version: "1.0.0" name: rfd4-proteina-setup description: > Prepare a machine to run RFD4-Proteina (RFD4-Proteina) — the Baker Lab x NVIDIA flow-matching protein design model. Reach for this skill for "set up rfproteina", "install rfproteina", "environment for rfproteina", "download the rfproteina checkpoints", "DATA_PATH not found", "atomworks not found", "which checkpoint do I use", "run rfproteina on a cluster", or before any generate/evaluate/search run. It installs the native Python environment and AtomWorks, wires DATA_PATH and runtime settings, downloads the model weights, and preflights GPU + checkpoints so downstream runs do not fail mid-job. Hand off to rfd4-proteina-generate once preflight is green. license: See the RFD4-Proteina repository (Baker Lab x NVIDIA collaboration)

RFD4-Proteina Setup

Get a host ready to run RFD4-Proteina inference: environment, model weights, and a green preflight. The two failure modes this skill exists to prevent are (1) importing the model before its environment and AtomWorks are installed, and (2) launching an hours-long run only to have it die because DATA_PATH or a checkpoint is missing. Preflight both up front.

Run everything from the repository root (the directory containing .project-root).

Step 1 — Install the runtime

From the repository root:

make setup
source .venv/bin/activate

Native setup creates .venv with the CUDA 13 model dependencies and installs the public AtomWorks version pinned in pyproject.toml; private setup retains its AtomWorks development submodule. It creates .env.local only if neither .env.local nor .env exists, preserves existing files, and does not edit shell startup files.

Step 2 — Check data paths and optional refolding

Setup downloads the configured native checkpoint pair to DATA_PATH, or caches an enabled Hugging Face bundle under HF_HOME. DATA_PATH defaults to ./data_path. Exported shell settings take precedence over the environment file. Use bash setup.sh --skip-weights for an environment-only installation.

configs/inference.yaml selects native checkpoints or a pinned model repository. Use HF_TOKEN or hf auth login while private; public downloads use the same code. To add OpenFold3 on a host with Apptainer:

./scripts/download_assets.sh ./data_path

Pass your configured DATA_PATH instead if it differs from ./data_path. Set the printed oracle paths in .env.local and use metrics/refolding_oracles=of3. See docs/installation.rst for the separate host oracle.

Step 3 — Preflight

bash skills/rfd4-proteina-setup/scripts/preflight.sh

This uses the same runtime dispatcher as model entry points and checks imports, CUDA visibility, and cached model files. Resolve failures before sampling.

Step 4 — Smoke test

rfd4-generate inputs=data/benchmarks/monomer \
    subset_to_examples=monomer-short num_replicates=1 out_dir=./runs/smoke ~metrics compile_lengths=null

A generated CIF under runs/smoke/inference_samples/ confirms the inference path. Hand off to rfd4-proteina-generate for real designs.

Optional host tools

Setup always creates .venv for SLURM configuration resolution. bash setup.sh --venv adds notebooks and development dependencies. Use the installed rfd4-* commands from the activated environment.

Troubleshooting

• Import failures: rerun setup or the pip installation in your selected environment.
• Missing model files: check HF_HOME and rerun setup. Optional data uses DATA_PATH.
• Missing GPU: verify the NVIDIA driver and the selected runtime's GPU support.
• Refolding failures: check the separate oracle image, weights, cache, and host Apptainer.
• SLURM host tooling missing: rerun setup, then configure scripts/.user_info_local.

For checkpoint details, see checkpoints.