2026 Sep 30
Hybrid KV Cache, Illustrated (Chinese) Hybrid KV Cache 图解
How vLLM, SGLang and Marconi do prefix caching for hybrid attention + recurrent (GDN / Mamba) models, with interactive labs. Written in Chinese.
Hybrid 模型（attention + GDN / Mamba）的 prefix cache：vLLM、SGLang 和 Marconi 各自怎么做，附交互实验。