Base-2 vs. base-e log-sum-exp mismatch silently corrupts attention in SGLang
By apsys · 2026-08-16 · 1 points · 0 comments
https://twitter.com/advprop/status/2089046843690729620
By apsys · 1 points · 0 comments · on Hacker News, read on BetterNews.
Open the full discussion on BetterNews