PagedAttention

Storing the KV cache in fixed-size blocks that need not be contiguous, the way an operating system pages memory. It removes the fragmentation that otherwise wastes most of the cache, so many more requests fit at once.

A definition is not an understanding. This is taught properly in Serving open models yourself, part of Deployment & serving.