PYTORCH_MODEL with CUDA 11.2

[alexberlaga]

Dear Plumed Team,

I am using PLUMED 2.10b with CUDA 11.2 and I am attempting to enable libtorch to use the Pytorch module. I have tried with multiple versions of libtorch but it seems like one of two things is happening:

(1) I use Cxx11 ABI, and then ./configure cannot enable libtorch
(2) I use pre-Cxx11 ABI, and then the "make" step does not work. I figure this is due to some issue with my cluster being incompatible with pre-Cxx11 ABI, rather than a PLUMED thing.

My question is -- would there be any version of libtorch that is compatible both with all 3 of PLUMED 2.10b, Cxx11 ABI and CUDA 11.2? Thanks!

[luigibonati]

Hi Alex,
can you attach the configure log files in the two cases?
Furthermore, which LibTorch versions did you try? In principle, there should not be compatibility issues besides cluster compatibility. Note that LibTorch>2.1 requires C++17 support, but that is already enforced if you are using PLUMED>2.10

Luigi

[alexberlaga]

Hi, thanks for your help! I'll share log files here:

./configure --enable-libtorch --enable-modules=pytorch
config.log

./configure --enable-libtorch --enable-modules=pytorch CXXFLAGS="-D_GLIBCXX_USE_CXX11_ABI=0"
config2.log