Disaggregated Inference at Scale with PyTorch and VLLM