We propose Generalized Primal Averaging (GPA), an extension of Nesterov's method in its primal averaging formulation that addresses key limitations of recent averaging-based optimizers such as single-worker DiLoCo and Schedule-Free (SF) in the non-distributed setting. These two recent algorithmic approaches improve the performance of base optimizers, such as AdamW, through different iterate averaging strategies. Schedule-Free explicitly maintains a uniform average of past weights, while single-worker DiLoCo performs implicit averaging by periodically aggregating trajectories, called pseudo-gradients, to update the model parameters. However, single-worker DiLoCo's periodic averaging introduces a two-loop structure, increasing its memory requirements and number of hyperparameters. GPA overcomes these limitations by decoupling the interpolation constant in the primal averaging formulation of Nesterov. This decoupling enables GPA to smoothly average iterates at every step, generalizing and improving upon single-worker DiLoCo. Empirically, GPA consistently outperforms single-worker DiLoCo while removing the two-loop structure, simplifying hyperparameter tuning, and reducing its memory overhead to a single additional buffer. On the Llama-160M model, GPA provides a 24.22% speedup in terms of steps to reach the baseline (AdamW's) validation loss. Likewise, GPA achieves speedups of 12% and 27% on small and large batch setups, respectively, to attain AdamW's validation accuracy on the ImageNet ViT workload. Furthermore, we prove that for any base optimizer with regret bounded by $O(\sqrt{T})$, where $T$ is the number of iterations, GPA can match or exceed the convergence guarantee of the original optimizer, depending on the choice of interpolation constants.

The recently proposed schedule-free method has been shown to achieve strong performance when hyperparameter tuning is limited. The current theory for schedule-free only supports a constant learning rate, where-as the implementation used in practice uses a warm-up schedule. We show how to extend the last-iterate convergence theory of schedule-free to allow for any scheduler, and how the averaging parameter has to be updated as a function of the learning rate. We then perform experiments showing how our convergence theory has some predictive power with regards to practical executions on deep neural networks, despite that this theory relies on assuming convexity. When applied to the warmup-stable-decay (wsd) schedule, our theory shows the optimal convergence rate of $\mathcal{O}(1/\sqrt{T})$. We then use convexity to design a new adaptive Polyak learning rate schedule for schedule-free. We prove an optimal anytime last-iterate convergence for our new Polyak schedule, and show that it performs well compared to a number of baselines on a black-box model distillation task.

Recently, Zamani and Glineur (2023) and Defazio et al. (2023) showed that the exact worst-case Our approach uses an alternative form of momentum that replaces traditional momentum. So from this viewpoint, the Schedule-Free updates can be seen as a version of momentum that has the same immediate effect, but with a greater delay for adding in the remainder of the gradient.

First-order methods underpin most large-scale learning algorithms, yet their classical convergence guarantees hinge on carefully scheduled step-sizes that depend on the total horizon $T$, which is rarely known in advance. The Schedule-Free (SF) method promises optimal performance with hyperparameters that are independent of $T$ by interpolating between Polyak--Ruppert averaging and momentum, but nonconvex analysis of SF has been limited or reliant on strong global assumptions. We introduce a robust Lyapunov framework that, under only $L$-smoothness and lower-boundedness, reduces SF analysis to a single-step descent inequality. This yields horizon-agnostic bounds in the nonconvex setting: $O(1/\log T)$ for constant step + PR averaging, $O(\log T/T)$ for a linearly growing step-size, and a continuum of $O(T^{-(1-α)})$ rates for polynomial averaging. We complement these proofs with Performance Estimation Problem (PEP) experiments that numerically validate our rates and suggest that our $O(1/\log T)$ bound on the original nonconvex SF algorithm may tighten to $O(1/T)$. Our work extends SF's horizon-free guarantees to smooth nonconvex optimization and charts future directions for optimal nonconvex rates.

Existing learning rate schedules that do not require specification of the optimization stopping step T are greatly out-performed by learning rate schedules that depend on T. We propose an approach that avoids the need for this stopping time by eschewing the use of schedules entirely, while exhibiting state-of-the-art performance compared to schedules across a wide family of problems ranging from convex problems to large-scale deep learning problems. Our Schedule-Free approach introduces no additional hyper-parameters over standard optimizers with momentum. Our method is a direct consequence of a new theory we develop that unifies scheduling and iterate averaging.