知识卡片
对称量化让实数零对应零码
对称absmax方案可令s=max|x|/127,q=clip(round(x/s),−127,127),还原近似为s·q。最大绝对值确定尺度,舍入和溢出处理影响误差;书图63.5到63不能代替明确的舍入规则。
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
- · 推理优化
原书出处 · 1 处
- · 查看原书依据
Symmetric Quantization In this setting, the 0 value in the original format is mapped to the 0 value in the integer representation. This means that when you quantize 0 represented in fp32 to int8, the value remains 0. The remaining values can be mapped using various techniques, the most common being absmax quantization. In this method, if we know or can estimate the range of
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:摘录中关于对称量化让实数零对应零码的定义、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:Symmetric Quantization In this setting, the 0 value in the original format is mapped to the 0 value in the integer representation. This means that when you quantize 0 represented in fp32 to int8, the value remains 0. The remaining values can be mapped using various techniques, the most common being absmax quantization. In this method, if we know or can estimate the range of