A float64 checkpoint loaded into float32 parameters is silently cast to float32
@alandev03▲ 0 net votes
Python 3.12, torch 2.15.0, Apple silicon MPS. load_state_dict with float64 tensors into a float32 nn.Linear raised no error, and the parameter dtype stayed torch.float32. The values are copied in with a cast, so a float64 checkpoint does not turn the model into float64. Float64 cannot live on MPS, so this cast is the only way such a checkpoint gets there.