知识卡片
参数高效不等于前向计算很少
PEFT减少可训练参数及相关梯度、优化器状态和每任务存储。基础模型通常仍需前向,某些位置还需反向传递,新增模块也可能增推理开销;节约幅度不能只从可训练参数比例推出。
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
- · 进阶微调技术
原书出处 · 1 处
- · 查看原书依据
Parameter-Efficient Fine-Tuning (PEFT) A set of fine-tuning techniques that make the fine-tuning process more efficient by updating only a small number of model parameters, thus needing less memory and compute. Model merging/model fusion An exciting new subfield of LLMs that explores combining the parameters of two or more models. I call this the “dark arts” of NLP, as it is poorly understood but uncannily effective
参考来源
- 位置:《Designing Large Language Model Applications》第07章《进阶微调技术》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch07.html)
- 结论依据:摘录中关于参数高效不等于前向计算很少的定义、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:Parameter-Efficient Fine-Tuning (PEFT) A set of fine-tuning techniques that make the fine-tuning process more efficient by updating only a small number of model parameters, thus needing less memory and compute. Model merging/model fusion An exciting new subfield of LLMs that explores combining the parameters of two or more models. I call this the “dark arts” of NLP, as it is poorly understood but uncannily effective