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

2026-04-29

AI Summary

This work addresses the computational bottleneck posed by autoregressive rollout generation in post-training reinforcement learning. It systematically integrates state-of-the-art speculative decoding techniques—including Eagle3, pretrained multi-token prediction (MTP) heads, and compact draft models—into the RL training pipeline using the NeMo-RL framework with a vLLM inference backend. The approach enables both synchronous and asynchronous rollout inference while preserving the target model’s output distribution. Experimental results demonstrate a 1.8× increase in rollout throughput for synchronous training with an 8B-parameter model. When combined with asynchronous mechanisms, the method is projected to accelerate end-to-end training by up to 2.5× at the 235B-parameter scale.

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.

Problem

• RL post-training
• rollout acceleration
• autoregressive generation
• system bottleneck
• throughput

Innovation

• speculative decoding
• reinforcement learning post-training
• rollout acceleration
• asynchronous RL
• language model training