Yhn SportBaseball first seen 1 d ago, last 15 min ago, peak #3
FlashAttention kernel generated automatically from 25 lines of SQL
Original: Show HN: Automatic deduction of FlashAttention kernel from ~25 lines of SQL
A new demonstration shows FlashAttention, the memory-efficient attention kernel widely used in transformer models, being automatically deduced from roughly 25 lines of SQL. The project, shared via the xarray-sql repository on GitHub, compiles a high-level SQL description into an optimized kernel, suggesting a compiler-driven approach to writing GPU code. Readers are debating how far declarative languages like SQL can go in replacing hand-tuned CUDA kernels for machine learning workloads.
Why now: The claim that a complex, hand-optimized GPU kernel can be derived from a few lines of SQL challenges assumptions about how ML kernels must be written.
FlashAttentionxarray-sqlGitHubHacker News
Evidence
API: https://socialmediatrends-api.osmike.com/v1/trends/4283000