Reproduce “Robust Scheduling with GFlowNets” on OpenResearch

About this paper

David W. Zhang, Corrado Rainone, Markus Peschl, Roberto Bondesan

A generative flow network (GFlowNet) approach is introduced for robust compiler scheduling, producing diverse sets of high-performing schedules to mitigate inaccuracies in proxy performance metrics. The method achieves superior average target speedups and diversity over traditional heuristics and other ML-based methods on various computation graphs.

Key insights

• Generating a diverse set of high-quality schedules, as identified by a proxy, significantly improves robustness against errors in the proxy model when selecting for actual hardware performance.
• Generative Flow Networks can be effectively adapted for conditional generation of complex discrete structures like schedules, especially with training objectives tailored for varying conditional partition functions.
• A single neural policy can be conditioned on a temperature parameter to control the exploration-exploitation balance at inference time, providing flexibility in schedule generation.

Summary by alphaXiv.