Discussion on Per-Head Scaling in Multi-Head Attention
stochasticchasm · x · 2026-07-16
This reply discusses a parameterization detail in multi-head attention:
- The other party noted that in many LLMs, qnorm is simply applied to qproj(x), and the q/k vectors of each head are not scaled by \sqrt(dhead).
- The question raised here is why we don't use "per head independent gain" (e.g., nn.RMSNorm((nhead, dhead), ...)) to achieve per head gain.
- The response suggests this is related to the muP concept.
- Another reply adds that explaining attention's 1/d scaling based on per-head qk normalization could itself be a novel idea.
Related event: Researchers Debate 1/d-Style Attention Scaling in Modern LLMs(6 posts)→
More from Research
- Diamandis: AI Cancer Navigator and $100 Genome Sequencing Close the Care Gap — PeterDiamandis · 2026-10-27
- EMNLP Oral: training-free Nexus Sampling cuts KV-cache memory 10x; OpenResearcher tops BrowseComp-Plus at 54.8% — yuz9yuz · 2026-10-27
- ICML 2026 position paper: LLM 'thinking traces' are a harmful anthropomorphism — rao2z · 2026-10-27
- Forking gap follows a power law in context frequency; masking rare n-grams shrinks it — ZimingLiu11 · 2026-10-27
- How forking happens: n-gram tables split logits into non-interacting subspaces — ZimingLiu11 · 2026-10-27
- The best LoRA explainer video yet: rank constraints meet intrinsic dimension — sophiamyang · 2026-10-27