SkewAdam cuts MoE optimizer state from 50.6 GB to 1.29 GB on a 6.78B model
Nuemaan Malik · hf · 2026-07-22
A new MoE optimization paper argues that where optimizer state lives matters almost as much as how much of it there is.
The paper studies SkewAdam, which separates MoE parameters into three groups with different state needs:
- Dense backbone: float32 momentum plus factored second moment
- Experts: factored second moment only
- Router: exact second moment
On a 6.78B-parameter MoE language model, this cuts optimizer state from 50.6 GB to 1.29 GB, reducing peak training memory from 81.4 GB to 31.3 GB — low enough for a 40 GB accelerator.
In controlled runs over 82M tokens, SkewAdam reaches validation perplexity 108.4, better than AdamW (126.8), Muon (120.2), and Lion (393.7). The paper also reports that the memory savings come from the tiered allocation, while the accuracy gains come mainly from keeping momentum.
Related event: SkewAdam Slashes MoE Optimizer Memory by 97.4%(2 posts)→
More from Infra
- Anthropic locked into $475B compute deals, OpenAI lifts projected spend to ~$750B through 2030 — kyliebytes · 2026-10-27
- First systematic data center noise study: 17 of 25 hyperscalers drew no credible complaints — AndyMasley · 2026-10-27
- Fork builds engine to run Qwen at ~400 tok/s on three used V100s — jfinkpottery · 2026-10-27
- RFD4-Proteina runs on an RTX 3060: protein design from $0.001 per design — AllThingsApx · 2026-10-27
- NPM packages now take ~20 minutes to go live, dev suspects AI-attack throttling — cocktailpeanut · 2026-10-27
- Nature Physics: reprogrammable refractive index enables arbitrary multimode wave control for ML — MarioKrenn6240 · 2026-10-27