---
issue: 4
status: done
tags: systems
---
# fused-attention

A flash-style attention kernel with an online softmax.

## What happened

- 2026-10-01: 1.52 speed-up (+0.52 on the baseline)
- 2026-10-04: 1.47 speed-up (+0.47 on the baseline)

## Next

Search the tile sizes per shape.