disaggregated-training

DynaResize is a runtime GPU reallocation system that dynamically switches GPUs between Rollout and Training stages during RL-based LLM post-training, improving throughput by 66.5% and reducing execution time by 33% over static configurations.