PagedAttentionPagedAttention
PagedAttention
解决什么问题What problem it solves
借操作系统的分页思路管理 KV 缓存，让显存几乎不浪费。Manage the KV cache with OS paging ideas so memory is barely wasted.
这是什么What it is
Kwon 等人的定义：受操作系统虚拟内存与分页技术启发的注意力算法。
论文称基于它构建的 vLLM 在 KV 缓存上接近零浪费，吞吐提升 2–4 倍。
Kwon et al.'s definition: an attention algorithm inspired by classical virtual memory and paging in operating systems.
The paper reports vLLM achieving near-zero KV cache waste and 2-4x throughput.
为什么需要知道Why it matters
vLLM 快的真正原因。理解了它，才明白显存要算的是缓存而不只是权重。The real reason vLLM is fast. Memory budgeting is about cache, not only weights.