This is still pretty idealized. In my experience the pytorch abstraction leaks _constantly_. In particular, any interesting ML project probably pulls in at least a few dependencies with custom pytorch extensions somewhere.