I understand it may be inefficient sometimes. Perhaps it would be nice to add an argument to lru_cache to specify that we want to deep copy? Something like

def lru_cache(..., deepcopy=False):
   ...