Unverified50% confidenceFactExact time
Prompt Cache本质上是大模型推理架构中KV Cache机制的服务端延伸，缓存Transformer中每个token生成的Key和Value矩阵
1
Sources
50%
Confidence
Long-term
Relevance
7/20/2026
First Seen
Sources
Related Claims
UnverifiedAnthropic的提示缓存（Prompt Caching）利用Transformer架构中的KV Cache机制，通过缓存共享提示前缀对应的键值状态来降低延迟和成本78% similarUnverified提示缓存的核心原理是KV Cache的跨请求复用，当多个请求共享相同前缀时已计算的KV状态可被缓存复用71% similarVerifiedKV Cache通过将已计算的键值对持久化存储，使相同Prompt前缀的后续请求可直接复用，跳过重复的矩阵运算69% similarUnverifiedvLLM的前缀缓存（Prefix Caching）对KV Cache块计算哈希值，当新请求输入前缀与已缓存请求重叠时直接复用物理缓存块，跳过该部分prefill计算68% similarUnverifiedKV Cache（键值缓存）机制使得共享相同前缀的API请求可以复用已计算的Key和Value向量，从而大幅降低推理成本68% similar
Cite This Claim
Stable URI
https://kongchang.com/claim/573262API
curl https://kongchang.com/api/v1/knowledge/claims/573262MCP
get_claim(id=573262)