DynaResize: Runtime GPU Reallocation for Disaggregated LLM Post-Training

2026-06-15

AI Summary

This work addresses the severe pipeline bubbles and low GPU utilization in reinforcement learning (RL) post-training of large language models caused by long-tailed rollout latencies under static GPU partitioning. To mitigate this, the authors propose DynaResize, a system that dynamically reallocates GPUs between rollout and training phases at runtime, balancing their execution times while preserving RL semantics. The key innovation lies in decomposing GPU reallocation into fine-grained operations and integrating communicator reuse, bounded state staging, and hysteresis control to eliminate non-computational overhead from the critical path. Experiments demonstrate that DynaResize improves end-to-end throughput by 66.5%, reduces total execution time by 33%, and hides 27% of role-switching overhead compared to the best static configuration.

Abstract

RL-based LLM post-training increasingly disaggregates Rollout and Training across separate GPU resources, but static GPU partitioning suffers from severe pipeline bubbles under long-tail rollout latency. We present DynaResize, a runtime GPU reallocation system that dynamically switches GPUs between Rollout and Training to balance stage execution times without changing RL semantics. DynaResize 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. Experimental results show that DynaResize can improve end-to-end throughput by 66.5% and reduce total execution time by 33% over the optimal static configuration, while hiding 27% of role-switching overhead.

Problem

• LLM post-training
• GPU disaggregation
• pipeline bubbles
• rollout latency
• static GPU partitioning

Innovation

• GPU reallocation
• disaggregated training
• runtime optimization
• pipeline balancing
• LLM post-training