TensorRT 10.5.0 -- CPU Memory leak while using nvinfer1::createInferBuilder on 4060

[Iridium771110]

---

Description

when I using nvinfer::IBuilder = nvinfer1::createInferBuilder (logger) to create a model engine, it seems a CPU Memory leak.
the full code is shown below.

#include <iostream>
#include <string>
#include <cstring>
#include <NvInfer.h>

class Logger : public nvinfer1::ILogger{
    void log(Severity severity, const char* msg) noexcept override{
        if (severity <= Severity::kWARNING) std::cout<<msg<<std::endl;
    }
};

size_t physical_memory_used_by_process(){
    FILE* file = fopen("/proc/self/status", "r");
    int result = -1;
    char line[128];
    while (fgets(line, 128, file) != nullptr) {
        if (strncmp(line, "VmRSS:", 6) == 0) {
            int len = strlen(line);
            const char* p = line;
            for (; std::isdigit(*p) == false; ++p) {}
            line[len - 3] = 0;
            result = atoi(p);
            break;
        }
    }
    fclose(file);
    return result;
}

int main(int argc, char *argv[]){
    size_t gpu_total_byte_mem;
    size_t gpu_free_byte_mem;
    size_t gpu_alloced_byte_mem;
    cudaDeviceSynchronize();
    cudaMemGetInfo(&gpu_free_byte_mem, &gpu_total_byte_mem);
    gpu_alloced_byte_mem = gpu_total_byte_mem - gpu_free_byte_mem;
    std::cout<<"init status"<<std::endl;
    std::cout<<"used gpu mem: "<<double(gpu_alloced_byte_mem) / 1024.0 / 1024.0<<"MB"<<std::endl;
    std::cout<<"used cpu mem: "<<double(physical_memory_used_by_process()) / 1024.0 <<"MB"<<std::endl;

    Logger logger;

    cudaDeviceSynchronize();
    cudaMemGetInfo(&gpu_free_byte_mem, &gpu_total_byte_mem);
    gpu_alloced_byte_mem = gpu_total_byte_mem - gpu_free_byte_mem;
    std::cout<<"logger created"<<std::endl;
    std::cout<<"used gpu mem: "<<double(gpu_alloced_byte_mem) / 1024.0 / 1024.0<<"MB"<<std::endl;
    std::cout<<"used cpu mem: "<<double(physical_memory_used_by_process()) / 1024.0 <<"MB"<<std::endl;

    nvinfer1::IBuilder* builder = nvinfer1::createInferBuilder(logger);

    cudaDeviceSynchronize();
    cudaMemGetInfo(&gpu_free_byte_mem, &gpu_total_byte_mem);
    gpu_alloced_byte_mem = gpu_total_byte_mem - gpu_free_byte_mem;
    std::cout<<"builder created"<<std::endl;
    std::cout<<"used gpu mem: "<<double(gpu_alloced_byte_mem) / 1024.0 / 1024.0<<"MB"<<std::endl;
    std::cout<<"used cpu mem: "<<double(physical_memory_used_by_process()) / 1024.0 <<"MB"<<std::endl;

    delete builder;

    cudaDeviceSynchronize();
    cudaMemGetInfo(&gpu_free_byte_mem, &gpu_total_byte_mem);
    gpu_alloced_byte_mem = gpu_total_byte_mem - gpu_free_byte_mem;
    std::cout<<"builder deleted"<<std::endl;
    std::cout<<"used gpu mem: "<<double(gpu_alloced_byte_mem) / 1024.0 / 1024.0<<"MB"<<std::endl;
    std::cout<<"used cpu mem: "<<double(physical_memory_used_by_process()) / 1024.0 <<"MB"<<std::endl;

    return 0;
}

build the code and run, I see the memory printed as

so it seems like a CPU memory leak by createInferBuilder API or IBuilder pointer.
do you have any suggestion here? thanks.

Environment

TensorRT Version: 10.5.0

NVIDIA GPU: 4060

NVIDIA Driver Version: 560.35.03

CUDA Version: 12.4

CUDNN Version: --

Operating System:

Python Version (if applicable):

Tensorflow Version (if applicable):

PyTorch Version (if applicable):

Baremetal or Container (if so, version):

Steps To Reproduce

just build the code and run it.

[lix19937]

You had better to use Valgrind to detect memory management .

[asfiyab-nvidia]

@zhenhuaw-me can you please look into this?

[zhenhuaw-me]

@Iridium771110 Thank you for reporting this issue. Some CUDA resources such as kernels are released when the process terminated. So the measurement is inaccurate.

As @lix19937 , if you think there is a memory leak, please use valgrind to detect and confirm.

I am going to close this ticket since it's a false alarm. Feel free to reopen if any further concerns.