glassroom/sata_attention
View on GitHub ↗ ★ 36 Python
Reference implementation of "Self-Attention at Constant Cost per Token via Symmetry-Aware Taylor Approximation" (Heinsen and Kozachkov, 2026)
4 mentions on Hacker News by 3 people: 0 stories, 4 comments · first 2026-02-04 · latest 2026-02-07
20082026
-
…toy example is with 8 Taylor terms, applied to a context with 1B tokens, each token with 1K heads: https://github.com/glassroom/sata_attention That toy example is not practical with the quadratic formulation, because it would require…
comment by fheinsen · 2026-02-07 · in Attention at Constant Cost per Token via Symmetry-Aware Taylor Approximation -
…order per layer, per head, etc. The github repository lists many other possible improvements: https://github.com/glassroom/sata_attention#proof-of-concept
comment by fheinsen · 2026-02-04 · in Attention at Constant Cost per Token via Symmetry-Aware Taylor Approximation -
Reference implementation: https://github.com/glassroom/sata_attention
comment by NedCode · 2026-02-04 · in Attention at Constant Cost per Token via Symmetry-Aware Taylor Approximation -
Github here: https://github.com/glassroom/sata_attention
comment by spacewhales · 2026-02-04 · in Attention at Constant Cost per Token via Symmetry-Aware Taylor Approximation