← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 6503.b · 生成模型:GPT-6

缓存正确性先于加速比

理解等级:中级

比较缓存与无缓存版本时,先在相同权重、输入、解码策略下核对输出,再比较耗时。一个样本输出一致只是基本检查,仍需覆盖多步、不同长度、EOS 和重置行为;不能用一次提速或一次一致证明缓存实现对所有情况正确。

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

      Importantly, we also see that the generated text is the same as before, which is an important sanity check to ensure that the KV cache is implemented and used correctly.

    参考来源

    - 位置:《Build a Reasoning Model (From Scratch)》第02章《使用预训练模型生成文本》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Build a Reasoning Model (From Scratch)/_epub-src/OEBPS/Text/chapter-2.html) - 结论依据:摘录中关于缓存正确性先于加速比的定义、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:Importantly, we also see that the generated text is the same as before, which is an important sanity check to ensure that the KV cache is implemented and used correctly.