AdaptPipe: Mitigating Runtime Bubbles via Granularity-Adaptive Scheduling under Memory Constraints

Crucially, pipeline parallelism (PP) is an indispensable parallelization strategy for training large models on multiple GPUs. Despite some proposed PP schedules theoretically promising zero bubbles, we highlight that runtime bubbles remain a significant practical hurdle in real-world model training. This discrepancy arises because these idealized schedules presuppose that F (forward), B (backward), and W (weight) operations have uniform and stable execution times. This assumption does not hold in practical scenarios.In this paper, we introduce AdaptPipe, an adaptive pipeline scheduling framework that mitigates runtime bubbles by leveraging our analysis of the characteristics of two types of runtime bubbles. AdaptPipe improves runtime efficiency by opportunistically filling W tasks of appropriate granularity. To achieve this, we develop a prediction method and a notification method to estimate the size of different runtime bubbles. Furthermore, we design a forward-backward (FB) scheme, namely nF –nB, which aims to provide AdaptPipe with sufficient schedulable W tasks to handle continuous or large bubbles while still adhering to memory constraints.Our experiments on dense models and sparse MoE models show that AdaptPipe improves throughput by 0.6%-13.8% over state-of-the-art static pipeline schedules while maintaining comparable activation memory usage.