← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6635 · 生成模型:GPT-6

稀疏专家分开总容量与活跃量

理解等级:高级

MoE为不同输入选择部分专家计算,使总参数容量可大于每次活跃参数。所有专家的存储与部署需求仍存在,路由、负载和设备通信也有开销;不能说增加容量完全不增加计算或资源。

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

      conditional computation; each input (either a token or the entire sequence) sees a different subset of the model, interacting with only the parameters that are best suited to process it. This is achieved by composing the architecture to be made up of several components called experts, with only a subset of experts being activated for each input.

    参考来源

    - 位置:《Designing Large Language Model Applications》第04章《架构与学习目标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch04.html) - 结论依据:摘录中关于稀疏专家分开总容量与活跃量的定义、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:conditional computation; each input (either a token or the entire sequence) sees a different subset of the model, interacting with only the parameters that are best suited to process it. This is achieved by composing the architecture to be made up of several components called experts, with only a subset of experts being activated for each input.