60x Faster Cold Starts Treating Peer Gpus As Weight Servers

Runway reduced model-loading delays by letting a newly started GPU worker copy model parameters from a worker already holding them. Its NCCLBack engineering account describes replacing repeated cloud-storage downloads with direct GPU transfers, bringing reported startup times from minutes to seconds. Finding a nearby copy is only part of the work: peers must agree on the checkpoint, model configuration and how the model is divided across GPUs. The system checks peer responsiveness and samples transferred parameters for verification. Faster loading helps new workers become available during scaling and software rollouts.