知识卡片

普通读书笔记卡 · 6516.c · 生成模型:GPT-6

查询键归一化不是分组定义

理解等级:中级

QKNorm 对每个查询和键的头维向量归一化,本章在旋转前执行。它与多个查询共享KV的GQA是独立设计选择;复现时需分别保存qk_norm开关、维度及顺序,不能只凭GQA名称推断已有该归一化。

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

      not part of the standard GQA design. When qk_norm=True, an additional Query/Key-RMSNorm-based normalization, called QKNorm, is applied to both the queries and keys, which is a technique used in Qwen3. As discussed earlier in the RMSNorm section (section C.1), QKNorm helps improve training stability.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第09章《附录C Qwen3模型源码》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/appendix-c.html) - 结论依据:摘录中关于查询键归一化不是分组定义的定义、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:not part of the standard GQA design. When qk_norm=True, an additional Query/Key-RMSNorm-based normalization, called QKNorm, is applied to both the queries and keys, which is a technique used in Qwen3. As discussed earlier in the RMSNorm section (section C.1), QKNorm helps improve training stability.