Outer-Momentum Restarting in High-Dimensional Two-Phase Optimization

To make large-scale distributed training practical outside high-bandwidth datacenters, we must reduce blocking, high-volume synchronization. While DiLoCo communicates infrequently, its outer synchronization remains bandwidth-heavy and brittle to stragglers and transient failures. We relax exact synchronization to approximate synchronization via mixing/gossip, which degrades gracefully under delays and communication failures. This allows us to factorize DiLoCo synchronization into a non-blocking mixing step that overlaps computation with no staleness, and a blocking mixing step that tightens worker agreement, yielding a tunable trade-off between compute utilization and optimization stability. On up to billion-parameter language models in low-bandwidth settings, our framework substantially improves compute utilization compared to DiLoCo, with training progress ranging from comparable to closely matching it, and is more robust to failures.

Training large language models is generally done on clusters containing thousands of accelerators, communicating over a high-bandwidth interconnect. Scaling up these clusters is expensive and can become impractical, imposing limits on the size of models that can be trained. Several recent studies have proposed training methods that are less communication intensive, avoiding the need for compute clusters with extremely high interconnect speeds. These low communication training methods still employ a global synchronization step for model parameters, which can be too costly with a high number of participants, as the communication cost scales quadratically with group size. In this work, we propose a novel optimization method, NoLoCo, that does not explicitly synchronize all model parameters during training and does not require any collective communication. NoLoCo implicitly synchronizes model weights via a novel variant of the Nesterov momentum optimizer by partially averaging model weights within randomly selected subgroups. We provide both a theoretical convergence analysis of our optimizer and empirical results from language model training. Our method requires significantly less communication than fully sharded data parallel training and DiLoCo, a widely used low-communication baseline. Moreover, our method avoids global blocking communication, thereby reducing accelerator idle time. Our experiments show that NoLoCo is more communication-efficient than DiLoCo, improving final perplexity by up to $4\%$ and converging up to $4\times$ faster in wall-clock time across a range of worker counts, model sizes, and communication bandwidths.

Asynchronous DiLoCo systems may receive pseudo-gradients computed several outer rounds earlier, yet the standard Nesterov outer optimizer does not explicitly condition its update on per-update age. This can make the outer momentum buffer brittle under large controlled delays. We propose Cosine Gated Adam Decay (CGAD), a simple, drop-in, age-aware outer optimizer that scales each incoming pseudo-gradient by $σ(τ) = γ(τ) e^{-ατ}$ before it enters Adam's first- and second-moment buffers; the exponential models information decay and the cosine gate $γ(τ)$ smoothly zeroes contributions past a chosen cutoff. CGAD reduces to plain Adam at $τ=0$, adds two hyperparameters whose defaults transfer across scales, and extends to partial-sync schedulers via a per-fragment age-aware variant (PA-CGAD). For an idealized gated-adaptive update on smooth non convex objectives, we prove a non-asymptotic convergence bound whose staleness-bias term depends on $α$ alone, rather than on the realized maximum delay $τ_{\max}$; standard analyses of asynchronous momentum-SGD instead carry a $τ_{\max}^2$ factor. Empirically, on Llama style language model pretraining at 25M, 1B, and 7B parameters, CGAD trains stably across the controlled delays we sweep. The cosine cutoff acts as scale insurance: the closest baseline, Adam Decay (CGAD without the cutoff), is competitive at 25M but its seed-to-seed $σ$ at $τ=8$ grows 27x from 25M to 7B, pushing its single-shot risk (mean + $σ$) above the chance-level loss while CGAD's stays well below. The published Nesterov recipe is the least stable method on the full sweep.