torch.compile can achieve throughput improvements ranging from 30% to 200% for most PyTorch models without requiring any modifications to the model code.