Dynamic Node Scheduling for Delay Optimization in Federated Large Language Model Training

During the training process of Large Language Model, Federated Learning is introduced to solve the trade-off between data-sharing requirements and privacy protection. Due to the large number of model parameters in the large language model and the distributed architecture of federated learning, training delay cost produced by nodes during the training process is large and is an important issue to address. This paper focuses on minimizing the total training delay by balancing the local training delay of participating nodes through dynamic node scheduling. We propose a dynamic scheduling algorithm LTQA based on Bi-directional LSTM model and reinforcement learning theory. During the training process, we find the node with the highest delay cost in each round using a Bi-directional LSTM model. Reinforcement learning theory is used to schedule resources and tasks with other nodes for cooperation, balancing the local training delay of each node according to dynamic environment changes. We evaluate the performance of the algorithm through several experiments on two widely-used large language model. It is found that compared with state-of-the-art FL algorithms, LTQA can save training delay cost by 7.17%-25.8%, and the model accuracy is improved by 2.59%-13.62%.