FLASHATTENTION 3

Summary from the TLDR AI newsletter

FlashAttention is the foundation of modern fast language models. This new version uses 75% of H100 capacity (up from 35% before). There are a number of important system improvements that led to this capability gain.