June 24, 2026 Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving 面向 Kimi 的 KVCache 中心化解耦架构，通过分离 Prefill 和 Decode 集群实现 525% 吞吐量提升 llm-serving disaggregated-serving kv-cache system-design production-system moonshot-ai