LibTorch symbol lookup error in CPU

C++
1

I try LibTorch in CPU.
version 1: use cxx11 ABI

cmake -DCMAKE_PREFIX_PATH=/tools/libtorch-cxx11-abi-shared-with-deps-1.13.0+cpu ..
cmake --build . --config Release

version 2: use Pre-cxx11 ABI

cmake -DCMAKE_PREFIX_PATH=/tools/libtorch-shared-with-deps-1.13.0+cpu ..
cmake --build . --config Release

Both can build successfully, but when to run the release executable program come same error:

./test-model: symbol lookup error: ./test-model: undefined symbol: _ZN2at4_ops4ones4callEN3c108ArrayRefINS2_6SymIntEEENS2_8optionalINS2_10ScalarTypeEEENS6_INS2_6LayoutEEENS6_INS2_6DeviceEEENS6_IbEE

However when switch the DCMAKE_PREFIX_PATH to Pytorch install path, for example /opt/conda/lib/python3.8/site-packages/torch/share/cmake, the release executable program can run successfully.
Therefore it seems the LibTorch download from official can not be directly use in C++？ Why ?
My test C++ code snippet is simple below：

#include <torch/script.h>
#include <torch/torch.h>
#include <iostream>
#include <memory>
#include <numeric>

int main(int argc, const char* argv[]) {
    if (argc != 4) {
        std::cerr << "usage: test-model <path-to-exported-script-module>\n";
        return -1;
    }

Above example What should do if I want to use the official LibTorch lib ? Maybe I use wrong version for my CPU? By the way, my OS is “Ubuntu VERSION=20.04.4 LTS (Focal Fossa)” and CPU is AMD EPYC 7713.

2

For those who may have hit the same error: it seems to be caused by the executable using different libraries (e.g. system libraries) than those used during linking.

Run ldd on the executable to see what it’s using for runtime, and make sure those are the same as passed to g++ or ld.

Note also you can grep the libraries for missing symbols, including that super long “_ZN2at…” one in your error message, to make sure they exist.