Resident KV Claims: A New Contract for AI Cache Memory Under Pressure

Lukas Stepanek's arXiv paper (2605.24259) tackles a growing pain in LLM serving: KV-cache reuse mechanisms expose priority, offload, and routing hints, but lack a portable contract for what happens when 'resident KV' (cached prefixes for future reuse) and 'active KV' (current request compute) exceed available memory. Existing systems silently evict residents, breaking reuse guarantees. Stepanek's 'resident KV claims' formally bind future-reuse intent to a materialization predicate, lifecycle state, and feasibility outcome.

In vLLM allocator probes, a 60-block resident claim and a 70-block active prefill oversubscribe an 80-block pool. Write no-admit prevents the active request from becoming reusable but still allows eviction of residents. A minimal vLLM prototype shows that hard protected resident claims convert this failure into scheduler-visible active refusal with direct blocking-claim attribution. The result isn't a speedup—it's a runtime contract that turns unreported resident loss into reconstructable arbitration, with companion litmus tests distinguishing eviction, soft priority, demotion, expiry, and active refusal.