Optimizing Memory Usage in PyTorch Models