Paper Overview
- Research area: Training optimization
- Authors: Philip Zmushko, Egor Petrov, Nursultan Abdullaev
- Published: 2026-07-01
- arXiv: 2507.00005
Summary
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.
This work challenges that assumption, demonstrating that degradation under one-step delay depends strongly on optimizer choice rather than being an intrinsic limitation. The authors provide the first comprehensive empirical analysis showing that while AdamW — dominant when PipeDream-2BW was introduced — indeed suffers severe degradation, recent optimizers such as Muon exhibit strong robustness under one-step delay.
They further introduce an optimizer-agnostic, error-feedback-inspired correction to mitigate delay effects, and supply supporting theoretical analysis proving that Muon converges both with and without this correction.
Extensive evaluations on models up to 10B parameters confirm that these strategies bridge the performance gap with synchronous training, underscoring the practical potential of large-scale asynchronous pipeline parallelism.
--- *Auto-collected on 2026-07-01*