← 返回刚才阅读的位置

知识卡片

普通读书笔记卡 · 8660.c · 生成模型:GPT-6

分页优化器用主存缓解显存压力

理解等级:中级

分页机制可在需要时把部分优化器内存页迁移到CPU侧,以降低某些显存峰值压力。它仍需可用主存和传输成本,不能修复所有激活峰值,也不保证发生频繁换页时速度不受影响。

这张卡片看不懂?

已有反馈 · 0 条

    还没有人提交反馈。

    在整书笔记中的位置

    原书出处 · 1 处

    • · 查看原书依据

      Paged Optimizers Using AdamW as an optimizer requires eight bytes of memory per parameter, which is a significant drag on memory requirements. This affects the maximum sequence length that can be supported. This is where paged optimizers can come in handy. In cases where the GPU runs out of memory during fine-tuning, paged optimizers automatically transfer memory pages to CPU RAM, then transfer them back

    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:摘录中关于分页优化器用主存缓解显存压力的定义、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:Paged Optimizers Using AdamW as an optimizer requires eight bytes of memory per parameter, which is a significant drag on memory requirements. This affects the maximum sequence length that can be supported. This is where paged optimizers can come in handy. In cases where the GPU runs out of memory during fine-tuning, paged optimizers automatically transfer memory pages to CPU RAM, then transfer them back