AbMAP

secondsPaperCodeProduction Ready

Antibody language model

Readiness

Readiness: Production Ready
Smoke: passed
Artifacts: passed
Production: ready
Storage: <5
Strategy: shared python312_slim.sif + AbMAP staged source/env/checkpoint + shared CPU Torch layer
Next: Production Ready for scoped AbMAP checkpoint model-forward fixed/variable embedding generation. Optional hardening: add full sequence preprocessing via ProteinEmbedding/create_cdr_specific_embedding for user antibody sequences.

AbMAP source cloned from rs239/abmap rev 3d8ef30180d0bc2e2657ae38ae37c77228a0703b; footprint ~710 MB with pretrained models. Slurm jobs 392 and 395 attempted import/model smoke using shared site packages and failed on missing biotite/loguru; small dependency install completed afterward but no successful rerun yet. 2026-07-05 production-path evidence: direct backend Slurm job 419 used python312_slim.sif plus staged AbMAP env/checkpoints and completed an AbMAP model-load/embed smoke. Artifacts: embedding_fixed.npy, embedding_variable.npy, summary.json, results.csv, stdout/stderr. Summary reported abmap_version 0.1.6, torch 2.4.1+cpu, fixed_shape [1,512], variable_shape [1,32,256], runtime_seconds 17.882. 2026-07-07T21:18:05+08:00 SOP production validation: added real runner adapter for abmap scoped checkpoint model-forward embedding mode. Initial runner API job 07372ff5d7b9 / Slurm 515 produced finite embeddings but stderr had optional dependency import noise from package __init__; adapter was hardened to load abmap/model.py directly. Clean runner API job 914a2723830f / Slurm job 516 completed in 1.463 s with empty slurm.err. Output shapes: fixed embedding 1x512 and variable embedding 1x32x256, all finite, using AbMAP_beplerberger_H.pt checkpoint. Artifacts verified: summary.json, results.csv, abmap_runtime.json, embedding_fixed.npy, embedding_variable.npy, job.log, slurm.out, slurm.err. Production scope is checkpoint model-forward embedding; sequence-to-embedding preprocessing remains optional future expansion.