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
- Qwen3.8-Max takes No. 1 on FlashInfer after 500 runs and 30K tool calls — YouJiacheng · 2026-07-22
- Unsloth says it can fine-tune 7B models on a single RTX 4090 with 70% less VRAM — thisdudelikesAI · 2026-07-22
- Open-weight models may raise hardware demand by unlocking new AI workloads — bookwormengr · 2026-07-22
- Chinese model quality is no longer the surprise; a 2026 domestic compute cluster would be — teortaxesTex · 2026-07-22
- Mark Cuban says many AI data centers may end up as pickleball courts — 2C_ornot2C · 2026-07-22
- LLM inference benchmarks can mislead teams before production traffic hits — Suspicious_Orchid770 · 2026-07-22