Readiness
- Readiness
- Need >5GB
- Smoke
- not_tested
- Artifacts
- not_applicable
- Production
- need_gt5gb_orb_cuda_chemistry_stack_pending
- Storage
- 5-20GB
- Strategy
- deliberate shared Orb CUDA/Chemistry SIF; avoid automatic pip dependency resolution
- Next
- Build a deliberate shared Orb CUDA/Chemistry SIF with Python >=3.12, torch >=2.8, orb-models, ase, and nvalchemi-toolkit-ops, then run ASE energy/forces smoke with a real Orb-v3 checkpoint.
Orb Models 3 source verified as orbital-materials/orb-models rev 81b26ac5cbfbb33d4d0a61d680a932e59f8894ac; PyPI orb-models 0.7.0 requires Python >=3.12; Apache-2.0. Public S3 Orb-v3/OrbMol checkpoints returned 200 OK and are ~97-99 MB each, no token required. Shared Python 3.12/Singularity and torch sites exist, but orb_models/ase/nvalchemiops are missing. No smoke run to avoid duplicate heavy deps. 2026-07-05 SOP recheck: public checkpoints are small and token-free, but shared Python/Torch path lacks orb_models, ase, and nvalchemiops; no direct smoke has passed. Kept Smoke Pending. 2026-07-05 SOP batch recheck: public checkpoints are small/token-free, but current shared Python/Torch path lacks orb_models, ase, and nvalchemiops; no direct smoke has passed. Kept Smoke Pending. 2026-07-05 SOP recheck: public checkpoints are small/token-free, but current shared Python/Torch path lacks orb_models, ase, and nvalchemiops; no direct smoke has passed. Kept Smoke Pending. 2026-07-05 SOP batch recheck: public Orb v3 checkpoints are small/token-free, but current shared Python/Torch path lacks orb_models, ase, and nvalchemiops; no direct smoke has passed. Kept Smoke Pending. 2026-07-05 18:19 +08:00 backend re-audit: Orb source remains staged (~55 MB) and public checkpoints are small/token-free, but shared Python/Torch lacks orb_models, ase, nvalchemiops and no direct smoke has passed. Agents agreed Smoke Pending remains correct. 2026-07-06 20:00 +08:00 attempted small shared python312_slim overlay for Orb deps. Source and checkpoint URLs are staged, but current upstream dependency resolution for nvalchemi-toolkit-ops[torch]>=0.3.1 pulls Torch 2.12.1 plus CUDA 13 wheels (large CUDA toolkit/cudnn/cublas/cusolver/cusparse/nccl/triton stack) rather than a small CPU overlay; install was stopped before packages were written to /media/nik/seagate_nik/bio_server/envs/orb_models_py312_site. No Slurm inference smoke submitted. Keep Smoke Pending; next attempt should use a deliberate shared CUDA chemistry SIF or locate a CPU-only nvalchemi/Orb path, not automatic pip dependency resolution. 2026-07-06 21:34 +08:00 recheck: sidecar review confirmed current automatic pip path is not suitable because upstream Orb stack requires Python >=3.12, torch >=2.8/nvalchemi-toolkit-ops and prior dependency resolution pulled a large CUDA stack. No CPU-only Orb path proven in this pass. 2026-07-06 22:18 +08:00 reclassified from Smoke Pending to Need >5GB after two-pass review: upstream Orb requires Python >=3.12, torch >=2.8 and nvalchemi-toolkit-ops; the prior automatic install attempt pulled a large Torch/CUDA stack, and read-only sidecar agreed the fastest valid path is a deliberate shared CUDA/Chemistry SIF unless a CPU-only nvalchemi path is proven. No direct smoke submitted.