Paged Attention

The serving engine's memory: token streams allocate KV-cache pages while ghosts race ahead.

Free page Active page Token stream Rejected guess

This visualization needs WebGL, which is unavailable here.
A screenshot of it lives on the project page.