플래시 주의

Standard attention compares every token to every other token, producing an N-by-N score matrix that grows quadratically with sequence length. Naively, that matrix is written to and read back from GPU high-bandwidth memory (HBM), and that shuttling — not the multiplications — is the real bottleneck. Flash Attention, introduced by Tri Dao and colleagues in 2022, reorganizes the computation so the matrix is never fully stored. It processes queries, keys, and values in small tiles that fit in fast on-chip SRAM, computes partial results, and stitches them together using an online running-softmax trick. The algorithm computes the same attention operation while reducing high-bandwidth-memory traffic and intermediate storage. It can be faster, especially on long sequences, but measured gains depend on GPU, sequence length, dtype and implementation.