This is an old thread, but linked recently:
These days, there also is a save / load functionality in C++, but is based on the JIT file format. So you can save a traced model (but use the model, not the function) und load weights back into your PyTorch C++ API-built model.
Best regards
Thomas