Accelerating RL Post-Training Rollouts via System-Integrated Speculative Decoding

TL;DR

This paper introduces a speculative decoding method integrated into RL training systems to significantly accelerate language model rollouts, achieving up to 2.5x speedup at large scales.

Contribution

It demonstrates the first system-optimized speculative decoding approach for RL post-training, compatible with existing models and pipelines, boosting throughput without loss of fidelity.

Findings

Speculative decoding improves RL rollout throughput by 1.8x at 8B scale.

Combining speculative decoding with asynchronous RL can yield up to 2.5x end-to-end training speedup at 235B scale.

Implementation in NeMo-RL with vLLM backend supports both synchronous and asynchronous pipelines.

Abstract

RL post-training of frontier language models is increasingly bottlenecked by autoregressive rollout generation, making rollout acceleration a central systems challenge. Many existing efficiency methods improve throughput by changing the rollout or optimization regime, for example, through off-policy execution, replay, or lower-precision generation. We study speculative decoding as a lossless acceleration primitive for RL rollouts that preserves the target model's output distribution. We implement speculative decoding in NeMo-RL with a vLLM backend, supporting both synchronous and asynchronous pipelines and enabling speculation during RL rollouts. This benefit is realizable across speculation mechanisms, such as pretrained MTP heads, small external draft models or even techniques such as Eagle3, which are traditionally applied after RL phase. This yields a deployment path for state-of-the-art speculative decoding inside RL training. In a reasoning post-training workload at 8B scale under synchronous RL, speculative decoding improves rollout throughput by 1.8x. Using a high-fidelity performance simulator, we project that combining speculative decoding with asynchronous RL yields up to 2.5x end-to-end training speedup at 235B scale.