2/ Tail-Aware Scheduling: cut LLM serving tail latency without length prediction. How: no decode-length predictor; soft priority boosting + cache-aware preemption. Result: P99 TTLT -35-50% vs SRPT w/…