Skip to content
All cold checks
GPU kernelsExpertAlgorithm derivation9 min

Situation

You must explain why tiled attention can be exact even though each query tile sees the key and value blocks one at a time and the softmax denominator is not known in advance.

Derive the running state needed for online softmax and explain how an old output accumulator is corrected when a later tile contains a larger maximum. Then state what FlashAttention changes and does not change about asymptotic compute and memory traffic.

Keep going

A different failure surface.