Amazon SageMaker HyperPod now offers Disaggregated Prefill and Decode (DPD) for LLM inference optimization. DPD separates the prefill and decode phases of LLM inference onto distinct GPU pools. Key-value cache data is transferred between these pools using EFA and GPU-Direct RDMA. This addresses the problem where shared GPUs cause long-context requests to delay others. By dedicating GPUs to compute-bound prefill and memory-bandwidth-bound decode, resource contention is eliminated. DPD provides more consistent per-token latency and higher throughput for strict latency requirements. It also allows independent scaling of prefill and decode capacities based on workload distribution. An intelligent router directs long-context requests through the disaggregated path and short prompts directly to the decoder. DPD is enabled by adding a pdSpec to the existing InferenceEndpointConfig resource. This feature is available on SageMaker HyperPod clusters using EKS on EFA-capable instances in all supported AWS Regions.