New AI Training Trick Cuts Months Off Model Development

ASP is an offline rollout scheduler for multi-instance asynchronous reinforcement learning of large language models. It uses a single offline preprocessing pass to jointly reduce padding bubbles from intra-minibatch length heterogeneity, skewness bubbles from long-tail outputs, and sorting bubbles from fluctuations in training-queue occupancy caused by length-sorted dispatch.

The complete ASP improves end-to-end throughput by up to 3.22 times and reduces iteration time by up to 76.8% relative to MindSpeed-RL, while reducing iteration time by up to 62.5% relative to verl. Within ASP, scheduling reduces mean iteration time by 10.9%.

It also scales more robustly with additional generation resources than those baselines and shows no visible convergence degradation in the measured bounded-lag regime. In an internal telecom-domain RL post-training workflow, ASP reduced the reported training period from more than three months to about two months and produced a domain-enhanced model for knowledge answering and tool calling.