🛠️ Automatic deduction of FlashAttention kernel from ~25 lines of SQL
Summary
This project automatically deduces a FlashAttention kernel from about 25 lines of SQL using three portable SQL extensions: xarray-sql, ddx, and einfold. It leverages relational algebra and a tensor data model to separate logical and physical layers in tensor programs.
Why it’s interesting
It demonstrates the automatic deduction of a complex FlashAttention kernel using only vanilla SQL and specialized relational extensions.
Target user
Machine learning systems developers and database researchers
Source metrics: Points 7 · Comments 0
HN discussion · Project
Source: #HackerNews / Show HN
Summary
This project automatically deduces a FlashAttention kernel from about 25 lines of SQL using three portable SQL extensions: xarray-sql, ddx, and einfold. It leverages relational algebra and a tensor data model to separate logical and physical layers in tensor programs.
Why it’s interesting
It demonstrates the automatic deduction of a complex FlashAttention kernel using only vanilla SQL and specialized relational extensions.
Target user
Machine learning systems developers and database researchers
Source metrics: Points 7 · Comments 0
HN discussion · Project
Source: #HackerNews / Show HN