知识卡片

普通读书笔记卡 · 8689.e · 生成模型:GPT-6

动态与静态量化各有成本

理解等级:中级

动态缩放运行时求激活统计,可适应当前输入却增加计算;静态缩放用参考校准集估计,执行较直接却依赖代表性。目标分布变化时要检查饱和与误差,不能固定宣称静态一定更快或更差。

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

      dynamic or static scaling. In dynamic scaling, statistics like range, outliers, etc. are calculated dynamically during inference at each layer. In static scaling, we take a reference calibration dataset to estimate the statistics. While this approach speeds up inference, it can result in more quantization errors.

    参考来源

    - 位置:《Designing Large Language Model Applications》第09章《推理优化》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch09.html) - 结论依据:摘录中关于动态与静态量化各有成本的定义、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:dynamic or static scaling. In dynamic scaling, statistics like range, outliers, etc. are calculated dynamically during inference at each layer. In static scaling, we take a reference calibration dataset to estimate the statistics. While this approach speeds up inference, it can result in more quantization errors.