APM

5-30 minutesPaperCodeNeed >5GB

All-Atom Protein Generative Model

Readiness

Readiness: Need >5GB
Smoke: not_tested
Artifacts: not_applicable
Production: need_gt5gb_cuda_faesm_model_runtime
Storage: <5
Strategy: shared PyTorch SIF + tool ext3 overlay; avoid per-tool torch duplication
Next: Process in large GPU runtime round: build/reuse CUDA APM SIF, stage eval-only weights/data, then run PepBench Slurm smoke before wiring.

APM source staged from bytedance/apm rev a98e59b17df29aec9424b598c5fac4bdc492b3e7; staged size ~151 MB. README lists data/weights including APM_full_data_90.tar.gz 9.5 GB and swissprot_data.tar.gz 11.2 GB, plus CUDA PyTorch/FAESM/FlashAttention/torch-scatter/OpenFold/ProteinMPNN stack. Slurm job 396 static probe passed; job 397 dependency probe failed on missing torch/hydra/omegaconf/lightning/deepspeed/faesm/torch_scatter/biotite/mdtraj. No inference smoke. 2026-07-04 backend update: shared SIF inventory only has bare python312/ubuntu2404 and ProteinMPNN-ddG CPU SIF; APM still needs CUDA PyTorch/FAESM/FlashAttention/torch-scatter plus eval weights/data, so it is no longer a lightweight Smoke Pending candidate.