I'd like to reuse my Linux libtorch installation in Pytorch as well. The benefits would be:
- reducing unpacked Python package size by ~600MB of shared objects
- certainty that I'm using the same backend implementation in both Python and other frontends
Is this possible? Scanning the Pytorch code, I don't see an obvious way (it seems to load via CDLL from a fixed path relative to __init__.py), but I'm not certain.
Extra context: It wasn't too hard to make a custom Pytorch build that used system CUDA/CUDNN libraries, so I'm somewhat encouraged and hope this is possible for libtorch as well.