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)→

Original post →

More from Research

Research channel →