The Spack package manager is quite popular in the HPC / supercomputer space for scientific software. Even with the more selective niche than a typical general purpose OS package manager, large language models (LLMs) have already proven capable of being useful in generating new Spack packages. But there have also been some headaches involved too for Spack developers.
Caetano Melone with Lawrence Livermore National Laboratory presented last month at the High Performance Software Foundation (HPSF) conference in Chicago on LLM-written Spack packages. Long story short, with some “slight nudging” LLMs have proven useful to having AI tools write Spack packages. The LLM exploration has also revealed new opportunities for ways to improve Spack itself.
Spack developers found that using LLMs for writing packages was quite possible given sufficient context and structure provided to the large language model, or as one of the slides in the presentation put it: “LLMs are capable; they need structured guidance to perform reliably.”
LLMs can be a time-saver but the challenge with other software projects too is avoiding any additional burden on the upstream maintainers itself if the generated LLM assets aren’t good enough and the one interacting with the LLM having enough knowledge for correctly verifying the output. In any case the LLM-generated Spack packages can be effective and Caetano Melone argued that with LLMs can be valuable tools with the right structured inputs, representative examples, and human oversight.
Those wanting to learn more about LLM-generated Spack packages can do so via Melone’s presentation embedded below and more details over on the HPSF Conf 2026 site.
