Hacker NewsHN submitter: alxmrs
Original published:

Show HN: Automatic deduction of FlashAttention kernel from ~25 lines of SQL

A demo that automatically deduces a FlashAttention kernel from about 25 lines of SQL expressing vanilla attention, utilizing three portable SQL extensions: xarray-sql, ddx, and einfold.

Key points

  • •It uses the xarray-sql library to efficiently represent petabyte-scale array datasets as tables in the cloud.
  • •It assumes the ddx database extension, which adds JAX-style automatic differentiation to standard SQL.
  • •It incorporates einfold, a modular query planner designed for fast query processing.
#LLM / Model Trends#Research & Papers#FlashAttention#xarray-sql#ddx
See more