Nereus is a cost-aware runtime that adapts reinforcement-learning post-training jobs to changing GPU cluster conditions (resource availability, sequence length, memory pressure, stage bottlenecks) by dynamically switching execution plans mid-run. It represents each model-stage replica as an ‘Elastic Model Unit’ and uses a global transition graph plus a calibrated cost model to decide when a plan swap is worth the transition cost and to coordinate GPU transfers. In trace-based evaluation, online tensor/pipeline-parallel adaptation cut average step latency by 27.7% versus a fixed layout, with transition overhead of only 0.079% of total run time at 1,024-GPU scale. It improves end-to-end 8B-model PPO throughput by 2.14-7.27x over OpenRLHF and 1.10-1.47x over Verl across diverse clusters.
