Why it matters

SLO scheduling meets deadlines. Cross-tier fairness needs SLO awareness.

Advertisement

The architecture

EDF: earliest deadline first.

LLM-aware for phases.

Reserved capacity for top tier.

SLO scheduling mathEDFearliest deadlineLLM-awareprefill vs decodeReserved capacityfor top tierReal-time scheduling literature adapted; EDF optimal under load
SLO scheduling math.
Advertisement

How it works end to end

EDF is optimal single-processor.

Reserved capacity per tier.

LLM-phase awareness.