GLIBC mismatch for cxx11 libtorch

[ElliottKasoar]

This may not be an FTorch issue, but attempting to link to FTorch installed with the cxx11 libtorch (2.2.1) throws an error due to what seems to be a mismatch between the GLIBC version it references (2.29) and the installed version (ldd (GNU libc) 2.28).

The solution for now seems to be using the non-cxx11 version, as the steps described below work when the URL is swapped for https://download.pytorch.org/libtorch/cpu/libtorch-shared-with-deps-2.2.1%2Bcpu.zip.

Steps to reproduce (also attempted with gcc 9):

1. Login to CSD3 (login-icelake.hpc.cam.ac.uk)

2. Run commands along the lines of:

module load cmake/latest
module load gcc/11.3.0

cd /rds/user/ek440/hpc-work/

wget https://download.pytorch.org/libtorch/cpu/libtorch-cxx11-abi-shared-with-deps-2.2.1%2Bcpu.zip
unzip libtorch-cxx11-abi-shared-with-deps-2.2.1+cpu.zip

cd FTorch/src/build

cmake .. -DCMAKE_PREFIX_PATH=/rds/user/ek440/hpc-work/libtorch -DCMAKE_BUILD_TYPE=Release -DCMAKE_INSTALL_PREFIX=/rds/user/ek440/hpc-work/lib_libtorch -DCMAKE_C_COMPILER=gcc -DCMAKE_CXX_COMPILER=g++ -DCMAKE_Fortran_COMPILER=gfortran

cmake --build . --target install

cd ../../examples/2_ResNet18/build

cmake .. -DCMAKE_PREFIX_PATH=/rds/user/ek440/hpc-work/lib_libtorch/ -DCMAKE_BUILD_TYPE=Release -DCMAKE_Fortran_COMPILER=gfortran

cmake --build .

This throws an error for me:

//rds/user/ek440/hpc-work/libtorch/lib/libtorch_cpu.so: undefined reference to `log2@GLIBC_2.29'
//rds/user/ek440/hpc-work/libtorch/lib/libtorch_cpu.so: undefined reference to `pow@GLIBC_2.29'
//rds/user/ek440/hpc-work/libtorch/lib/libtorch_cpu.so: undefined reference to `exp2@GLIBC_2.29'
//rds/user/ek440/hpc-work/libtorch/lib/libtorch_cpu.so: undefined reference to `exp@GLIBC_2.29'
//rds/user/ek440/hpc-work/libtorch/lib/libc10.so: undefined reference to `log@GLIBC_2.29'
collect2: error: ld returned 1 exit status

I recall there can be issues with LD_LIBRARY_PATH being incorrectly set, but that does not appear to be an the case here, as it seems to find libtorch_cpu.so fine (and I've tried setting it), but also if I set it differently:

LD_LIBRARY_PATH=/rds/user/ek440/hpc-work/libtorch-non-cxx/lib/

I get a different error on the last step:

/rds/user/ek440/hpc-work/lib_libtorch/lib64/libftorch.so: undefined reference to `c10::ListType::get(std::__cxx11::basic_string<char, std::char_traits<char>, std::allocator<char> > const&, c10::Type::SingletonOrSharedTypePtr<c10::Type>)'

It's odd to me that it gets so far before the error occurs.

The main alternative suggestion other than different versions of libtorch/glibc I've seen is to export CC and CXX, which didn't seem to solve the problem.

I don't plan to look into this much further, but I thought it might be useful to be aware of if nothing else.

[jatkinson1000]

I suggest someone check this in a future hackathon session to see if the issue is still there in the latest versions, then we close the issue as I see little we can do except have this as a point of reference/information for future users that may encounter it.