Platform for AI (PAI) - EAS launches Prefill-Decode (PD) separation feature
Aug 07 2025
Platform for AI (PAI)EAS now supports Prefill-Decode (PD) separation, offering both static and dynamic deployment modes. It supports multiple inference engines including vLLM, SGLang, and BladeLLM, enabling significant reduction in inference latency, especially for large LLM workloads.
Content
Target customers: Designed for customers building LLM-driven applications and services on the EAS platform. 1. Enterprises running high-traffic consumer-facing applications: Improved user experience with reduced Time to First Token (TTFT) and Time Per Output Token (TPOT). 2. Organizations processing long-context workloads: Reduced end-to-end latency for long input sequences. New Feature/Specification: EAS supports enabling Prefill-Decode (PD) separation during LLM service deployment. This feature divides the inference task into two independent phases, Prefill and Decode, and allocates them to their own computing resources for execution. This significantly improves system throughput while meeting strict latency requirements.
