15-year-old builds attention kernel: no OOM at 32k, 11x faster

A 15-year-old developer released a custom attention kernel that keeps early tokens as memory anchors and uses entropy-decayed rolling windows for the rest, avoiding OOM at 32k tokens while running 11x faster than standard attention on a Tesla T4.

2026-11-04 ~ 2026-11-04 · 2 related posts

• 15-year-old's attention kernel survives 32k tokens where standard PyTorch OOMs, 11x faster — vankownlaw · 2026-11-04

1 near-duplicate retellings: vankownlaw