MDGen

1-5 minutesPaperCodeNeed >5GB

Generative modeling of molecular dynamics trajectories

Readiness

Readiness: Need >5GB
Smoke: not_tested
Artifacts: not_applicable
Production: need_gt5gb_pytorch_openmm_legacy_runtime
Storage: <5
Strategy: shared python312_slim.sif + tool ext3 overlay
Next: Build/reuse legacy GPU PyTorch/OpenMM SIF and fetch one required checkpoint, then run minimal MDGen inference smoke before evaluating web use.

MDGen source resolved to bjing2016/mdgen rev 81482a403b91c5a8437046da1d8d321ba97089cc; catalog URL was malformed. MIT license, but README notes research reproducibility/not intended for application workflows. Public HF model repo has five checkpoints totaling ~686 MB; large tetrapeptide dataset not downloaded. Source ~27 MB. Slurm job 399 did syntax/top-level probe only; no model inference. Requires old torch==1.12.1+cu113, mdtraj, openmm, pyEMMA, fair-esm, pytorch_lightning/OpenFold-style stack. 2026-07-04 backend update: available shared SIFs do not include torch/openmm/mdtraj/pyEMMA/fair-esm; MDGen requires old CUDA PyTorch/OpenMM stack plus checkpoints, so it should move to >5GB runtime phase.