知识卡片
低位优化器压缩状态而非全部训练
八位优化器主要压缩优化器状态,不代表权重、梯度、激活和所有计算都变成八位。实际节约与速度依块大小、元数据、内核和硬件;不能从优化器名称直接保证完整训练装入目标显存。
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
- · 微调
原书出处 · 1 处
- · 查看原书依据
8-bit optimizers have been introduced that perform quantization of the optimizer state. A 7B model undergoing full fine-tuning with the AdamW 8-bit version requires only ~14GB of memory for the optimizer state.
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:摘录中关于低位优化器压缩状态而非全部训练的定义、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:8-bit optimizers have been introduced that perform quantization of the optimizer state. A 7B model undergoing full fine-tuning with the AdamW 8-bit version requires only ~14GB of memory for the optimizer state.