Reproduce “Architecture-Aware Mixed Precision for Replica-Exchange Training of Geometry-Aware Neural Operators with Uncertainty Quantification” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Ujin Jeon, Michael Penwarden, Wenqian Chen, Guang Lin

The authors introduce AA-MP, a role-based precision policy for replica-exchange training of ArGEnT–DeepONet. It keeps the cold refinement replica in FP32 and uses BF16 computation in the hot exploratory replica, while retaining FP32 updates and reductions. On AirfRANS, compared with FP32/FP32, it reduced hot-replica peak GPU memory by 46.9% and step time by 25.0%, with comparable accuracy; these are not whole-job savings.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.