vLLM with torch.compile: Efficient LLM inference on PyTorch