SkewAdam cuts MoE optimizer memory by 97.4% and fits 6.78B on a 40GB GPU
Kooky-Ad-4124 · reddit · 2026-07-22
A new preprint introduces SkewAdam, a tiered optimizer designed to slash the memory cost of Mixture-of-Experts training.
- The core idea is to allocate optimizer state by parameter type instead of treating everything equally.
- In the proposed setup, the backbone gets momentum plus factored second moment, experts get factored second moment only, and the router gets exact second moment.
- The paper reports optimizer-state memory falling from 50.6 GB to 1.29 GB, a 97.4% reduction.
- Peak training memory drops from 81.4 GB to 31.3 GB, allowing a 6.78B MoE to fit on a single 40GB GPU without hurting convergence or router stability.
Related event: SkewAdam Slashes MoE Optimizer Memory by 97.4%(2 posts)→
More from Infra
- Alchemy adds effect/Workflow support across Durable Objects, Workflows and Lambda Durable Functions — samgoodwin89 · 2026-10-27
- Alchemy's Lambda Web Functions run plain Node web servers on Lambda, no OpenNext needed — samgoodwin89 · 2026-10-27
- Foldkit Adds SSR Support With a Few Lines of Config on Cloudflare and AWS — samgoodwin89 · 2026-10-27
- Alchemy Cuts Cloudflare Bindings CPU Time by Up to 166× Via Profiling — samgoodwin89 · 2026-10-27
- Importing all 438 AWS services now takes 13.9 MB and 0.14s, down from 227 MB and 1.61s — samgoodwin89 · 2026-10-27
- Alchemy makes all ~2,850 AWS bindings work from any cloud, not just S3/Lambda/DynamoDB — samgoodwin89 · 2026-10-27