您好，我们已经收到了您的问题，会安排技术人员尽快解答您的问题，请耐心等待。请您再次检查是否提供了清晰的问题描述、复现代码、环境&版本、报错信息等。同时，您也可以通过查看官网API文档、常见问题、历史Issue、AI社区来寻求解答。祝您生活愉快～

Hi! We've received your issue and please be patient to get responded. We will arrange technicians to answer your questions as soon as possible. Please make sure that you have posted enough message to demo your request. You may also check out the API，FAQ，Github Issue and AI community to get the answer.Have a nice day!

您好，TryShrinkMemory是释放临时变量的api，释放模型是在m_Predictor的析构函数。

我的测试代码如下：
#include "paddle/include/paddle_inference_api.h"
#include
#include <gflags/gflags.h>
#include <glog/logging.h>
#include
#include
#include
#include
using namespace std;

namespace paddle_infer
{

void PrepareConfig(Config* config, int num) 
{
    config->EnableUseGpu(100, 0);
    config->DisableGlogInfo();
    config->EnableMemoryOptim();
    config->EnableGpuMultiStream();
    config->EnableTensorRtEngine(1 << 30, 4096, 3, paddle::AnalysisConfig::Precision::kFloat32, true, false);
    if (num == 0)
    {
        config->SetModel("F:/L1/MobileNetV3_small_x0_35/inference.pdmodel", "F:/L1/MobileNetV3_small_x0_35/inference.pdiparams");
        config->SetOptimCacheDir("F:/L1/MobileNetV3_small_x0_35/");
    }

    if (num == 1)
    {
        config->SetModel("F:/L2/MobileNetV3_small_x0_35/inference.pdmodel", "F:/L2/MobileNetV3_small_x0_35/inference.pdiparams");
        config->SetOptimCacheDir("F:/L2/MobileNetV3_small_x0_35/");
    }

    if (num == 2)
    {
        config->SetModel("F:/L4/MobileNetV3_small_x0_35/inference.pdmodel", "F:/L4/MobileNetV3_small_x0_35/inference.pdiparams");
        config->SetOptimCacheDir("F:/L4/MobileNetV3_small_x0_35/");
    }
}

void Run(std::shared_ptr<Predictor> predictor, Config* config)
{
    int batchsize = 4096;

    // 准备输入数据
    std::vector<int>   input_shape = { batchsize, 1, 124, 84 };
    std::vector<float> input_data(batchsize * 1 * 124 * 84, 1.0);
    std::vector<float> out_data;
    vector<double>     counts;
    for (size_t i = 0; i < 200; i++)
    {
        auto curTime      = std::chrono::steady_clock::now();
        int  input_num    = std::accumulate(input_shape.begin(), input_shape.end(), 1, std::multiplies<int>());
        auto input_names  = predictor->GetInputNames();
        auto input_tensor = predictor->GetInputHandle(input_names[0]);
        input_tensor->Reshape(input_shape);
        input_tensor->CopyFromCpu(input_data.data());

        // 执行预测
        predictor->Run();

        // 获取预测输出
        auto output_names  = predictor->GetOutputNames();
        auto output_tensor = predictor->GetOutputHandle(output_names[0]);
        std::vector<int> output_shape = output_tensor->shape();
        int  out_num = std::accumulate(output_shape.begin(), output_shape.end(), 1, std::multiplies<int>());

        out_data.resize(out_num);
        output_tensor->CopyToCpu(out_data.data());

        auto   endTime = std::chrono::steady_clock::now();
        if (i > 10)
        {
            counts.push_back(std::chrono::duration<double, std::milli>(endTime - curTime).count());
        }

        if (i % 100 == 0)
        {
            int    gpu_id = config->gpu_device_id();
            string param_files = config->params_file();
            cout << "gpu id = " << gpu_id << ", " << "param  = " << param_files << ", " << "idx    = " << i << endl;
        }
    }

    double cout_ave = accumulate(counts.begin(), counts.end(), 0) / counts.size();
    std::cout << "bachsize = " << batchsize << ", gpu id = " << config->gpu_device_id() << " avg classify img time : " << cout_ave / batchsize << "ms!" << endl;
}

}

int main(int argc, char**argv)
{
paddle_infer::Config config[3];
std::shared_ptr<paddle_infer::Predictor> predictor[3];
for (size_t i = 0; i < 3; i++)
{
paddle_infer::PrepareConfig(&config[i], i);
predictor[i] = paddle_infer::CreatePredictor(config[i]);
}

std::cout << "start running" << endl;
auto curTime = std::chrono::steady_clock::now();
std::vector<std::thread> threads;
for (int i = 0; i < 3; ++i) 
{
    threads.emplace_back(paddle_infer::Run, predictor[i], &config[i]);
}

for (int i = 0; i < 3; ++i) 
{
    threads[i].join();
}

auto endTime  = std::chrono::steady_clock::now();
auto lastTime = std::chrono::duration<double, std::milli>(endTime - curTime).count();
std::cout << "end runing, ave time = " << lastTime / (4096 * 600) << "ms!" << endl;

}

但是运行完毕后出现以下错误：

C++ Traceback (most recent call last):

Not support stack backtrace yet.

Error Message Summary:

ExternalError: Cuda error(4), driver shutting down.
[Advise: Please search for the error code(4) on website( https://docs.nvidia.com/cuda/archive/9.0/cuda-runtime-api/group__CUDART__TYPES.html#group__CUDART__TYPES_1g3f51e3575c2178246db0a94a430e0038 ) to get Nvidia's official solution about CUDA Error.] (at C:\home\workspace\Paddle_release5\paddle\fluid\platform\gpu_info.cc:275)

麻烦帮忙看下怎么处理呢？

您好，有一个这个api的测试案例，可以看一下

https://github.com/PaddlePaddle/Paddle-Inference-Demo/tree/master/c%2B%2B/test/shrink_memory

https://github.com/PaddlePaddle/Paddle-Inference-Demo/blob/master/c%2B%2B/test/shrink_memory/multi_thread_test.cc
跑多线程示例可以运行么？
另外，您的环境下，单线程可以正常析构么