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
- SGLang MegaKernel + PD Pushes the TCO-Latency Frontier on GB300, SemiAnalysis Benchmarks Show — AccBalanced · 2026-10-27
- Anthropic hits $16.3B H1 revenue as compute efficiency jumps 5.2x in 18 months — daniel_mac8 · 2026-10-27
- Tenstorrent runs Proteina protein model at high throughput, claims no HBM shortage — gajesh · 2026-10-27
- YC-backed Hopper launches voice-agent inference cloud: 80 ms TTFT at $0.30/1M tokens — ycombinator · 2026-10-27
- A100 rental prices still rising in 2026 per a16z data, Jevons paradox hits compute — VraserX · 2026-10-27
- Anthropic S-1 reveals $800B+ in off-books future compute commitments — GaryMarcus · 2026-10-27