Skip to visualization
visualization / Flash Attention

minGPT · Flash Attention

tiled exact attention
◆ FLASH
T=8
i8
i8

🔍 Variable Inspector

click any code line
Click a code line to inspect its variables.
About this lesson
Quick answer

What FlashAttention explains

Tile exact attention with online softmax so the full score matrix is never materialized in high-bandwidth memory.

Key concepts

Primary research

FlashAttention · Dao et al.