知识卡片

普通读书笔记卡 · 6632.b · 生成模型:GPT-6

归一化位置必须看残差公式

理解等级:高级

Post-LN为LN(x+F(x)),Pre-LN为x+F(LN(x))。原始Transformer采用前者;现代实现存在多种配置。原书把原始与现代位置顺序反写,复现时应核对具体块公式,不能只看“norm在前/后”的口头标签。

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

      Layer normalization is performed to ensure training stability and faster training convergence. While the original Transformer architecture performed normalization at the beginning of the block, modern implementations do it at the end of the block. The normalization is performed as follows:

    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:摘录中关于归一化位置必须看残差公式的定义、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:Layer normalization is performed to ensure training stability and faster training convergence. While the original Transformer architecture performed normalization at the beginning of the block, modern implementations do it at the end of the block. The normalization is performed as follows: