知识卡片
数值格式改变范围和精度
FP16与BF16虽同为16位,指数和尾数分配不同,分别影响动态范围与有效精度。权重压缩还要看内核与硬件支持;“Int8固定慢20%”属于特定条件经验,不能作为通用排序。
这张卡片看不懂?
这张卡片已通过人工审查,不再接受失败反馈。
已有反馈 · 0 条
还没有人提交反馈。
在整书笔记中的位置
- · 按任务适配模型
原书出处 · 1 处
- · 查看原书依据
Only 5 bits are reserved for the exponent as opposed to 8 bits in Float32. This means that using Float16 comes with overflow/underflow problems for very large and small numbers.
参考来源
- 位置:《Designing Large Language Model Applications》第05章《按任务适配模型》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch05.html)
- 结论依据:摘录中关于数值格式改变范围和精度的定义、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:Only 5 bits are reserved for the exponent as opposed to 8 bits in Float32. This means that using Float16 comes with overflow/underflow problems for very large and small numbers.