Memory-Efficient Model Weight Loading in PyTorch