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
- Engram's random reads don't suit SSDs; CPU-memory over NVLink could serve all 72 GPUs — bookwormengr · 2026-09-11
- 80% of the DIY LLM inference hype posters have already quit — it's brutally hard systems work — abhijithneil · 2026-09-11
- Hugging Face's Ultra Scale Playbook: a free book on training LLMs on GPU clusters — mdancho84 · 2026-09-11
- Is inference latency becoming the biggest bottleneck for production AI agents? — Euphoric_Sea632 · 2026-09-11
- LLM Serving Metrics Thread: Why TPOT and Uptime Make or Break User Experience — abhijithneil · 2026-09-11
- PlanetScale launches sharded Postgres: 768 servers acting as one, 1PB scale — dhruv2038 · 2026-09-11