知识卡片
整数量化必须保存尺度约定
整数码配合scale和zero-point近似原实数,不是简单丢弃所有小数。完整有符号int8通常为−128到127,某些对称方案选择−127到127;码值范围和舍入、裁剪约定应明确记录。
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
The process of converting representation of a number from a higher-precision format to a lower-precision format is called quantization. We can quantize 32-bit values to 8-bit integer formats as well. This reduces memory requirements by a factor of 4, at the cost of even more precision. With 8-bit quantization, we can represent numbers between –127 and 127, without any decimal point.
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:摘录中关于整数量化必须保存尺度约定的定义、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:The process of converting representation of a number from a higher-precision format to a lower-precision format is called quantization. We can quantize 32-bit values to 8-bit integer formats as well. This reduces memory requirements by a factor of 4, at the cost of even more precision. With 8-bit quantization, we can represent numbers between –127 and 127, without any decimal point.