PEARC DragonHPC Tutorial

Incorporating AI capabilities into traditional HPC workflows introduces many challenges for application developers, as it brings together two very different computational paradigms. Python can serve as a common language to bridge these domains, but, as a general purpose language, the exact interfaces to scientific applications, workflow design motifs, and runtimes that enable performance and scalability are still being explored and developed. This tutorial will help attendees develop skills for building AI-coupled HPC workflows using the DragonHPC project along with GPU-accelerated tools like PyTorch and vLLM. Attendees will first learn the fundamentals of using DragonHPC for Python multiprocessing on GPUs across multiple nodes. Hands-on exercises will then introduce them to aspects of the DragonHPC API for orchestrating both Python and traditional HPC (MPI-based) applications, leveraging GPU acceleration, coupled through fast data exchange with DragonHPC’s in-memory distributed dictionary. From there, we will explore a number of execution and data-transfer motifs common to hybrid AI-HPC workflows. attendees will interact with lead developers from DragonHPC and learn about its applications in the scientific workflows community.