DynaResize: Runtime GPU Reallocation for Disaggregated LLM Post-Training
DynaResize, an innovative system, tackles inefficiencies in pipeline processes during the post-training phase of RL-based LLMs by reallocating GPUs in real-time between the Rollout and Training stages. This method breaks down resizing into detailed operations and eliminates non-essential tasks from the critical path via communicator reuse, bounded state staging, and hysteresis-driven resizing. Experimental findings indicate a 66.5% enhancement in overall throughput and a 33% decrease in total execution time compared to the best static setup, while concealing 27% of the overhead associated with role-switching. The full paper can be found on arXiv.
Key facts
- DynaResize is a runtime GPU reallocation system for disaggregated LLM post-training.
- It dynamically switches GPUs between Rollout and Training stages.
- It decomposes resizing into fine-grained operations.
- It uses communicator reuse, bounded state staging, and hysteresis-based resizing.
- Experimental results show 66.5% improvement in end-to-end throughput.
- Total execution time reduced by 33% over optimal static configuration.
- 27% of role-switching overhead is hidden.
- The paper is submitted to arXiv.
Entities
Institutions
- arXiv