Mooncake Store for Agentic RL in Disaggregated Architectures

Talk overview

Mooncake primarily provides a high-performance data plane in reinforcement learning (RL): it optimizes rollout data transmission via structured objects, DataProto/typed-ragged, BufferPool, and RDMA, and integrates with Miles to replace the Ray Object Store while supporting both synchronous and fully asynchronous modes. For weight updates, it supports NCCL broadcast, Mooncake TransferEngine P2P RDMA, and disk-delta, achieving approximately a 7x speedup with P2P on Kimi-K2. Additionally, it supports heterogeneous resharding for TP/PP/EP/DP, Store weight snapshots, and MoE sparse deltas. Its production capabilities are utilized across ecosystems such as Miles, SGLang, and vLLM.