verl-pipe A reinforcement-learning training extension that pipelines model sampling, reward calculation, and training updates to reduce idle compute during language-model post-training.