知识卡片
稠密模型转专家还要继续训练
upcycling可复制预训练前馈层作为多个专家的初始化,复用已有稠密模型知识。复制只提供起点,路由与专家差异还需后续学习;不能把参数副本数量当成已经获得同倍独立知识。
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
- · 架构与学习目标
原书出处 · 1 处
- · 查看原书依据
upcycling method that can be used to add an MoE component to an already pre-trained model. This is done by making N copies of the feedforward layers, one for each expert, and using the original parameters of these layers as the initialization parameters for the MoE model.
参考来源
- 位置:《Designing Large Language Model Applications》第04章《架构与学习目标》(来源项目:a19c7933-4eda-4176-aebd-30fe7e1ad292;源文件:extracted/Designing Large Language Model Applications/_epub-src/OEBPS/ch04.html)
- 结论依据:摘录中关于稠密模型转专家还要继续训练的定义、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:upcycling method that can be used to add an MoE component to an already pre-trained model. This is done by making N copies of the feedforward layers, one for each expert, and using the original parameters of these layers as the initialization parameters for the MoE model.