Paged Attention Deep Dive

Why it matters

Paged Attention shapes vLLM. Understanding shapes serving.

The architecture

KV cache in fixed-size blocks.

Block table maps logical to physical.

Copy-on-write for shared prefixes.

How it works end to end

Kwon et al. 2023.

Block size 16-32 typical.

Prefix sharing.

Swap to CPU / disk possible.