Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. This repository hosts its technical report and will also be utilized for the forthcoming open sourcing of traces. Stay tuned!

• June 27, 2024: We present a series of Chinese blogs with more discussions on zhihu 1, 2, 3.
• June 26, 2024: Initial technical report release.

Mooncake features a KVCache-centric disaggregated architecture that separates the prefill and decoding clusters. It also leverages the underutilized CPU, DRAM, and SSD resources of the GPU cluster to implement a disaggregated cache of KVCache.

The core of Mooncake is its KVCache-centric scheduler, which balances maximizing overall effective throughput while meeting latency-related Service Level Objectives (SLOs) requirements. Unlike traditional studies that assume all requests will be processed, Mooncake faces challenges due to highly overloaded scenarios. To mitigate these, we developed a prediction-based early rejection policy. Experiments show that Mooncake excels in long-context scenarios. Compared to the baseline method, Mooncake can achieve up to a 525% increase in throughput in certain simulated scenarios while adhering to SLOs. Under real workloads, Mooncake’s innovative architecture enables Kimi to handle 75% more requests.

@article{qin2024mooncake,
  title        = {Mooncake: A KVCache-centric Disaggregated Architecture for LLM Serving},
  author       = {Ruoyu Qin, Zheming Li, Weiran He, Mingxing Zhang, Yongwei Wu, Weimin Zheng, and Xinran Xu},
  year         = {2024},
  url          = {https://arxiv.org/abs/2407.00079}
}