Show HN: Automatic deduction of FlashAttention kernel from ~25 lines of SQL
Summary
Show HN detailing the automatic deduction of a FlashAttention kernel from about 25 lines of SQL within xarray-sql. It describes how FlashAttention is realized via streaming blocks in SQL, how einfold rewrites the plan to fuse softmax-weighted operations, and includes forward, gradient tests, and a Nadarya-Watson control for comparison.