Artifact for CGO 2027: "Polydron: Legality-Guided Code Optimization with a Specialized Reasoning LLM"
Artifact of the CGO 2027 paper "Polydron: Legality-Guided Code Optimization with a Specialized Reasoning LLM". polydron-artifact.zip: the Polydron search system with the Tiramisu compiler interface (TiraLib), the 150 Tiramisu PolyBench kernels and PolyBench/C 4.2.1, the logs of the paper's runs, analysis and run scripts, README. polydron-env.sif: Apptainer/Singularity container with GCC 9.2, Python 3.10, and a prebuilt Tiramisu. polydron-env-docker.tar.gz: the same environment as a Docker image (docker load -i polydron-env-docker.tar.gz). polydron-llm.tar: Polydron-LLM weights (Hugging Face format, float16, 28 GB; needed only for live runs on a GPU). Requires an x86-64 CPU with AVX (any recent Intel or AMD CPU). Start with README.md in polydron-artifact.zip and run_test.sh.
Authors
- Syed Umair Hussain Shah (ORCID: https://orcid.org/0009-0006-9463-1875)
Institutions
- New York University (US)
Publication Details
- Journal
- Zenodo (CERN European Organization for Nuclear Research)
- Published
- 2026-11-16
- DOI
- https://doi.org/10.5281/zenodo.24281828
- Primary Topic
- Parallel Computing and Optimization Techniques
- Type
- article
- Field-Weighted Citation Impact
- 0.00