Nous Research Pre-Trains 15B Model with DeMo and Heterogeneous Hardware

Dec 2, 2024, 05:03 PM

Nous Research has announced the successful pre-training of a 15-billion parameter language model using distributed training over the internet. The model, trained with Nous DisTrO (now DeMo), utilized heterogeneous hardware contributed by partners such as Oracle, LambdaAPI, Northern Data Group, Crusoe Cloud, and the Andromeda Cluster. The training achieved a loss curve and convergence rate that meets or exceeds those of centralized training methods, demonstrating the viability of decentralized training at scale.