Unverified50% confidenceFactExact time
PagedAttention divides the KV Cache into fixed-size blocks, with each block typically storing the key-value pairs of 16 tokens.
1
Sources
50%
Confidence
Long-term
Relevance
8/7/2026
First Seen
Sources
vLLM架构深度解析：PagedAttention如何实现高吞吐LLM推理
Related Claims
UnverifiedIn vLLM, each memory page typically stores KV vectors for a fixed number of tokens (e.g., 16), allowing KV Cache from different requests to be stored non-contiguously in physical memory.65% similarUnverified差分启发式需要O(K×N)的存储空间（K个地标、N个节点），例如100万节点的图使用16个地标、每个距离值4字节存储，总内存约64MB64% similarUnverifiedvLLM的PagedAttention将KV Cache划分为固定大小物理块，通过块表实现地址映射，将碎片化率从60-80%降至不足4%，吞吐量提升可达24倍64% similarUnverifiedA 200K token context request requires approximately 25× the KV-Cache memory compared to an 8K token request.63% similarVerifiedPagedAttention技术借鉴操作系统虚拟内存的分页思想，将KV Cache切分为固定大小的物理块动态分配63% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/705558API
curl https://kongchang.com/api/v1/knowledge/claims/705558MCP
get_claim(id=705558)