Use libtorch 1.10.2+cu102 vs libtorch 1.9.0+cu102

C++
1

I try to build a conan package for libtorch in ubuntu 18.04, Environment is:
NVIDIA-SMI 460.32.03 Driver Version: 460.32.03 CUDA Version: 11.2 GeForce RTX 2080

test_package.cpp:

#include <torch/cuda.h>
#include <iostream>

int main(void) {
  std::cout << "CUDA DEVICE COUNT: " << torch::cuda::device_count() << std::endl;
  if (torch::cuda::is_available()) {
    std::cout << "CUDA is available! Training on GPU." << std::endl;
  }
  return 0;
}

When I test libtorch 1.10.2+cu102， it returns

libtorch/1.10.2+cu102 (test package): Running test()
CUDA DEVICE COUNT: 1
CUDA is available! Training on GPU.

But when I test libtorch 1.9.0+cu102(or libtorch 1.9.1+cu102), it returns:

libtorch/1.9.1+cu102 (test package): Running test()
CUDA DEVICE COUNT: 0

how does this happen? I can’t figure out why.

2

BTW, Python version is 3.6.9