--- license: apache-2.0 tags: - partidesign - partiscore - molecular-design - pytorch - checkpoints --- # PartiDesign and PartiScore checkpoints This model repository distributes the original PartiDesign and PartiScore checkpoint files under Apache-2.0. The matching deployment source remains uncommitted for the maintainer's review; weight publication does not imply that the pending source changes are already on the official main branch. This weight-only package contains the original PartiDesign generation and PartiScore interface-scoring checkpoint binaries. No third-party weights, inference service, model implementation, environment or dataset is included. Use the [PartiDesign project](https://github.com/MegaGatlingPea/PartiDesign-official) to run the checkpoints. ## Files and integrity | Path | Use | Bytes | SHA256 | | --- | --- | ---: | --- | | `ckpts/partidesign.ckpt` | Molecular generation | 4,909,825,762 | `ade1da422f89c3817d616195380c68a77ba31a2f71f877803680c6ee09b54944` | | `ckpts/partiscore.ckpt` | Molecular-interface scoring | 25,350,931 | `8961ba6612883d0e8a30265cc643ab48ccc97abb24fa25684736327438587506` | Both files retain their existing binary format and filename. They were copied and checked byte-for-byte through matching source/copy SHA256 values; no checkpoint conversion, extraction or repackaging was performed. ## Compatible source and environments The tested source package is **PartiDesign 0.1.0**. Its current deployment candidate is based on source commit `d062407fddba918bce4de9c3d7feadd0bb28702d` with uncommitted patch SHA256 `c77d5655db1e460722c8b6c7ed3d9be8465e64073c4202df099f1581c3bdfe16`. That patch fingerprint identifies the reviewed working tree; it is not a new Git commit or a Hugging Face revision. The final source release reference will be recorded when the maintainer accepts and publishes it. The project links in this card are navigation references: they do not mean that the uncommitted installer changes are already on the official repository's main branch. Use the matching reviewed source and installation guide. | Profile | Python | PyTorch | JAX / GPU plugins | cuEquivariance | | --- | --- | --- | --- | --- | | CUDA 12 | 3.12.6 | 2.9.1+cu128 | 0.10.2 / CUDA 12 | 0.10.0 | | CUDA 13 | 3.12.6 | 2.11.0+cu130 | 0.10.2 / CUDA 13 | 0.13.0 | These are full Linux/NVIDIA environments. The native layer retains OpenMM 8.4.0 and Conda openmmforcefields `0.15.1=pyhcf101f3_1`. BioIR is included only with CUDA 13. The private project model implementations remain part of the source/environment; similarly named PyPI projects are not substitutes. The default installer uses uv/CUDA 12. **uv and Conda or Mamba are both prerequisites**: uv manages the locked Python/model stack and a paired native Conda prefix provides Python, chemistry libraries and build/search tools. ```bash bash install.sh source .venv/bin/activate ``` For CUDA 13, keep a separate virtual environment and native prefix: ```bash bash install.sh --cuda 13 \ --venv .venv-cu13 --native-prefix .cache/environments/native-cu13 source .venv-cu13/bin/activate ``` The named Conda route remains available explicitly: ```bash bash install.sh --installer conda --cuda 12 conda activate partidesign ``` See the matching [environment guide](https://github.com/MegaGatlingPea/PartiDesign-official/blob/main/docs/environments.md) for driver requirements, custom paths, updates and recovery. Keep each uv/native pair intact; never combine the two CUDA profiles in one environment. ## Downloading and running Use a full immutable Hugging Face commit hash from this repository's Files and versions history. The matching project download catalog pins its selected release revision. A mutable branch name is not an immutable release identifier; check the matching source documentation for the rollout's current catalog. Keep the two files under a model root: ```text /path/to/partidesign-models/ └── ckpts/ ├── partidesign.ckpt └── partiscore.ckpt ``` Use `--model-dir` or `PARTIDESIGN_MODEL_DIR` to select that root. From the matching source checkout, the public command entry points are: ```bash MODEL_DIR=/path/to/partidesign-models partidesign design configs/examples/protein.yaml \ --task protein_fixture --batch-size 1 \ --model-dir "$MODEL_DIR" --ckpt "$MODEL_DIR/ckpts/partidesign.ckpt" \ --out inference/hf_example # Score an existing directory of molecular complexes: partidesign score /path/to/complexes \ --model-dir "$MODEL_DIR" --ckpt "$MODEL_DIR/ckpts/partiscore.ckpt" \ --device cpu ``` The YAML task controls generation inputs, constraints and sampling. PartiScore accepts direct-child PDB/CIF files. It needs eligible contacting partners for an interface score; a monomer can report no eligible interface. See the matching [design](https://github.com/MegaGatlingPea/PartiDesign-official/blob/main/docs/design.md) and [scoring](https://github.com/MegaGatlingPea/PartiDesign-official/blob/main/docs/score.md) guides. The ordinary shipped pipeline uses ESMFold2-fast for iteration and ESMFold2-standard, with five samples, for final validation. Standard uses MSA for fixed protein context without templates; designed chains and tasks without eligible protein context remain query-only. The explicit PPI preset instead uses AF2-IG models [1,2], one complex/monomer pair each. AF3 and Mol-1 remain optional. Those upstream model weights are excluded from this package and must be prepared separately for the chosen route. ## Deployment verification and limits The deployment candidate was checked on fresh full Conda CUDA12/13 and uv CUDA12/13 environments on an RTX 4080 SUPER host. Installation/native GPU and backend checks completed; 48 real pipeline commands exited successfully. Four ordinary tasks under the original thresholds eliminated normally at round 1; their unexecuted later stages remain pending in those runs. Separate explicitly labeled deployment execution smokes covered the complete three-round and final stages. Fixed-context native MSA use, query-only cases, cache reuse/cache-only, completed resume and explicit search failures were checked independently. Eight explicit AF3/Mol-1 public validation calls also completed across the four environments, with finite native/public structures and complete input mapping. Those calls used the recorded standalone one-sequence/one-complex budgets and MSA off; they do not establish optional pipeline-final-five or experimental quality. Deployment checks establish software execution, not binding or activity. No additional benchmark accuracy or training-data claims are made by this card. Transport and source-catalog integration are checked separately from the software deployment results above. Those pre-publication deployment checks do not by themselves establish an anonymous cold download. Use the matching source release's download instructions and acceptance record for that separate gate. ## License The two project-owned checkpoints are prepared for release under **Apache-2.0**, as specified by the maintainer. The accompanying `LICENSE` is copied unchanged from the PartiDesign source repository. Third-party model parameters retain their own upstream terms and are not included here.