Study note
Paged Attention from First Principles: A View Inside vLLM
Properties
- Type
- Blogs
- Status
- 待读
- Domain
- AI / ML
- Category
- LLM 推理与服务
- Source
- hamzaelshafie.bearblog.dev
- Vault note
library/articles/ai_ml/Paged-Attention-from-First-Principles-A-View-Inside-vLLM-ed6d17cd1f84407d.md
Summary
从训练与推理差异、朴素 KV cache 的浪费讲到操作系统分页类比和 vLLM PagedAttention。
Highlights
适合在读论文或源码前建立分页 KV 缓存解决碎片与共享问题的直观模型。
Comments
Loading comments...