Search papers, labs, and topics across Lattice.
KVServe addresses the KV cache communication bottleneck in disaggregated LLM serving by introducing a service-aware and adaptive compression framework. It unifies KV compression techniques into a modular strategy space, employs a Bayesian Profiling Engine to efficiently identify Pareto-optimal compression profiles, and uses a Service-Aware Online Controller to dynamically select profiles based on real-time service conditions. Experiments demonstrate significant JCT speedups (up to 9.13x) in PD-separated serving and TTFT reduction (up to 32.8x) in KV-disaggregated serving when integrated with vLLM.
Forget static KV cache compression – KVServe dynamically adapts compression strategies to your service context, slashing latency by up to 32.8x in disaggregated LLM serving.
LLMs are widely adopted in production, pushing inference systems to their limits. Disaggregated LLM serving (e.g., PD separation and KV state disaggregation) improves scalability and cost efficiency, but it also turns KV into an explicit payload crossing network and storage boundaries, making KV a dominant end-to-end bottleneck. Existing KV compression are typically static runtime configurations, despite production service context varies over time in workload mix, bandwidth, and SLO/quality budgets. As a result, a fixed choice can be suboptimal or even increase latency. We present \emph{KVServe}, the first service-aware and adaptive KV communication compression framework for disaggregated LLM serving: KVServe (1) unifies KV compression into a modular strategy space with new components and cross-method recomposition; (2) introduces Bayesian Profiling Engine that efficiently searches this space and distills a 3D Pareto candidate set, reducing $50\times$ offline search overhead; and (3) deploys a Service-Aware Online Controller that combines an analytical latency model with a lightweight bandit to select profiles under constraints and correct offline-to-online mismatch. Integrated into vLLM and evaluated across datasets, models, GPUs and networks, KVServe achieves up to $9.13\times$ JCT speedup in PD-separated serving and up to $32.8\times$ TTFT reduction in KV-disaggregated serving.
