ReviveMoE: Fast Recovery for Hardware Failures in Large-Scale MoE LLM Inference Deployments
Haley Li, Xinglu Wang, Cong Feng, Chunxu Zuo, Yanan Wang, Hei Lo, Yufei Cui, Bingji Wang, Duo Cui, Shuming Jing, Yizhou Shan, Ying Xiong, Jiannan Wang, Yong Zhang, Zhenan Fan · 2026
Read ReviveMoE: Fast Recovery for Hardware Failures in Large-Scale MoE LLM Inference Deployments by Haley Li, Xinglu Wang, Cong Feng, Chunxu Zuo, Yanan Wang, Hei Lo, Yufei Cui, Bingji Wang, Duo Cui, Shuming Jing, Yizhou Shan, Ying Xiong, Jiannan Wang, Yong Zhang, Zhenan Fan (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2602-21140.txt.
Description
Large language-model deployments fail often enough that recovery speed matters, yet restarting a serving instance reloads weights and recompiles graphs. Haley Li and fourteen co-authors propose a faster path in this 2026 arXiv paper: ReviveMoE, rapid failure recovery without restarting the instance. It supports both colocated architectures, where mixture-of-experts and attention share hardware, and disaggregated ones that separate them.
Built into a cloud model-as-a-service platform on an existing serving stack and communications library, ReviveMoE treats failure recovery as a systems problem rather than a restart problem. Readers get a design for keeping large inference fleets responsive when individual hardware inevitably fails, instead of paying full restart costs on every fault.
Catalog record
| CREATOR | Haley Li, Xinglu Wang, Cong Feng, Chunxu Zuo, Yanan Wang, Hei Lo, Yufei Cui, Bingji Wang, Duo Cui, Shuming Jing, Yizhou Shan, Ying Xiong, Jiannan Wang, Yong Zhang, Zhenan Fan |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 37 KB |
| VERSION | 1.0 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2602-21140