Workshops Archive

LLMs as Packagers of HPC Software


Workshop: AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC

Authors: Caetano Melone, Daniel Nichols, Konstantinos Parasyris, Todd Gamblin, and Harshitha Menon (Lawrence Livermore National Laboratory)

Abstract: High performance computing (HPC) software ecosystems are heterogeneous, with applications depending on hundreds of external packages across diverse build systems and configuration options. Package managers such as Spack automate dependency resolution, but rely on manually authored recipes encoding intricate build knowledge that becomes increasingly labor-intensive to maintain. Large language models (LLMs) have shown promise for software engineering tasks, but synthesizing HPC package recipes remains challenging: the output must capture dependency, configuration, and platform constraints that are often fragmented across repositories and prior packaging practice.

We present a systematic study of what context and feedback LLMs need to synthesize valid Spack recipes, isolating the effects of repository metadata extraction, reference retrieval, and iterative repair. Structured build information improves installation success from approximately 43% in minimal-context settings to 100% in the best configuration, showing that metadata distillation, reference packages, and diagnostic feedback are essential for HPC package synthesis.


Back to AgenticAI4HPC’26: The First International Workshop on Agentic AI for HPC Archive Listing

Back to Full Workshop Archive Listing