Hi, I am experimenting with giant models (3B, 11B) parameters, and I noticed that when loading state dict we must first create the model instance and only then load the new state dict I want which takes 2x time and 2x memory.
Is there some workaround?