FlashAttention
Attention computed in tiles to avoid materialising the full matrix.
Aperçu
FlashAttention computes attention in blocks that fit in fast on-chip memory, never writing the full attention matrix to high-bandwidth memory. It solves the problem that attention's cost is dominated by memory traffic rather than arithmetic, so a mathematically identical computation can be several times faster when scheduled differently. It exists as the clearest case of a systems change producing a large speed-up with no change to the model.
Graphe des technologies
Comment FlashAttention se rattache au reste de la pile.
Appliqué à
- Inference concept Infrastructure FlashAttention used for Inference High ✓ source
Outils
Logiciels qui implémentent FlashAttention ou s'y intègrent.