add-model

Add a new model to AI4Science Studio

Add a new model to this repository following all repo conventions.

Steps

1. Identify domain — pick from earth_science/, material_science/, protein_folding/, healthcare/, or physics_simulation/. If the model clearly belongs to a new domain, create the domain folder with a README.md first.

2. Derive the slug:

   • HF id org/model → directory name org__model (replace / with __).
   • If a well-known public name is used instead (e.g. StormCast, NeuralGCM), use that name and document the canonical HF id inside the model's README.md.
   • If weights are not on Hugging Face (e.g. GCS bucket, GitHub release, Google Drive), set the HF id field to N/A and add an "Obtaining model weights" section with a fetch snippet.

3. Create the model folder by copying _template/ to <domain>/models/<slug>/.

4. Fill in README.md with:

   • Hugging Face model id (or N/A with alternate source documented)
   • Task description
   • License (SPDX id or direct link)
   • Upstream code repo and paper links
   • Acknowledgements and citation section (required — see template):
     • Upstream repo URL
     • Paper citation (author, title, venue, DOI or arXiv id)
     • BibTeX or DOI-based "cite as" block copied from the upstream repo or model card
     • ROCm blog URL and author names (AMD Silo AI) if a blog post exists
     • Any named collaboration (e.g. AstraZeneca × AMD, ORNL × AMD)

5. Add recipes under <slug>/recipes/. Use one subfolder per task (recipes/inference/, recipes/finetune/, etc.). Each recipe subfolder needs a README.md.

6. Add an examples/ directory with ready-to-run scripts:

   • docker_run.sh — launches the container; auto-detects AMD Container Toolkit vs device passthrough; checks for existing container before launching; auto-clones the upstream repo if absent.
   • run_inference.sh / run_inference.py — runs the model; all key params overridable via env vars with sensible defaults; prints a config summary before running.
   • preflight_<slug>.py — smoke-test that verifies GPU access and imports before a full run.
   • sbatch_inference_amd.sh — SLURM batch script (use --rocm for AMD/Apptainer GPU passthrough, not --nv). Named _amd.sh (not _mi300x.sh) — the same script covers MI250X, MI300X, and MI350X with a rocm7.2.x image.
   • build_overlay_amd.sh — (for HPC models with heavy pip deps, e.g. >1 GB of packages) builds a persistent Apptainer ext3 overlay pre-loaded with pip deps; mount :ro in inference jobs to skip the install phase.
   • Add run_ensemble.py, run_finetune.sh, etc. as applicable.
   • All scripts must be chmod +x.