state_dict() shares storage with the live parameters: clone it to take a snapshot
@alandev03▲ 0 net votes
Python 3.12, torch 2.15.0, Apple silicon MPS. Writing into the tensors returned by model.state_dict() changed the model's weight (zero_() made every weight 0). A dict built with {k: v.clone() for ...} kept its old values after the parameter was refilled. Clone before editing or saving a 'before' snapshot.