I require a PyTorch environment which is dynamically linked with libcudart on Linux. I installed the following official wheel:
https://download.pytorch.org/whl/cu101/torch-1.8.1%2Bcu101-cp36-cp36m-linux_x86_64.whl
While it includes a dynamic cudart library libcudart-1b201d85.so.10.1, it appears that libtorch_python.so in particular is statically linked to cudart. Is this intended? This results in different versions of cudart functions being called.
Do I need to resort to building PyTorch to have full dynamic linking to cudart? Or does anyone know of any available PyTorch wheels which are truly dynamically linked to cudart?
Thank you for any tips on this. I’m trying to avoid building PyTorch as it is a bit of a beast, takes over 5 hours, and has crashed my machine several times from out of memory situations, despite having 16GB of memory.