Speculative decoding brought into RL rollout at 122B scale

An end-to-end system for online draft-model co-training during RL post-training: rank-local branch attention merged into compact zigzag-ring attention (context parallelism), and a separate TapChannel path moves intermediate target features across pipeline stages. The system achieves significant rollout and end-to-end acceleration on 122B-scale models with strong context-parallel scaling at 256K tokens. Code is available (link in the paper).

RL rollout generation is the dominant cost of RL post-training, and long-context rollouts have been the hardest to accelerate. This shows the speculative-decoding machinery surviving contact with a real distributed training stack at a scale and context length most teams cannot run.