DynaServe: Unified and Elastic Execution for Dynamic Disaggregated LLM Serving

Chaoyi Ruan, Yinhe Chen, Dongqi Tian +4

LLM inference must meet strict latency SLOs (e.g., 100 ms P99 time-between-tokens) while maximizing goodput. Yet, real-world variability in prompt and response lengths skews comput…