FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness

This is part of the architecture of MosaicML (https://news.ycombinator.com/item?id=35829800).