What does FlashAttention change about transformer training and serving?
Simple meaning
FlashAttention computes attention in GPU SRAM tiles so it avoids materializing the full n-by-n matrix in HBM.
Open the full page for Why, Steps, Example and Key takeaway.