Multi-Head Attention Scaling and Per-Head Gain
SeunghyunSEO7 · x · 2026-07-16
During a discussion on multi-head attention, someone pointed out that many LLMs currently apply qnorm directly to qproj(x), rather than scaling each head's q/k by \sqrt(dhead).
Replies mentioned that the 1/dhead scaling concept originates from muP. The reason for following a Law of Large Numbers (LLN) intuition is that q and k might already be aligned from the start. Another viewpoint suggested that if q and k are first scaled by a factor like \sqrt{dhead} gain, their dot product becomes dhead (qgain, kgain) cos(θ), meaning the "per head gain" design warrants further scrutiny.
Related event: Researchers Debate 1/d-Style Attention Scaling in Modern LLMs(6 posts)→
More from Research
- Datology webinar: Curation Studio matched top data baselines with under 20% compute — josh_wills · 2026-10-27
- KANLab: an interactive browser playground for Kolmogorov-Arnold Networks — Lukas Niehaus · 2026-10-27
- CMU's LLM2Cert synthesizes stability certificates verified via SMT and Lean — CarnegieMellonU · 2026-10-27
- NVIDIA open-sources FLORA, a robotics data engine ~125x cheaper than GPT-6 Astra — nvidia · 2026-10-27
- RSNA 2026 Deep Learning Labs return: eight hands-on AI classes over four days in Chicago — PRouzrokh · 2026-10-27
- Steering two orthogonal concept neurons makes a model compose "gnome" — Sauers_ · 2026-10-27