General Non-Clairvoyant KV-Cache Scheduling via Regime-Aware Routing
For LLM serving systems, this provides a theoretically optimal scheduling policy under memory constraints, solving a key bottleneck in batched inference.
This paper presents the first constant-competitive algorithm for non-clairvoyant KV-cache scheduling in LLM inference, achieving optimal total completion time against a clairvoyant baseline without assumptions on prompt or response lengths.
We study non-clairvoyant scheduling for batched Large Language Model (LLM) inference under a hard Key-Value (KV) cache memory budget. Each request has a known prompt length but an unknown response length, and its memory footprint comprises a fixed prompt component together with a response component that grows with each decoded token. At each decoding round, the scheduler chooses a feasible batch of active requests; evicting a request discards its accumulated cache states, wasting prior computation. The goal is to minimize total completion time against the optimal clairvoyant schedule that knows all response lengths. We present the first constant-competitive algorithm for arbitrary prompt lengths and arbitrary response lengths with no additional assumptions. Rather than relying on a single universal scheduling policy, our algorithm is built on a novel regime-aware routing framework. Specialized sub-schedulers handle different memory-growth geometries, while a meta-scheduler time-shares the memory budget across them and dynamically routes each job as its execution progressively reveals its behavior. This framework also yields constant-competitive guarantees for makespan and for total completion time under online arrivals.