According to Google DeepMind on X, Decoupled DiLoCo investigates how to maintain continuous large scale training even when individual chips fail by decoupling strict synchronization across identical accelerators. As reported by Google DeepMind, frontier model training often stalls because a single device failure halts synchronized all-reduce steps; Decoupled DiLoCo aims to tolerate faults while preserving throughput. According to Google DeepMind, the approach explores relaxing lockstep coordination and allowing progress despite stragglers or dropouts, which could cut downtime and hardware underutilization in multi node GPU and TPU clusters. As reported by Google DeepMind, the business impact includes higher cluster efficiency, fewer restarts, and lower cost per training run for large language model and multimodal model training workloads that require thousands of accelerators.