One-Step Gradient Delay is Not a Barrier for Large-Scale Asynchronous Pipeline Parallel LLM Pretraining

Modern large-scale LLM pretraining benefits from utilizing Pipeline Parallelism; however, synchronous implementations leave GPUs idle during pipeline bubbles, wasting computational resources. Asynchronous Pipeline Parallelism eliminates these bubbles, maximizing throughput at the cost of gradient staleness. Among asynchronous schedules, PipeDream-2BW is particularly appealing: unlike the original PipeDream schedule, it ensures a constant one-step gradient delay regardless of pipeline depth. However, its adoption remains limited due to the common belief that optimizing under staleness is fundamentally unstable. In this work, we challenge this assumption, demonstrating that degradation under one-step delay depends strongly on optimizer choice rather than being an intrinsic limitation. We provide the first comprehensive empirical analysis showing that while AdamW, the predominant optimizer at the time when PipeDream-2BW was introduced, indeed suffers from severe degradation, recent methods like Muon exhibit strong robustness under a one-step delay. We introduce an optimizer-agnostic Error Feedback-inspired correction to further mitigate delay effects. We provide supporting theoretical analysis demonstrating convergence for Muon with and without this correction. Extensive evaluation on models up to 10B parameters confirms that our strategies bridge the performance gap with synchronous training, highlighting the practical potential of asynchronous pipeline parallelism at scale.

Training very large language models requires spreading the work across many GPUs. One common way to do this makes some GPUs wait while others finish their part, wasting expensive computing time. An alternative is to let GPUs keep working without waiting, but then parts of training use slightly outdated information, which has often been viewed as risky. We show that this risk depends strongly on the training method. Older standard choices can perform poorly with outdated information, while newer methods such as Muon remain much more stable. We also propose a simple correction that helps compensate for this outdated information and further closes the gap with standard training. We test this across many model sizes, including a large 10-billion-parameter model, and find that asynchronous training with our correction can match the final quality of standard training. These results suggest that large language models can be trained more efficiently without sacrificing model quality.