FedQueue

FedQueue aims to improve wall-clock time-to-quality by predicting queue delays and adapting local work to a synchronization horizon. FedQueue achieved the best final test losses in the production deployment and was the only method to reduce loss below 0.4. FedQueue's advantage is not confined to synthetic queue models, according to the real-world LLaMA2-7B experiment. FedQueue combines online queue prediction, adaptive work budgeting, and staleness-aware admission and aggregation. FedQueue uses scheduler delay as an online signal for work budgeting, admission control, and aggregation in cross-facility federated learning.