← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8701.a · 生成模型:GPT-6

向量存储先算基础字节量

理解等级:中级

N个d维FP32向量的纯数值存储为N·d·4字节;1亿个768维向量约307.2GB十进制。索引、元数据、副本和运行缓冲另计,不能把这个下界直接当数据库总资源需求。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      dimension. Therefore, storing 100 million vectors generated from the all-mpnet-base-v2 model, which has 768 dimensions, needs close to 300 GB of memory! It is not uncommon to represent a single sentence, almost always no longer than 40 tokens, with a 768-dimension vector. Do we really need 768 dimensions to represent 40 tokens? The reality is that embedding training is very inefficient, and a large number

    参考来源

    - 位置:《Designing Large Language Model Applications》第11章《表示学习与嵌入》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch11.html) - 结论依据:摘录中关于向量存储先算基础字节量的定义、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:dimension. Therefore, storing 100 million vectors generated from the all-mpnet-base-v2 model, which has 768 dimensions, needs close to 300 GB of memory! It is not uncommon to represent a single sentence, almost always no longer than 40 tokens, with a 768-dimension vector. Do we really need 768 dimensions to represent 40 tokens? The reality is that embedding training is very inefficient, and a large number