersionedInferEngine::thrd_initialize_impl() { return -1; }
+int VersionedInferEngine::thrd_finalize_impl() { return -1; }
+int VersionedInferEngine::thrd_clear_impl() { return -1; }
+int VersionedInferEngine::proc_finalize_impl() { return -1; }
+int VersionedInferEngine::infer_impl(const void* in,
+                                     void* out,
+                                     uint32_t batch_size) {
+  return -1;
+}
+int VersionedInferEngine::task_infer_impl(const BatchTensor& in,
+                                          BatchTensor& out) {  // NOLINT
+  return -1;
+}
+
+int InferManager::proc_initialize(const char* path, const char* file) {
+  ModelToolkitConf model_toolkit_conf;
+  if (configure::read_proto_conf(path, file, &model_toolkit_conf) != 0) {
+    LOG(ERROR) << "failed load infer config, path: " << path << "/" << file;
+    return -1;
+  }
+  size_t engine_num = model_toolkit_conf.engines_size();
+  for (size_t ei = 0; ei < engine_num; ++ei) {
+    LOG(INFO) << "model_toolkit_conf.engines(" << ei
+              << ").name: " << model_toolkit_conf.engines(ei).name();
+    std::string engine_name = model_toolkit_conf.engines(ei).name();
+    VersionedInferEngine* engine = new (std::nothrow) VersionedInferEngine();
+    if (!engine) {
+      LOG(ERROR) << "Failed generate versioned engine: " << engine_name;
+      return -1;
+    }
+    if (engine->proc_initialize(model_toolkit_conf.engines(ei)) != 0) {
+      LOG(ERROR) << "Failed initialize version engine, name:" << engine_name;
+      return -1;
+    }
+    auto r = _map.insert(std::make_pair(engine_name, engine));
+    if (!r.second) {
+      LOG(ERROR) << "Failed insert item: " << engine_name;
+      return -1;
+    }
+    LOG(WARNING) << "Succ proc initialize engine: " << engine_name;
+  }
+  return 0;
+}
+
+int InferManager::thrd_initialize() {
+  for (auto it = _map.begin(); it != _map.end(); ++it) {
+    if (it->second->thrd_initialize() != 0) {
+      LOG(ERROR) << "Failed thrd initialize engine, name: " << it->first;
+      return -1;
+    }
+    LOG(WARNING) << "Succ thrd initialize engine, name: " << it->first;
+  }
+  return 0;
+}
+
+int InferManager::thrd_clear() {
+  for (auto it = _map.begin(); it != _map.end(); ++it) {
+    if (it->second->thrd_clear() != 0) {
+      LOG(ERROR) << "Failed thrd clear engine, name: " << it->first;
+      return -1;
+    }
+  }
+  return 0;
+}
+
+int InferManager::reload() {
+  for (auto it = _map.begin(); it != _map.end(); ++it) {
+    if (it->second->reload() != 0) {
+      LOG(ERROR) << "Failed reload engine, name: " << it->first;
+      return -1;
+    }
+  }
+  return 0;
+}
+
+int InferManager::thrd_finalize() {
+  for (auto it = _map.begin(); it != _map.end(); ++it) {
+    if (it->second->thrd_finalize() != 0) {
+      LOG(ERROR) << "Failed thrd finalize engine, name: " << it->first;
+      return -1;
+    }
+    LOG(WARNING) << "Succ thrd finalize engine, name: " << it->first;
+  }
+  return 0;
+}
+
+int InferManager::proc_finalize() {
+  for (auto it = _map.begin(); it != _map.end(); ++it) {
+    if (it->second->proc_finalize() != 0) {
+      LOG(ERROR) << "Failed proc finalize engine, name: " << it->first;
+      return -1;
+    }
+    LOG(WARNING) << "Succ proc finalize engine, name: " << it->first;
+  }
+  _map.clear();
+  return 0;
+}
+
+// Inference interface
+int InferManager::infer(const char* model_name,
+                        const void* in,
+                        void* out,
+                        uint32_t batch_size) {
+  auto it = _map.find(model_name);
+  if (it == _map.end()) {
+    LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
+    return -1;
+  }
+  return it->second->infer(in, out, batch_size);
+}
+
+template <typename T>
+T* InferManager::get_core(const char* model_name) {
+  auto it = _map.find(model_name);
+  if (it == _map.end()) {
+    LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
+    return NULL;
+  }
+  auto infer_engine =
+      dynamic_cast<DBReloadableInferEngine<T>*>(it->second->default_engine());
+  if (infer_engine) {
+    return infer_engine->get_core();
+  }
+  LOG(WARNING) << "fail to get core for " << model_name;
+  return NULL;
+}
+
+// Versioned inference interface
+int InferManager::infer(const char* model_name,
+                        const void* in,
+                        void* out,
+                        uint32_t batch_size,
+                        uint64_t version) {
+  auto it = _map.find(model_name);
+  if (it == _map.end()) {
+    LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
+    return -1;
+  }
+  return it->second->infer(in, out, batch_size, version);
+}
+
+template <typename T>
+T* InferManager::get_core(const char* model_name, uint64_t version) {
+  auto it = _map.find(model_name);
+  if (it == _map.end()) {
+    LOG(WARNING) << "Cannot find engine in map, model name:" << model_name;
+    return NULL;
+  }
+  return it->second->get_core<T>(version);
+}
+
+int InferManager::query_version(const std::string& model, uint64_t& version) {
+  auto it = _map.find(model);
+  if (it == _map.end()) {
+    LOG(WARNING) << "Cannot find engine in map, model name:" << model;
+    return -1;
+  }
+  auto infer_engine = it->second->default_engine();
+  if (!infer_engine) {
+    LOG(WARNING) << "Cannot get default engine for model:" << model;
+    return -1;
+  }
+  version = infer_engine->version();
+  LOG(INFO) << "Succ get version: " << version << " for model: " << model;
+  return 0;
+}
+
+}  // namespace predictor
+}  // namespace paddle_serving
+}  // namespace baidu

core/predictor/framework/resource.cpp

@@ -27,22 +27,9 @@ namespace predictor {
 using configure::ResourceConf;
 using configure::GeneralModelConfig;
 using rec::mcube::CubeAPI;
-// __thread bool p_thread_initialized = false;
 
-static void dynamic_resource_deleter(void* d) {
-#if 1
-  LOG(INFO) << "dynamic_resource_delete on " << bthread_self();
-#endif
-  delete static_cast<DynamicResource*>(d);
-}
-
-DynamicResource::DynamicResource() {}
-
-DynamicResource::~DynamicResource() {}
-
-int DynamicResource::initialize() { return 0; }
-
-std::vector<std::shared_ptr<PaddleGeneralModelConfig> > Resource::get_general_model_config() {
+std::vector<std::shared_ptr<PaddleGeneralModelConfig>>
+Resource::get_general_model_config() {
   return _configs;
 }
 
@@ -96,8 +83,6 @@ void Resource::print_general_model_config(
   }
 }
 
-int DynamicResource::clear() { return 0; }
-
 int Resource::initialize(const std::string& path, const std::string& file) {
   ResourceConf resource_conf;
   if (configure::read_proto_conf(path, file, &resource_conf) != 0) {
@@ -150,29 +135,25 @@ int Resource::initialize(const std::string& path, const std::string& file) {
 
   if (FLAGS_enable_model_toolkit) {
     size_t model_toolkit_num = resource_conf.model_toolkit_path_size();
-    for (size_t mi=0; mi < model_toolkit_num; ++mi) {
+    for (size_t mi = 0; mi < model_toolkit_num; ++mi) {
       std::string model_toolkit_path = resource_conf.model_toolkit_path(mi);
       std::string model_toolkit_file = resource_conf.model_toolkit_file(mi);
 
       if (InferManager::instance().proc_initialize(
               model_toolkit_path.c_str(), model_toolkit_file.c_str()) != 0) {
         LOG(ERROR) << "failed proc initialize modeltoolkit, config: "
-                  << model_toolkit_path << "/" << model_toolkit_file;
+                   << model_toolkit_path << "/" << model_toolkit_file;
         return -1;
       }
 
       if (KVManager::instance().proc_initialize(
               model_toolkit_path.c_str(), model_toolkit_file.c_str()) != 0) {
         LOG(ERROR) << "Failed proc initialize kvmanager, config: "
-                  << model_toolkit_path << "/" << model_toolkit_file;
+                   << model_toolkit_path << "/" << model_toolkit_file;
       }
     }
   }
 
-  if (THREAD_KEY_CREATE(&_tls_bspec_key, dynamic_resource_deleter) != 0) {
-    LOG(ERROR) << "unable to create tls_bthread_key of thrd_data";
-    return -1;
-  }
   // init rocksDB or cube instance
   if (resource_conf.has_cube_config_file() &&
       resource_conf.has_cube_config_path()) {
@@ -225,18 +206,16 @@ int Resource::general_model_initialize(const std::string& path,
     return -1;
   }
   size_t general_model_num = resource_conf.general_model_path_size();
-  for (size_t gi=0; gi < general_model_num; ++gi) {
-
-
+  for (size_t gi = 0; gi < general_model_num; ++gi) {
     std::string general_model_path = resource_conf.general_model_path(gi);
     std::string general_model_file = resource_conf.general_model_file(gi);
 
     GeneralModelConfig model_config;
     if (configure::read_proto_conf(general_model_path.c_str(),
-                                  general_model_file.c_str(),
-                                  &model_config) != 0) {
-      LOG(ERROR) << "Failed initialize model config from: " << general_model_path
-                << "/" << general_model_file;
+                                   general_model_file.c_str(),
+                                   &model_config) != 0) {
+      LOG(ERROR) << "Failed initialize model config from: "
+                 << general_model_path << "/" << general_model_file;
       return -1;
     }
     auto _config = std::make_shared<PaddleGeneralModelConfig>();
@@ -249,7 +228,7 @@ int Resource::general_model_initialize(const std::string& path,
     _config->_is_lod_feed.resize(feed_var_num);
     _config->_capacity.resize(feed_var_num);
     _config->_feed_shape.resize(feed_var_num);
-    for (int i=0; i < feed_var_num; ++i) {
+    for (int i = 0; i < feed_var_num; ++i) {
       _config->_feed_name[i] = model_config.feed_var(i).name();
       _config->_feed_alias_name[i] = model_config.feed_var(i).alias_name();
       VLOG(2) << "feed var[" << i << "]: " << _config->_feed_name[i];
@@ -265,7 +244,7 @@ int Resource::general_model_initialize(const std::string& path,
         VLOG(2) << "var[" << i << "] is tensor";
         _config->_capacity[i] = 1;
         _config->_is_lod_feed[i] = false;
-        for (int j=0; j < model_config.feed_var(i).shape_size(); ++j) {
+        for (int j = 0; j < model_config.feed_var(i).shape_size(); ++j) {
           int32_t dim = model_config.feed_var(i).shape(j);
           VLOG(2) << "var[" << i << "].shape[" << i << "]: " << dim;
           _config->_feed_shape[i].push_back(dim);
@@ -279,7 +258,7 @@ int Resource::general_model_initialize(const std::string& path,
     _config->_fetch_name.resize(fetch_var_num);
     _config->_fetch_alias_name.resize(fetch_var_num);
     _config->_fetch_shape.resize(fetch_var_num);
-    for (int i=0; i < fetch_var_num; ++i) {
+    for (int i = 0; i < fetch_var_num; ++i) {
       _config->_fetch_name[i] = model_config.fetch_var(i).name();
       _config->_fetch_alias_name[i] = model_config.fetch_var(i).alias_name();
       _config->_fetch_name_to_index[_config->_fetch_name[i]] = i;
@@ -290,7 +269,7 @@ int Resource::general_model_initialize(const std::string& path,
         _config->_is_lod_fetch[i] = true;
       } else {
         _config->_is_lod_fetch[i] = false;
-        for (int j=0; j < model_config.fetch_var(i).shape_size(); ++j) {
+        for (int j = 0; j < model_config.fetch_var(i).shape_size(); ++j) {
           int dim = model_config.fetch_var(i).shape(j);
           _config->_fetch_shape[i].push_back(dim);
         }
@@ -316,36 +295,6 @@ int Resource::thread_initialize() {
     return -1;
   }
 
-  DynamicResource* p_dynamic_resource =
-      reinterpret_cast<DynamicResource*>(THREAD_GETSPECIFIC(_tls_bspec_key));
-  if (p_dynamic_resource == NULL) {
-    p_dynamic_resource = new (std::nothrow) DynamicResource;
-    if (p_dynamic_resource == NULL) {
-      LOG(ERROR) << "failed to create tls DynamicResource";
-      return -1;
-    }
-    if (p_dynamic_resource->initialize() != 0) {
-      LOG(ERROR) << "DynamicResource initialize failed.";
-      delete p_dynamic_resource;
-      p_dynamic_resource = NULL;
-      return -1;
-    }
-
-    if (THREAD_SETSPECIFIC(_tls_bspec_key, p_dynamic_resource) != 0) {
-      LOG(ERROR) << "unable to set tls DynamicResource";
-      delete p_dynamic_resource;
-      p_dynamic_resource = NULL;
-      return -1;
-    }
-  }
-#if 0
-    LOG(INFO) << "Successfully thread initialized dynamic resource";
-#else
-  LOG(INFO) << bthread_self()
-            << ": Successfully thread initialized dynamic resource "
-            << p_dynamic_resource;
-
-#endif
   return 0;
 }
 
@@ -362,26 +311,6 @@ int Resource::thread_clear() {
     LOG(ERROR) << "Failed thrd clear infer manager";
     return -1;
   }
-
-  DynamicResource* p_dynamic_resource =
-      reinterpret_cast<DynamicResource*>(THREAD_GETSPECIFIC(_tls_bspec_key));
-  if (p_dynamic_resource == NULL) {
-#if 0
-    LOG(ERROR) << "tls dynamic resource shouldn't be null after "
-        << "thread_initialize";
-#else
-    LOG(ERROR)
-        << bthread_self()
-        << ": tls dynamic resource shouldn't be null after thread_initialize";
-#endif
-    return -1;
-  }
-  if (p_dynamic_resource->clear() != 0) {
-    LOG(ERROR) << "Failed to invoke dynamic resource clear";
-    return -1;
-  }
-
-  // ...
   return 0;
 }
 size_t Resource::get_cube_quant_bits() { return this->cube_quant_bits; }

core/predictor/framework/resource.h

@@ -54,15 +54,6 @@ class PaddleGeneralModelConfig {
 };
 
 class BaseRdDict;
-struct DynamicResource {
-  DynamicResource();
-
-  ~DynamicResource();
-
-  int initialize();
-
-  int clear();
-};
 
 class Resource {
  public:
@@ -94,20 +85,17 @@ class Resource {
 
   int finalize();
 
-  std::vector<std::shared_ptr<PaddleGeneralModelConfig> > get_general_model_config();
+  std::vector<std::shared_ptr<PaddleGeneralModelConfig>>
+  get_general_model_config();
 
   void print_general_model_config(
       const std::shared_ptr<PaddleGeneralModelConfig>& config);
 
-  DynamicResource* get_dynamic_resource() {
-    return reinterpret_cast<DynamicResource*>(
-        THREAD_GETSPECIFIC(_tls_bspec_key));
-  }
   size_t get_cube_quant_bits();
 
  private:
   int thread_finalize() { return 0; }
-  std::vector<std::shared_ptr<PaddleGeneralModelConfig> > _configs;
+  std::vector<std::shared_ptr<PaddleGeneralModelConfig>> _configs;
   std::string cube_config_fullpath;
   int cube_quant_bits;  // 0 if no empty
 

doc/BAIDU_KUNLUN_XPU_SERVING.md

@@ -59,11 +59,11 @@ For example, after the Server Compiation step，the whl package will be produced
 # Request parameters description
 In order to deploy serving
  service on the arm server with Baidu Kunlun xpu chips and use the acceleration capability of Paddle-Lite，please specify the following parameters during deployment。
-|param|param description|about|
-|:--|:--|:--|
-|use_lite|using Paddle-Lite Engine|use the inference capability of Paddle-Lite|
-|use_xpu|using Baidu Kunlun for inference|need to be used with the use_lite option|
-|ir_optim|open the graph optimization|refer to[Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite)|
+| param    | param description                | about                                                              |
+| :------- | :------------------------------- | :----------------------------------------------------------------- |
+| use_lite | using Paddle-Lite Engine         | use the inference capability of Paddle-Lite                        |
+| use_xpu  | using Baidu Kunlun for inference | need to be used with the use_lite option                           |
+| ir_optim | open the graph optimization      | refer to[Paddle-Lite](https://github.com/PaddlePaddle/Paddle-Lite) |
 # Deplyment examples
 ## Download the model
 ```
@@ -78,15 +78,15 @@ There are mainly three deployment methods：
     
 The first two deployment methods are recommended。
 
-Start the rpc service, deploying on ARM server with Baidu Kunlun chips，and accelerate with Paddle-Lite and Baidu Kunlun xpu.
+Start the rpc service, deploying on ARM server with Baidu Kunlun chips，and accelerate with Paddle-Lite and Baidu Kunlun xpu.
 ```
 python3 -m paddle_serving_server_gpu.serve --model uci_housing_model --thread 6 --port 9292 --use_lite --use_xpu --ir_optim
 ```
-Start the rpc service, deploying on ARM server，and accelerate with Paddle-Lite.
+Start the rpc service, deploying on ARM server，and accelerate with Paddle-Lite.
 ```
 python3 -m paddle_serving_server_gpu.serve --model uci_housing_model --thread 6 --port 9292 --use_lite --ir_optim
 ```
-Start the rpc service, deploying on ARM server.
+Start the rpc service, deploying on ARM server.
 ```
 python3 -m paddle_serving_server_gpu.serve --model uci_housing_model --thread 6 --port 9292
 ```
@@ -103,7 +103,7 @@ fetch_map = client.predict(feed={"x": np.array(data).reshape(1,13,1)}, fetch=["p
 print(fetch_map)
 ```
 Some examples are provided below, and other models can be modifed with reference to these examples。
-|sample name|sample links|
-|:-----|:--|
-|fit_a_line|[fit_a_line_xpu](../python/examples/xpu/fit_a_line_xpu)|
-|resnet|[resnet_v2_50_xpu](../python/examples/xpu/resnet_v2_50_xpu)|
+| sample name | sample links                                                |
+| :---------- | :---------------------------------------------------------- |
+| fit_a_line  | [fit_a_line_xpu](../python/examples/xpu/fit_a_line_xpu)     |
+| resnet      | [resnet_v2_50_xpu](../python/examples/xpu/resnet_v2_50_xpu) |

doc/COMPILE.md

@@ -170,7 +170,7 @@ cmake -DPYTHON_INCLUDE_DIR=$PYTHON_INCLUDE_DIR/ \
 make -j10
 ```
 
-**Note:** After the compilation is successful, you need to set the `SERVING_BIN` path, see the following [Notes](COMPILE.md#Notes) ).
+**Note:** After the compilation is successful, you need to set the `SERVING_BIN` path, see the following [Notes](https://github.com/PaddlePaddle/Serving/blob/develop/doc/COMPILE.md#Notes).
 
 ## Compile Client
 

doc/LOW_PRECISION_DEPLOYMENT.md

+# Low-Precision Deployment for Paddle Serving
+(English|[简体中文](./LOW_PRECISION_DEPLOYMENT_CN.md))
+
+Intel CPU supports int8 and bfloat16 models, NVIDIA TensorRT supports int8 and float16 models.
+
+## Obtain the quantized model through PaddleSlim tool
+Train the low-precision models please refer to [PaddleSlim](https://paddleslim.readthedocs.io/zh_CN/latest/tutorials/quant/overview.html).
+
+## Deploy the quantized model from PaddleSlim using Paddle Serving with Nvidia TensorRT int8 mode
+
+Firstly, download the [Resnet50 int8 model](https://paddle-inference-dist.bj.bcebos.com/inference_demo/python/resnet50/ResNet50_quant.tar.gz) and convert to Paddle Serving's saved model。
+```
+wget https://paddle-inference-dist.bj.bcebos.com/inference_demo/python/resnet50/ResNet50_quant.tar.gz
+tar zxvf ResNet50_quant.tar.gz
+
+python -m paddle_serving_client.convert --dirname ResNet50_quant
+```
+Start RPC service, specify the GPU id and precision mode
+```
+python -m paddle_serving_server.serve --model serving_server --port 9393 --gpu_ids 0 --use_gpu --use_trt --precision int8 
+```
+Request the serving service with Client
+```
+from paddle_serving_client import Client
+from paddle_serving_app.reader import Sequential, File2Image, Resize, CenterCrop
+from paddle_serving_app.reader import RGB2BGR, Transpose, Div, Normalize
+
+client = Client()
+client.load_client_config(
+    "resnet_v2_50_imagenet_client/serving_client_conf.prototxt")
+client.connect(["127.0.0.1:9393"])
+
+seq = Sequential([
+    File2Image(), Resize(256), CenterCrop(224), RGB2BGR(), Transpose((2, 0, 1)),
+    Div(255), Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], True)
+])
+
+image_file = "daisy.jpg"
+img = seq(image_file)
+fetch_map = client.predict(feed={"image": img}, fetch=["score"])
+print(fetch_map["score"].reshape(-1))
+```
+
+## Reference
+* [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim)
+* [Deploy the quantized model Using Paddle Inference on Intel CPU](https://paddle-inference.readthedocs.io/en/latest/optimize/paddle_x86_cpu_int8.html)
+* [Deploy the quantized model Using Paddle Inference on Nvidia GPU](https://paddle-inference.readthedocs.io/en/latest/optimize/paddle_trt.html)
\ No newline at end of file

doc/LOW_PRECISION_DEPLOYMENT_CN.md

+# Paddle Serving低精度部署
+(简体中文|[English](./LOW_PRECISION_DEPLOYMENT.md))
+
+低精度部署, 在Intel CPU上支持int8、bfloat16模型，Nvidia TensorRT支持int8、float16模型。
+
+## 通过PaddleSlim量化生成低精度模型
+详细见[PaddleSlim量化](https://paddleslim.readthedocs.io/zh_CN/latest/tutorials/quant/overview.html)
+
+## 使用TensorRT int8加载PaddleSlim Int8量化模型进行部署
+首先下载Resnet50 [PaddleSlim量化模型](https://paddle-inference-dist.bj.bcebos.com/inference_demo/python/resnet50/ResNet50_quant.tar.gz),并转换为Paddle Serving支持的部署模型格式。
+```
+wget https://paddle-inference-dist.bj.bcebos.com/inference_demo/python/resnet50/ResNet50_quant.tar.gz
+tar zxvf ResNet50_quant.tar.gz
+
+python -m paddle_serving_client.convert --dirname ResNet50_quant
+```
+启动rpc服务, 设定所选GPU id、部署模型精度
+```
+python -m paddle_serving_server.serve --model serving_server --port 9393 --gpu_ids 0 --use_gpu --use_trt --precision int8 
+```
+使用client进行请求
+```
+from paddle_serving_client import Client
+from paddle_serving_app.reader import Sequential, File2Image, Resize, CenterCrop
+from paddle_serving_app.reader import RGB2BGR, Transpose, Div, Normalize
+
+client = Client()
+client.load_client_config(
+    "resnet_v2_50_imagenet_client/serving_client_conf.prototxt")
+client.connect(["127.0.0.1:9393"])
+
+seq = Sequential([
+    File2Image(), Resize(256), CenterCrop(224), RGB2BGR(), Transpose((2, 0, 1)),
+    Div(255), Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], True)
+])
+
+image_file = "daisy.jpg"
+img = seq(image_file)
+fetch_map = client.predict(feed={"image": img}, fetch=["score"])
+print(fetch_map["score"].reshape(-1))
+```
+
+## 参考文档
+* [PaddleSlim](https://github.com/PaddlePaddle/PaddleSlim)
+* PaddleInference Intel CPU部署量化模型[文档](https://paddle-inference.readthedocs.io/en/latest/optimize/paddle_x86_cpu_int8.html)
+* PaddleInference NV GPU部署量化模型[文档](https://paddle-inference.readthedocs.io/en/latest/optimize/paddle_trt.html)
\ No newline at end of file

paddle_inference/paddle/include/paddle_engine.h

@@ -57,9 +57,9 @@ PrecisionType GetPrecision(const std::string& precision_data) {
 }
 
 // Engine Base
-class PaddleEngineBase {
+class EngineCore {
  public:
-  virtual ~PaddleEngineBase() {}
+  virtual ~EngineCore() {}
   virtual std::vector<std::string> GetInputNames() {
     return _predictor->GetInputNames();
   }
@@ -107,7 +107,7 @@ class PaddleEngineBase {
 };
 
 // Paddle Inference Engine
-class PaddleInferenceEngine : public PaddleEngineBase {
+class PaddleInferenceEngine : public EngineCore {
  public:
   int create(const configure::EngineDesc& engine_conf) {
     std::string model_path = engine_conf.model_dir();

python/examples/fit_a_line/benchmark.py

@@ -25,19 +25,24 @@ args = benchmark_args()
 
 
 def single_func(idx, resource):
+    train_reader = paddle.batch(
+        paddle.reader.shuffle(
+            paddle.dataset.uci_housing.train(), buf_size=500),
+        batch_size=1)
+    total_number = sum(1 for _ in train_reader())
+
     if args.request == "rpc":
         client = Client()
         client.load_client_config(args.model)
         client.connect([args.endpoint])
-        train_reader = paddle.batch(
-            paddle.reader.shuffle(
-                paddle.dataset.uci_housing.train(), buf_size=500),
-            batch_size=1)
         start = time.time()
         for data in train_reader():
+            #new_data = np.zeros((1, 13)).astype("float32")
+            #new_data[0] = data[0][0]
+            #fetch_map = client.predict(feed={"x": new_data}, fetch=["price"], batch=True)
             fetch_map = client.predict(feed={"x": data[0][0]}, fetch=["price"])
         end = time.time()
-        return [[end - start]]
+        return [[end - start], [total_number]]
     elif args.request == "http":
         train_reader = paddle.batch(
             paddle.reader.shuffle(
@@ -49,7 +54,7 @@ def single_func(idx, resource):
                 'http://{}/uci/prediction'.format(args.endpoint),
                 data={"x": data[0]})
         end = time.time()
-        return [[end - start]]
+        return [[end - start], [total_number]]
 
 
 multi_thread_runner = MultiThreadRunner()

python/examples/xpu/bert/README.md

+
+## Prepare
+### convert model
+```
+python -m paddle_serving_client.convert --dirname infer_bert-base-chinese_ft_model_4000.pdparams
+```
+
+## RPC Service
+
+### Start Service
+
+```
+pytyon bert_web_service.py serving_server 7703
+```
+
+### Client Prediction
+
+```
+python bert_client.py
+```

python/examples/xpu/bert/bert_client.py

+# coding:utf-8
+# pylint: disable=doc-string-missing
+# Copyright (c) 2020 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.
+
+import sys
+from paddle_serving_client import Client
+from paddle_serving_client.utils import benchmark_args
+from chinese_bert_reader import ChineseBertReader
+import numpy as np
+args = benchmark_args()
+
+reader = ChineseBertReader({"max_seq_len": 128})
+fetch = ["save_infer_model/scale_0.tmp_1"]
+endpoint_list = ['127.0.0.1:7703']
+client = Client()
+client.load_client_config(args.model)
+client.connect(endpoint_list)
+
+for line in sys.stdin:
+    feed_dict = reader.process(line)
+    for key in feed_dict.keys():
+        feed_dict[key] = np.array(feed_dict[key]).reshape((128, 1))
+    #print(feed_dict)
+    result = client.predict(feed=feed_dict, fetch=fetch, batch=False)
+print(result)

python/examples/xpu/bert/bert_web_service.py

+# coding=utf-8
+# Copyright (c) 2020 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.
+# pylint: disable=doc-string-missing
+from paddle_serving_server.web_service import WebService
+from paddle_serving_app.reader import ChineseBertReader
+import sys
+import os
+import numpy as np
+
+
+class BertService(WebService):
+    def load(self):
+        self.reader = ChineseBertReader({
+            "vocab_file": "vocab.txt",
+            "max_seq_len": 128
+        })
+
+    def preprocess(self, feed=[], fetch=[]):
+        feed_res = []
+        is_batch = False
+        for ins in feed:
+            feed_dict = self.reader.process(ins["words"].encode("utf-8"))
+            for key in feed_dict.keys():
+                feed_dict[key] = np.array(feed_dict[key]).reshape(
+                    (len(feed_dict[key]), 1))
+            feed_res.append(feed_dict)
+        return feed_res, fetch, is_batch
+
+
+bert_service = BertService(name="bert")
+bert_service.load()
+bert_service.load_model_config(sys.argv[1])
+bert_service.prepare_server(
+    workdir="workdir", port=int(sys.argv[2]), use_lite=True, use_xpu=True, ir_optim=True)
+bert_service.run_rpc_service()
+bert_service.run_web_service()

python/examples/xpu/bert/chinese_bert_reader.py

+#   Copyright (c) 2020 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.
+# coding=utf-8
+from paddle_serving_app.reader.bert_base_reader import BertBaseReader
+from paddle_serving_app.reader.batching import pad_batch_data
+from paddle_serving_app.reader.tokenization import FullTokenizer, convert_to_unicode
+
+
+class ChineseBertReader(BertBaseReader):
+    """
+    ChineseBertReader handles the most traditional Chinese Bert
+    preprocessing, a user can define the vocab file through initialization
+
+    Examples:
+    from paddle_serving_app import ChineseBertReader
+
+    line = ["this is China"]
+    reader = ChineseBertReader()
+    reader.process(line[0])
+
+    """
+
+    def __init__(self, args={}):
+        super(ChineseBertReader, self).__init__()
+        vocab_file = ""
+        if "vocab_file" in args:
+            vocab_file = args["vocab_file"]
+        else:
+            vocab_file = self._download_or_not()
+
+        self.tokenizer = FullTokenizer(vocab_file=vocab_file)
+        if "max_seq_len" in args:
+            self.max_seq_len = args["max_seq_len"]
+        else:
+            self.max_seq_len = 20
+        self.vocab = self.tokenizer.vocab
+        self.pad_id = self.vocab["[PAD]"]
+        self.cls_id = self.vocab["[CLS]"]
+        self.sep_id = self.vocab["[SEP]"]
+        self.mask_id = self.vocab["[MASK]"]
+        self.feed_keys = [
+            "input_ids", "token_type_ids"
+        ]
+
+    """
+    inner function
+    """
+
+    def _download_or_not(self):
+        import os
+        import paddle_serving_app
+        module_path = os.path.dirname(paddle_serving_app.__file__)
+        full_path = "{}/tmp/chinese_bert".format(module_path)
+        os.system("mkdir -p {}".format(full_path))
+        if os.path.exists("{}/vocab.txt".format(full_path)):
+            pass
+        else:
+            url = "https://paddle-serving.bj.bcebos.com/reader/chinese_bert/vocab.txt"
+            r = os.system("wget --no-check-certificate " + url)
+            os.system("mv vocab.txt {}".format(full_path))
+            if r != 0:
+                raise SystemExit('Download failed, please check your network')
+        return "{}/vocab.txt".format(full_path)
+
+    """
+    inner function
+    """
+
+    def _pad_batch(self, token_ids, text_type_ids):
+        batch_token_ids = [token_ids]
+        batch_text_type_ids = [text_type_ids]
+
+        padded_token_ids, input_mask = pad_batch_data(
+            batch_token_ids,
+            max_seq_len=self.max_seq_len,
+            pad_idx=self.pad_id,
+            return_input_mask=True)
+        padded_text_type_ids = pad_batch_data(
+            batch_text_type_ids,
+            max_seq_len=self.max_seq_len,
+            pad_idx=self.pad_id)
+        return padded_token_ids,  padded_text_type_ids
+
+    """
+    process function deals with a raw Chinese string as a sentence
+    this funtion returns a feed_dict
+    default key of the returned feed_dict: input_ids, position_ids, segment_ids, input_mask
+    """
+
+    def process(self, line):
+        text_a = convert_to_unicode(line)
+        tokens_a = self.tokenizer.tokenize(text_a)
+        if len(tokens_a) > self.max_seq_len - 2:
+            tokens_a = tokens_a[0:(self.max_seq_len - 2)]
+        tokens = []
+        text_type_ids = []
+        tokens.append("[CLS]")
+        text_type_ids.append(0)
+        for token in tokens_a:
+            tokens.append(token)
+            text_type_ids.append(0)
+        token_ids = self.tokenizer.convert_tokens_to_ids(tokens)
+        #position_ids = list(range(len(token_ids)))
+        p_token_ids, p_text_type_ids= \
+            self._pad_batch(token_ids, text_type_ids)
+        feed_result = {
+            self.feed_keys[0]: p_token_ids.reshape(-1).tolist(),
+            #self.feed_keys[1]: p_pos_ids.reshape(-1).tolist(),
+            self.feed_keys[1]: p_text_type_ids.reshape(-1).tolist(),
+            #self.feed_keys[3]: input_mask.reshape(-1).tolist()
+        }
+        return feed_result

python/examples/xpu/ernie/README.md

+
+## Prepare
+### convert model
+```
+python3 -m paddle_serving_client.convert --dirname erine
+```
+
+## RPC Service
+
+### Start Service
+
+```
+python3 -m paddle_serving_server.serve --model serving_server --port 7704 --use_lite --use_xpu --ir_optim
+```
+
+### Client Prediction
+
+```
+head data-c.txt | python3 ernie_client.py --model serving_client/serving_client_conf.prototxt
+```

python/examples/xpu/ernie/chinese_ernie_reader.py

+#   Copyright (c) 2020 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.
+# coding=utf-8
+from paddle_serving_app.reader.bert_base_reader import BertBaseReader
+from paddle_serving_app.reader.batching import pad_batch_data
+from paddle_serving_app.reader.tokenization import FullTokenizer, convert_to_unicode
+
+
+class ChineseErnieReader(BertBaseReader):
+    """
+    ChineseErnieReader handles the most traditional Chinese Bert
+    preprocessing, a user can define the vocab file through initialization
+
+    Examples:
+    from paddle_serving_app import ChineseErnieReader
+
+    line = ["this is China"]
+    reader = ChineseErnieReader()
+    reader.process(line[0])
+
+    """
+
+    def __init__(self, args={}):
+        super(ChineseErnieReader, self).__init__()
+        vocab_file = ""
+        if "vocab_file" in args:
+            vocab_file = args["vocab_file"]
+            print("vocab")
+        else:
+            vocab_file = self._download_or_not()
+
+        self.tokenizer = FullTokenizer(vocab_file=vocab_file)
+        print(self.tokenizer)
+        if "max_seq_len" in args:
+            self.max_seq_len = args["max_seq_len"]
+        else:
+            self.max_seq_len = 20
+        self.vocab = self.tokenizer.vocab
+        self.pad_id = self.vocab["[PAD]"]
+        self.cls_id = self.vocab["[CLS]"]
+        self.sep_id = self.vocab["[SEP]"]
+        self.mask_id = self.vocab["[MASK]"]
+        self.feed_keys = [
+            "placeholder_0", "placeholder_1", "placeholder_2", "placeholder_3"
+        ]
+
+    """
+    inner function
+    """
+
+    def _download_or_not(self):
+        import os
+        import paddle_serving_app
+        module_path = os.path.dirname(paddle_serving_app.__file__)
+        full_path = "{}/tmp/chinese_bert".format(module_path)
+        os.system("mkdir -p {}".format(full_path))
+        if os.path.exists("{}/vocab.txt".format(full_path)):
+            pass
+        else:
+            url = "https://paddle-serving.bj.bcebos.com/reader/chinese_bert/vocab.txt"
+            r = os.system("wget --no-check-certificate " + url)
+            os.system("mv vocab.txt {}".format(full_path))
+            if r != 0:
+                raise SystemExit('Download failed, please check your network')
+        return "{}/vocab.txt".format(full_path)
+
+    """
+    inner function
+    """
+
+    def _pad_batch(self, token_ids, text_type_ids, position_ids):
+        batch_token_ids = [token_ids]
+        batch_text_type_ids = [text_type_ids]
+        batch_position_ids = [position_ids]
+
+        padded_token_ids, input_mask = pad_batch_data(
+            batch_token_ids,
+            max_seq_len=self.max_seq_len,
+            pad_idx=self.pad_id,
+            return_input_mask=True)
+        padded_text_type_ids = pad_batch_data(
+            batch_text_type_ids,
+            max_seq_len=self.max_seq_len,
+            pad_idx=self.pad_id)
+        padded_position_ids = pad_batch_data(
+            batch_position_ids,
+            max_seq_len=self.max_seq_len,
+            pad_idx=self.pad_id)
+        return padded_token_ids, padded_position_ids, padded_text_type_ids, input_mask
+
+    """
+    process function deals with a raw Chinese string as a sentence
+    this funtion returns a feed_dict
+    default key of the returned feed_dict: input_ids, position_ids, segment_ids, input_mask
+    """
+
+    def process(self, line):
+        text_a = convert_to_unicode(line)
+        tokens_a = self.tokenizer.tokenize(text_a)
+        if len(tokens_a) > self.max_seq_len - 2:
+            tokens_a = tokens_a[0:(self.max_seq_len - 2)]
+        tokens = []
+        text_type_ids = []
+        tokens.append("[CLS]")
+        text_type_ids.append(0)
+        for token in tokens_a:
+            tokens.append(token)
+            text_type_ids.append(0)
+        token_ids = self.tokenizer.convert_tokens_to_ids(tokens)
+        position_ids = list(range(len(token_ids)))
+        p_token_ids, p_pos_ids, p_text_type_ids, input_mask = \
+            self._pad_batch(token_ids, text_type_ids, position_ids)
+        feed_result = {
+            self.feed_keys[0]: p_token_ids.reshape(-1).tolist(),
+            self.feed_keys[1]: p_pos_ids.reshape(-1).tolist(),
+            self.feed_keys[2]: p_text_type_ids.reshape(-1).tolist(),
+            self.feed_keys[3]: input_mask.reshape(-1).tolist()
+        }
+        return feed_result

python/examples/xpu/ernie/erine_web_service.py

+# coding=utf-8
+# Copyright (c) 2020 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.
+# pylint: disable=doc-string-missing
+from paddle_serving_server.web_service import WebService
+from paddle_serving_app.reader import ChineseBertReader
+import sys
+import os
+import numpy as np
+
+
+class BertService(WebService):
+    def load(self):
+        self.reader = ChineseBertReader({
+            "vocab_file": "vocab.txt",
+            "max_seq_len": 128
+        })
+
+    def preprocess(self, feed=[], fetch=[]):
+        feed_res = []
+        is_batch = False
+        for ins in feed:
+            feed_dict = self.reader.process(ins["words"].encode("utf-8"))
+            for key in feed_dict.keys():
+                feed_dict[key] = np.array(feed_dict[key]).reshape(
+                    (len(feed_dict[key]), 1))
+            feed_res.append(feed_dict)
+        return feed_res, fetch, is_batch
+
+
+bert_service = BertService(name="bert")
+bert_service.load()
+bert_service.load_model_config(sys.argv[1])
+bert_service.prepare_server(
+    workdir="workdir", port=int(sys.argv[2]), use_lite=True, use_xpu=True, ir_optim=True)
+bert_service.run_rpc_service()
+bert_service.run_web_service()

python/examples/xpu/ernie/ernie_client.py

+# coding:utf-8
+# pylint: disable=doc-string-missing
+# Copyright (c) 2020 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.
+
+import sys
+from paddle_serving_client import Client
+from paddle_serving_client.utils import benchmark_args
+from chinese_ernie_reader import ChineseErnieReader
+import numpy as np
+args = benchmark_args()
+
+reader = ChineseErnieReader({"max_seq_len": 128})
+fetch = ["save_infer_model/scale_0"]
+endpoint_list = ['127.0.0.1:7704']
+client = Client()
+client.load_client_config(args.model)
+client.connect(endpoint_list)
+
+for line in sys.stdin:
+    feed_dict = reader.process(line)
+    for key in feed_dict.keys():
+        feed_dict[key] = np.array(feed_dict[key]).reshape((128, 1))
+   # print(feed_dict)
+    result = client.predict(feed=feed_dict, fetch=fetch, batch=False)
+    print(result)

python/examples/xpu/vgg19/README.md

+
+## Prepare
+### convert model
+```
+python -m paddle_serving_client.convert --dirname VGG19
+```
+
+## RPC Service
+
+### Start Service
+
+```
+python -m paddle_serving_server.serve --model serving_server --port 7702 --use_lite --use_xpu --ir_optim
+```
+
+### Client Prediction
+
+```
+python vgg19_client.py
+```

python/examples/xpu/vgg19/vgg19_client.py

+# Copyright (c) 2020 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.
+
+from paddle_serving_client import Client
+from paddle_serving_app.reader import Sequential, File2Image, Resize, CenterCrop
+from paddle_serving_app.reader import RGB2BGR, Transpose, Div, Normalize
+
+client = Client()
+client.load_client_config(
+    "serving_client/serving_client_conf.prototxt")
+client.connect(["127.0.0.1:7702"])
+
+seq = Sequential([
+    File2Image(), Resize(256), CenterCrop(224), RGB2BGR(), Transpose((2, 0, 1)),
+    Div(255), Normalize([0.485, 0.456, 0.406], [0.229, 0.224, 0.225], True)
+])
+
+image_file = "daisy.jpg"
+img = seq(image_file)
+fetch_map = client.predict(feed={"image": img}, fetch=["save_infer_model/scale_0"])
+#print(fetch_map)
+print(fetch_map["save_infer_model/scale_0"].reshape(-1))

python/gen_version.py

@@ -47,3 +47,5 @@ elif package_name.endswith('xpu'):
 path = "paddle_serving_" + sys.argv[1]
 commit_id = subprocess.check_output(['git', 'rev-parse', 'HEAD'])
 update_info(path + "/version.py", "commit_id", commit_id)
+
+update_info(path + "/version.py", "version_tag", "${VERSION_TAG}")

python/paddle_serving_app/__init__.py

@@ -12,3 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 from .models import ServingModels
+from . import version
+
+__version__ = version.version_tag

python/paddle_serving_app/version.py

@@ -12,5 +12,5 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ Paddle Serving App version string """
-serving_app_version = "0.0.0"
+version_tag = "0.0.0"
 commit_id = ""

python/paddle_serving_client/__init__.py

@@ -17,4 +17,4 @@ from . import version
 from . import client
 from .client import *
 
-__version__ = version.serving_client_version
+__version__ = version.version_tag

python/paddle_serving_client/client.py

@@ -155,7 +155,7 @@ class Client(object):
         file_path_list = []
         for single_model_config in model_config_path_list:
             if os.path.isdir(single_model_config):
-                file_path_list.append("{}/serving_server_conf.prototxt".format(
+                file_path_list.append("{}/serving_client_conf.prototxt".format(
                     single_model_config))
             elif os.path.isfile(single_model_config):
                 file_path_list.append(single_model_config)
@@ -574,7 +574,7 @@ class MultiLangClient(object):
         file_path_list = []
         for single_model_config in model_config_path_list:
             if os.path.isdir(single_model_config):
-                file_path_list.append("{}/serving_server_conf.prototxt".format(
+                file_path_list.append("{}/serving_client_conf.prototxt".format(
                     single_model_config))
             elif os.path.isfile(single_model_config):
                 file_path_list.append(single_model_config)

python/paddle_serving_client/version.py

@@ -12,7 +12,6 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ Paddle Serving Client version string """
-serving_client_version = "0.0.0"
-serving_server_version = "0.0.0"
+version_tag = "0.0.0"
 module_proto_version = "0.0.0"
 commit_id = ""

python/paddle_serving_server/__init__.py

@@ -31,4 +31,4 @@ from paddle_serving_server import (
 from .dag import *
 from .server import *
 
-__version__ = version.serving_server_version
+__version__ = version.version_tag

python/paddle_serving_server/rpc_service.py

+# Copyright (c) 2021 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.
+
 import sys
 import os
-
+import numpy as np
 import google.protobuf.text_format
 
 from .proto import general_model_config_pb2 as m_config
@@ -9,6 +23,7 @@ sys.path.append(
     os.path.join(os.path.abspath(os.path.dirname(__file__)), 'proto'))
 from .proto import multi_lang_general_model_service_pb2_grpc
 
+
 class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
                                      MultiLangGeneralModelServiceServicer):
     def __init__(self, model_config_path_list, is_multi_model, endpoints):
@@ -31,7 +46,7 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
             model_config_path_list = [model_config_path_list]
         elif isinstance(model_config_path_list, list):
             pass
-        
+
         file_path_list = []
         for single_model_config in model_config_path_list:
             if os.path.isdir(single_model_config):
@@ -57,7 +72,7 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
             f = open(file_path_list[-1], 'r')
             model_conf = google.protobuf.text_format.Merge(
                 str(f.read()), model_conf)
-        
+
         self.fetch_names_ = [var.alias_name for var in model_conf.fetch_var]
         self.fetch_types_ = {}
         for i, var in enumerate(model_conf.fetch_var):
@@ -86,11 +101,11 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
                 v_type = self.feed_types_[name]
                 data = None
                 if is_python:
-                    if v_type == 0:# int64
+                    if v_type == 0:  # int64
                         data = np.frombuffer(var.data, dtype="int64")
-                    elif v_type == 1:# float32
+                    elif v_type == 1:  # float32
                         data = np.frombuffer(var.data, dtype="float32")
-                    elif v_type == 2:# int32
+                    elif v_type == 2:  # int32
                         data = np.frombuffer(var.data, dtype="int32")
                     else:
                         raise Exception("error type.")
@@ -99,7 +114,7 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
                         data = np.array(list(var.int64_data), dtype="int64")
                     elif v_type == 1:  # float32
                         data = np.array(list(var.float_data), dtype="float32")
-                    elif v_type == 2:# int32
+                    elif v_type == 2:  # int32
                         data = np.array(list(var.int_data), dtype="int32")
                     else:
                         raise Exception("error type.")
@@ -155,7 +170,8 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
         # This porcess and Inference process cannot be operate at the same time.
         # For performance reasons, do not add thread lock temporarily.
         timeout_ms = request.timeout_ms
-        self._init_bclient(self.model_config_path_list, self.endpoints_, timeout_ms)
+        self._init_bclient(self.model_config_path_list, self.endpoints_,
+                           timeout_ms)
         resp = multi_lang_general_model_service_pb2.SimpleResponse()
         resp.err_code = 0
         return resp
@@ -176,4 +192,4 @@ class MultiLangServerServiceServicer(multi_lang_general_model_service_pb2_grpc.
         #dict should be added when graphMaker is used.
         resp = multi_lang_general_model_service_pb2.GetClientConfigResponse()
         resp.client_config_str_list[:] = self.model_config_path_list
-        return resp
\ No newline at end of file
+        return resp

python/paddle_serving_server/serve.py

@@ -103,7 +103,7 @@ def serve_args():
 
 
 def start_standard_model(serving_port):  # pylint: disable=doc-string-missing
-    args = parse_args()
+    args = serve_args()
     thread_num = args.thread
     model = args.model
     port = serving_port
@@ -410,6 +410,7 @@ if __name__ == "__main__":
             use_lite=args.use_lite,
             use_xpu=args.use_xpu,
             ir_optim=args.ir_optim,
+            thread_num=args.thread,
             precision=args.precision,
             use_calib=args.use_calib)
         web_service.run_rpc_service()

python/paddle_serving_server/server.py

@@ -22,7 +22,7 @@ from .proto import general_model_config_pb2 as m_config
 from .proto import multi_lang_general_model_service_pb2_grpc
 import google.protobuf.text_format
 import time
-from .version import serving_server_version, version_suffix, device_type
+from .version import version_tag, version_suffix, device_type
 from contextlib import closing
 import argparse
 
@@ -369,7 +369,7 @@ class Server(object):
         version_file = open("{}/version.py".format(self.module_path), "r")
 
         folder_name = "serving-%s-%s" % (self.get_serving_bin_name(),
-                                         serving_server_version)
+                                         version_tag)
         tar_name = "%s.tar.gz" % folder_name
         bin_url = "https://paddle-serving.bj.bcebos.com/bin/%s" % tar_name
 

python/paddle_serving_server/version.py

@@ -12,10 +12,9 @@
 # See the License for the specific language governing permissions and
 # limitations under the License.
 """ Paddle Serving Server version string """
-serving_client_version = "0.0.0"
-serving_server_version = "0.0.0"
-module_proto_version = "0.0.0"
+version_tag = "0.0.0"
 version_suffix = ""
+module_proto_version = "0.0.0"
 device_type = "0"
 cuda_version = "9"
 commit_id = ""

python/paddle_serving_server/web_service.py

@@ -101,7 +101,7 @@ class WebService(object):
 
         self.fetch_vars = {var.name: var for var in model_conf.fetch_var}
         if client_config_path == None:
-            self.client_config_path = self.server_config_dir_paths
+            self.client_config_path = file_path_list
 
     def set_gpus(self, gpus):
         print("This API will be deprecated later. Please do not use it")
@@ -176,10 +176,12 @@ class WebService(object):
                        use_xpu=False,
                        ir_optim=False,
                        gpuid=0,
+                       thread_num=2,
                        mem_optim=True):
         print("This API will be deprecated later. Please do not use it")
         self.workdir = workdir
         self.port = port
+        self.thread_num = thread_num
         self.device = device
         self.gpuid = gpuid
         self.port_list = []
@@ -197,7 +199,7 @@ class WebService(object):
                     self.workdir,
                     self.port_list[0],
                     -1,
-                    thread_num=2,
+                    thread_num=self.thread_num,
                     mem_optim=mem_optim,
                     use_lite=use_lite,
                     use_xpu=use_xpu,
@@ -211,7 +213,7 @@ class WebService(object):
                         "{}_{}".format(self.workdir, i),
                         self.port_list[i],
                         gpuid,
-                        thread_num=2,
+                        thread_num=self.thread_num,
                         mem_optim=mem_optim,
                         use_lite=use_lite,
                         use_xpu=use_xpu,

python/setup.py.app.in

@@ -22,7 +22,7 @@ import os
 from setuptools import setup, Distribution, Extension
 from setuptools import find_packages
 from setuptools import setup
-from paddle_serving_app.version import serving_app_version
+from paddle_serving_app.version import version_tag
 from pkg_resources import DistributionNotFound, get_distribution
         
 def python_version():
@@ -78,7 +78,7 @@ package_dir={'paddle_serving_app':
 
 setup(
     name='paddle-serving-app',
-    version=serving_app_version.replace('-', ''),
+    version=version_tag.replace('-', ''),
     description=
     ('Paddle Serving Package for saved model with PaddlePaddle'),
     url='https://github.com/PaddlePaddle/Serving',

python/setup.py.client.in

@@ -22,7 +22,7 @@ import sys
 from setuptools import setup, Distribution, Extension
 from setuptools import find_packages
 from setuptools import setup
-from paddle_serving_client.version import serving_client_version
+from paddle_serving_client.version import version_tag
 import util
 
 py_version = sys.version_info
@@ -79,7 +79,7 @@ package_dir={'paddle_serving_client':
 
 setup(
     name='paddle-serving-client',
-    version=serving_client_version.replace('-', ''),
+    version=version_tag.replace('-', ''),
     description=
     ('Paddle Serving Package for saved model with PaddlePaddle'),
     url='https://github.com/PaddlePaddle/Serving',

python/setup.py.in

@@ -19,7 +19,7 @@ from __future__ import print_function
 from setuptools import setup, Distribution, Extension
 from setuptools import find_packages
 from setuptools import setup
-from paddle_serving.version import serving_client_version
+from paddle_serving.version import version_tag
 from grpc_tools import protoc
 import util
 
@@ -43,7 +43,7 @@ package_dir={'paddle_serving.serving_client':
 
 setup(
     name='paddle-serving-client',
-    version=serving_client_version.replace('-', ''),
+    version=version_tag.replace('-', ''),
     description=
     ('Paddle Serving Package for saved model with PaddlePaddle'),
     url='https://github.com/PaddlePaddle/Serving',

python/setup.py.server.in

@@ -19,10 +19,10 @@ from __future__ import print_function
 from setuptools import setup, Distribution, Extension
 from setuptools import find_packages
 from setuptools import setup
-from paddle_serving_server.version import serving_server_version, version_suffix
+from paddle_serving_server.version import version_tag, version_suffix
 import util
 
-package_version = serving_server_version.replace('-', '')
+package_version = version_tag.replace('-', '')
 if version_suffix != "":
     version_suffix = "post" + version_suffix
     package_version = package_version + "." + version_suffix