← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8662.b · 生成模型:GPT-6

混合精度不同于低位量化

理解等级:中级

BF16/FP16浮点训练与四位权重量化是不同机制。量化存储、反量化计算类型、非量化模块和训练状态可使用不同精度,应分别记录;不能把一个bf16开关称为完成四位训练。

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

      A very effective form of saving memory is through quantization, as introduced in Chapter 5. We will go through quantization techniques in more detail in Chapter 9. For our use case, we will use bf16 as it represents a sound tradeoff between memory savings and performance.

    参考来源

    - 位置:《Designing Large Language Model Applications》第06章《微调》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch06.html) - 结论依据:摘录中关于混合精度不同于低位量化的定义、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:A very effective form of saving memory is through quantization, as introduced in Chapter 5. We will go through quantization techniques in more detail in Chapter 9. For our use case, we will use bf16 as it represents a sound tradeoff between memory savings and performance.