BubbleSpec: Turning Long-Tail Bubbles into Speculative Rollout Drafts for Synchronous Reinforcement Learning

Reinforcement Learning (RL) has become a cornerstone for improving the performance of Large Language Models (LLMs). However, its rollout phase constitutes a significant efficiency bottleneck, mainly arising from the long-tail bubbles across data parallel ranks, particularly in long-context scenarios where faster GPUs remain idle while waiting for stragglers. Existing solutions, such as partial rollout or asynchronous RL, mitigate these bubbles by compromising the algorithm's strict synchronous nature. Instead, we propose **BubbleSpec**, a novel framework that accelerates RL rollouts while strictly keeping the mathematical exactness. Instead of attempting to eliminate bubbles, BubbleSpec exploits them. We exploit the idle time windows of faster ranks to pre-generate rollout results for subsequent steps, serving as drafts for speculative decoding. Unlike prior speculative methods that rely on historical epoch similarity and warm-ups, BubbleSpec is agnostic to dataset size and provides immediate acceleration from the onset of training. Extensive evaluations demonstrate that BubbleSpec reduces decoding steps by **$\sim$50\%** and increases rollout throughput by up to **1.8$\times$**. Critically, BubbleSpec is seamlessly compatible with various RL frameworks and strategies as it sustains the strict synchronous property of RL algorithms.

A major challenge in training large language models with reinforcement learning is that expensive GPUs often sit waiting instead of working. In synchronous training, some machines finish generating answers quickly, while others take longer, creating “dead time” that slows the whole system down. Our method, BubbleSpec, makes use of this dead time instead of wasting it. When a GPU finishes early, it uses the spare time to prepare possible future pieces of text. These guesses are then efficiently checked and reused to speed up later generation. This approach preserves the exact behavior of the original training method, which is important for keeping training stable and reliable. In experiments on several models, BubbleSpec reduced the number of decoding steps by around 50% and increased rollout speed by up to 1.8×. By making reinforcement learning training faster and cheaper, our work helps researchers use computing resources more effectively.