Flash attention

Flash attention: An input-output-aware algorithm that computes exact attention without storing the full attention matrix in memory, greatly speeding training and inference.

Browse the full glossary →