Skip to content

Sparse attention

The efficiency lineage · 2025

The middle door of the taxonomy: fewer comparisons, same growing cache.

The full grid, then the window that misses “animal”, then top-k that keeps the near-tie — and the cache that still grows.

Sparse attention — fewer compares, same growing cache · 21s, companion clip.
What it fixedAttend only to the top-k relevant tokens - compute ~ N x k.
What it still costsMemory still grows with N. The residual that motivates the linear branch.

Worth having first: The memory wall

A window that drops the first tokens is the catastrophic case this page already showed. Sliding window with sinks keeps those first tokens on purpose — they stabilize softmax; they do not remember “animal”.