LiveResize: Runtime GPU Reallocation for Disaggregated LLM Post-Training

Abstract

RL-based LLM post-training increasingly disaggregates rollout and training across separate GPU resources, but static GPU partitioning can suffer from pipeline bubbles under long-tail rollout latency. We present LiveResize, a runtime GPU reallocation system that dynamically switches GPUs between rollout and training to balance stage execution times without changing RL semantics. LiveResize decomposes resizing into fine-grained operations and removes non-startup-critical work from the critical path through communicator reuse, bounded state staging, and hysteresis-based resizing. LiveResize achieves a 1.26× throughput end-to-end speedup over a representative fixed 6:2 train/rollout baseline, compared with 1.24× for our unoptimized in-job resize baseline, while reducing isolated role-switching latency by 27%.