MHC-Fine

5-30 minutesPaperCodeNeed >5GB

Glukhov E et. al

Readiness

Readiness: Need >5GB
Smoke: failed
Artifacts: failed
Production: need_gt5gb_cuda_pytorch_mhc_fine_runtime_pending
Storage: 5-20GB
Strategy: large shared CUDA PyTorch SIF/overlay for official MHC-Fine path
Next: Build or reuse a CUDA PyTorch SIF/overlay that satisfies the hard-coded cuda path, rerun the official MHC-Fine model probe through Slurm, then add runner/API only after pass.

SOP agent staged official Bitbucket mhc-fine revision 192dae21b047a6f3d2e2ab7ecd7ecdceb0a17955 at /media/nik/seagate_nik/bio_server/tools/mhc-fine/official/source plus supplemental alphafold_finetune revision b85e577659d726154cac6b0c22908e6aa11cb4fd. Tool folder about 773M; official weight mhc_fine_weights.pt is 388,268,218 bytes. Reused palm_py312_site with tiny local absl-py/gdown deps. Slurm job 378 ran model probe for 00:11:44 and failed at AssertionError: Torch not compiled with CUDA enabled because official code hard-codes cuda. Evidence log: /media/nik/seagate_nik/bio_server/tools/mhc-fine/reports/evidence_20260701_worker_mhc_fine.md. 2026-07-07 16:05 +0800: SOP Runner Adapter Pending audit: demoted from Runner Adapter Pending because the official Slurm probe failed after 00:11:44 with Torch not compiled with CUDA enabled; source and weights are staged, but CUDA runtime validation is not complete. 2026-07-07 17:16 +0800: SOP Runner Adapter/Blocked pass: moved from Production Blocked to Need >5GB because the official Slurm probe failed on CPU-only torch while upstream hard-codes CUDA; shared CUDA PyTorch runtime work should be handled in the large-runtime phase.