—Christoph Thaiss
My best theory: the fused standard path wins because XLA sees the entire softmax(Q @ K.T) @ V expression at once and compiles it into one optimized kernel — no intermediate matrices spilling to HBM. My flash attention uses fori_loop, which XLA likely compiles as a generic sequential loop. It probably can’t fuse across iterations, can’t pipeline memory loads, can’t interleave independent work. (I haven’t dumped the HLO to verify this — it’s an inference from the benchmark numbers and XLA’s documented behavior.)
，这一点在heLLoword翻译中也有详细论述
Follow our Australia news live blog for latest updates
Последние новости
What to look for in an air purifier