AdaLRS: Loss-Guided Adaptive Learning Rate Search for Efficient Foundation Model Pretraining

Learning rate is widely regarded as crucial for effective foundation model pretraining. Recent research explores and demonstrates the transferability of learning rate configurations across varying model and dataset sizes, etc. Nevertheless, these approaches are constrained to specific training scenarios and typically necessitate extensive hyperparameter tuning on proxy models. In this work, we propose AdaLRS, a plug-in-and-play adaptive learning rate search algorithm that conducts online optimal learning rate search via optimizing loss descent velocities. We provide theoretical and experimental analyzes to show that foundation model pretraining loss and its descent velocity are both convex and share the same optimal learning rate. Relying solely on training loss dynamics, AdaLRS involves few extra computations to guide the search process, and its convergence is guaranteed via theoretical analysis. Experiments on both LLM and VLM pretraining show that AdaLRS adjusts suboptimal learning rates to the neighborhood of optimum with marked efficiency and effectiveness, with model performance improved accordingly. We also show the robust generalizability of AdaLRS across varying training scenarios, such as different model sizes, training paradigms, base learning rate scheduler choices, and hyperparameter settings. 1 max(λ t β,1) , where α, β > 1 are two multiplicatively independent real numbers and λ ∈ (0, 1) is a decay factor. We validate the multiplicatively independent design of LR scaling factors (for all integers m, n, α m = β n =⇒ m = n = 0) in Appendix B.

Workflow. During model training, AdaLRS monitors the loss curve slope v t with the least squares method [7], and attempts to upscale the learning rate when the loss curve slope decays. After the upscaling adjustment, we compare the loss curve slope with that before upscaling. As shown in Equation 1, if the estimated loss slope increases more than 2e after the adjustment, the upscaling is regarded as valid and the adjustment will be retained. On the other hand, once the validation fails