PyTorch/Libtorch: can't find GLIBC on Fedora when making project

Question

I'm trying to build a C++ project that uses Libtorch on a new laptop and am keep getting the following error when I run Make:

/home/labuser/anaconda3/envs/Research/bin/../lib/gcc/x86_64-conda-linux-gnu/13.1.0/../../../../x86_64-conda-linux-gnu/bin/ld: /home/labuser/anaconda3/envs/Research/lib/python3.11/site-packages/torch/lib/libc10.so: undefined reference to `memcpy@GLIBC_2.14'
collect2: error: ld returned 1 exit status
make[2]: *** [CMakeFiles/learn.dir/build.make:101: learn] Error 1
make[1]: *** [CMakeFiles/Makefile2:83: CMakeFiles/learn.dir/all] Error 2
make: *** [Makefile:91: all] Error 2

For reference, my operating system is Fedora 38 with glibc 2.37, and I installed PyTorch/Libtorch and GXX using Conda. The project itself is build using CMake, and then running Make shows the above error.

This issue doesn't happen on my Fedora 37 desktop, where I have glibc 2.36 installed.

Things I've tried:

• Updated glibc using dnf update glibc
• Updated my PyTorch installation, etc. in Conda
• I also tried linking directly with Libtorch header files (both non-ABI and ABI) downloaded from the PyTorch website.

I'm rather confused on how to proceed from here.

One more thing: running CMakeList did throw a warning. It doesn't on my other Fedora computer.

-- Caffe2: Protobuf version 3.20.3
CMake Warning at /home/labuser/anaconda3/envs/Research/lib/python3.11/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:22 (message):
  static library kineto_LIBRARY-NOTFOUND not found.
Call Stack (most recent call first):
  /home/labuser/anaconda3/envs/Research/lib/python3.11/site-packages/torch/share/cmake/Torch/TorchConfig.cmake:127 (append_torchlib_if_found)
  CMakeLists.txt:8 (find_package)

How can I proceed? I found some advice on Stack Exchange for glibc when writing own code, but I am unsure how to work around the PyTorch/Libtorch one.

I updated PyTorch and GXX in my Conda install and updated glibc using DNF.

Answer 1

I found an answer to my issue, posting it here for people who are likely to have this issue as well.

It's anaconda. The newest Anaconda changes the default g++ and gcc to it's own, instead of using the system one.This is fine, except here it breaks on Fedora the ability to link to glibc via libtorch. Solution:

1. Use dnf or yum to install the correct system compiler.
2. Set the CC and CXX environment variables to point to the system compiler in /usr/bin . On installation, Anaconda sets these to point to its own, and these are the environment variables CMake uses to choose its compilers.
3. Profit lol.

Doing so led to successful linking, even with Anaconda's own torch libraries.