Workshops Archive

Coconut: An HPC-Native Inference Framework for Neural Surrogates in MPI Simulations


Workshop: AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications

Authors: Wenke Du (INRIA, IFPEN); Jean-Marc Gratien and Raphaël Gayno (IFPEN); and Bruno Raffin (INRIA)

Abstract: Neural surrogates can accelerate hybrid PDE solvers, but whether a surrogate delivers a speedup can only be determined in situ, because deployment configuration strongly affects its performance, and current HPC infrastructures support this exploration poorly. We present Coconut, an HPC-native inference framework that fits the MPI programming model and exposes deployment options at two levels: batch size, numerical precision, device placement, and inference backend at the model level, and process topology at the architecture level. A microbenchmark quantifies the cost of the embedded Python interpreter required by the PyTorch Inductor backend, showing when this model-safe execution path becomes efficient. In an industrial thermal multiphase-flow simulator, a thermodynamic surrogate moves from a 2.4× kernel-level slowdown to a 34.12× speedup over the original kernel and 2.51× over a multithreaded batched baseline, carrying through to application-level throughput gains. Deployment choices decide whether a surrogate pays off; Coconut enables their in situ evaluation.


Back to AI4S: 8th Workshop on Artificial Intelligence and Machine Learning for Scientific Applications Archive Listing

Back to Full Workshop Archive Listing