知识卡片
缓存容量与生命周期共同管理
K/V缓存会随token数、层数、键值头和精度增长,还与并发序列有关。应考虑可复用程度、有效期和淘汰,而非无限保存;服务缓存收费与TTL属版本化规则,书中“只收费一次”不能当当前或普遍定价。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 推理优化
原书出处 · 1 处
- · 查看原书依据
Note that in the caching strategy, we are trading compute for additional storage. K-V caches can get unfeasibly large, especially at longer sequence lengths. To keep costs under control, LLM providers typically limit the age of the cache to a short period or price users by caching duration. As an example, let’s look at a request to Anthropic’s Claude suite of models that utilizes context
参考来源
- 位置:《Designing Large Language Model Applications》第09章《推理优化》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch09.html)
- 结论依据:摘录中关于缓存容量与生命周期共同管理的定义、比较或条件支撑卡片;原书表述不严谨的范围已在覆盖记录逐项限定,不把案例结果外推为普遍结论。
- 原始内容:Note that in the caching strategy, we are trading compute for additional storage. K-V caches can get unfeasibly large, especially at longer sequence lengths. To keep costs under control, LLM providers typically limit the age of the cache to a short period or price users by caching duration. As an example, let’s look at a request to Anthropic’s Claude suite of models that utilizes context