• 版本、环境信息：

   1）PaddlePaddle版本：1.3
   2）CPU：预测若用CPU，请提供CPU型号，MKL/OpenBlas/MKLDNN/等数学库使用情况
   3）GPU：V100
   4）系统环境：Ubuntu 5.4.0-6ubuntu1~16.04.5, Python 3.7.1
-预测信息
   1）C++预测：预测库安装包的version.txt文件

GIT COMMIT ID: 4b3f9e5c61d687ea90e6599bf9494df92ed088fb
      WITH_MKL: ON
      WITH_MKLDNN: ON
      WITH_GPU: ON
      CUDA version: 9.0
      CUDNN version: v7

 2）CMakeLists.txt

MAKE_MINIMUM_REQUIRED(VERSION 3.2)
PROJECT(inference_demo)

add_subdirectory(pybind11)
pybind11_add_module(example example.cpp)

SET(CMAKE_C_COMPILER gcc)
SET(CMAKE_CXX_COMPILER g++)
ADD_COMPILE_OPTIONS(-std=c++11 -D_GLIBCXX_USE_CXX11_ABI=0)

SET(FLUID_INFER_LIB ${CMAKE_SOURCE_DIR}/fluid_inference)
SET(FLUID_INC_PATH ${FLUID_INFER_LIB}/paddle/include)
SET(FLUID_LIB_PATH ${FLUID_INFER_LIB}/paddle/lib)

INCLUDE_DIRECTORIES(${FLUID_INC_PATH})

LINK_DIRECTORIES(${FLUID_LIB_PATH})

TARGET_LINK_LIBRARIES(example PRIVATE absolute/path/to/libpaddle_fluid.so)

 3）API信息 CreatePaddlePredictor
   4）预测库来源：官网下载
5) 代码

# include <paddle_inference_api.h>

# include <unistd.h>

# include <iostream>

# include <string>

# include <vector>

# include <pybind11/pybind11.h>

namespace py = pybind11;

void CreateErnieInstance() {
  std::string model_dir = "/path/to/model/dir";
  bool use_gpu = true;
  const char *file_path = model_dir.c_str();
  if (access(file_path, F_OK) == 0){
       std::cout << "model dir exists!" << std::endl;
  }

  paddle::NativeConfig config;
  config.model_dir = model_dir;
  if (use_gpu) {
      config.use_gpu = true;
      config.fraction_of_gpu_memory = 0.15;
      config.device = 0;
  }

  std::cout << "before creating predictor" << std::endl;
  auto predictor = CreatePaddlePredictor(config);
  std::cout << "after creating predictor" << std::endl;
}

PYBIND11_MODULE(example, m) {
   m.doc() = "ernie inference core";
   m.def("create_ernie_instance", &CreateErnieInstance, "A function which creates ernie inference instance");
}

• 报错信息:

执行到"auto predictor = CreatePaddlePredictor(config);"时候会报错: Segmentation fault (core dumped)

1. 报错复现步骤

2. 下载fluid_inference库到example.cpp与CMakeLists.txt的路径

3. git clone pybind11到example.cpp的路径

4. 修改CMakeLists.txt最后一行的最后一个参数为libpaddle_fluid.so的绝对路径

5. 编译完成后执行python -c "import example; example.create_ernie_instance()"

6. 问题描述：
   项目背景是部署ERNIE的inference model在服务器，之前用python的load_inference_model api发现显存占用和training差不多, 都是十几个G，而用C++基于CreatePaddlePredictor API做inference的时候显存只占用3个G。所以尝试用pybind11封装CreatePaddlePredictor函数，以便在python中调用，然后就出现了以上报错信息。请问这个问题是什么原因造成的？另外如果已经有python的api是基于CreatePaddlePredictor实现的也烦请告知。谢谢！