Getting the Best of Both Worlds: Bridging Fortran and Pytorch for the CPU and the GPU HPC Simulations
DescriptionOver the past years, machine learning (ML) has been attracting rapidly increasing interest in the computational science. Many developers are adding ML models to their traditional simulation pipelines, and yet more are willing to follow. However, integration of such models into an existing application is often a technical challenge, which hinders research progress and demands domain scientists to deal with the programming issues. These complications are especially pronounced when the application is written in a language uncommon and not well supported in the ML community, such as Fortran. We present a lightweight library that enables seamless invocation of the Pytorch models from the high-performance Fortran codes. Users can run zero-copy inference and training from within their Fortran applications with the model defined via Pytorch Python scripts. The library supports CPU and GPU backends and is compatible with OpenACC-accelerated codes. We believe that our work could facilitate ML research in the Fortran computational community.
TimeWednesday, June 2814:30 - 15:00 CEST
Climate, Weather and Earth Sciences