Profluent E1

1-5 minutesPaperCodeProduction Ready

Protein language model

Readiness

Readiness: Production Ready
Smoke: passed
Artifacts: passed
Production: production_ready_api_slurm_profluent_e1_150m_masked_lm
Storage: Production runtime staged: 607MB E1 dependency layer plus 295MB E1-150m HF cache, reusing shared Torch layer
Strategy: Production Ready using shared python312_slim.sif, shared boltz_py312_site Torch layer, Profluent E1 source at /media/nik/seagate_nik/bio_server/repos/profluent_e1, dependency layer /media/nik/seagate_nik/bio_server/envs/profluent_e1_py312_site, and HF cache /media/nik/seagate_nik/bio_server/cache/profluent_e1.
Next: Production Ready for E1-150m masked protein-token prediction and embeddings. Optional hardening: expose model size selector, retrieval-augmented multi-sequence inputs, GPU/BF16 path on larger GPUs, and batch scoring.

large protein language model weights likely >5GB 2026-07-17T01:18:00+08:00 SOP Need >5GB next-10 audit: remains Need >5GB. Code link exists, but runner audit is not_configured and backend inventory found no Profluent E1 source/model/runtime assets. Needs large protein language model runtime/model bundle plus real Slurm benchmark. 2026-07-18T21:37:01+08:00 SOP Need >5GB batch source-staging: Official repo cloned to backend /media/nik/seagate_nik/bio_server/repos/profluent_e1 at bfd2620 (776KB). Import from source works with shared Torch layer, but a Profluent-specific dependency install for Transformers/HF utilities stalled before writing files; no model checkpoint or Slurm benchmark yet. Public HF weights are E1-150m/E1-300m/E1-600m; start with E1-150m. 2026-07-18T22:34:12+08:00 SOP Need >5GB continuation batch: No production validation completed in this continuation. Prior Profluent E1 dependency install attempts on a Python 3.12 overlay stalled before writing the needed Transformers/HF layer; next attempt should use a prebuilt wheelhouse/conda layer or SIF while reusing shared Torch. 2026-07-19T02:36:30+08:00 SOP production validation: promoted from Need >5GB to Production Ready. Installed non-Torch E1 dependency layer at /media/nik/seagate_nik/bio_server/envs/profluent_e1_py312_site (607MB), reused shared Torch 2.13.0+cu130 from boltz_py312_site, staged official source /media/nik/seagate_nik/bio_server/repos/profluent_e1 rev bfd2620, and cached public Hugging Face model Profluent-Bio/E1-150m under /media/nik/seagate_nik/bio_server/cache/profluent_e1 (295MB). Direct benchmark using official E1BatchPreparer passed on masked sequence MKT?AYIAKQRQISFVKSHFSRQLEERLGLIEVQANL with finite logits, embedding shape [1,41,768], top token L probability 0.1129051372. Added real cluster runner adapter and registry entry. Runner API job a42c3e7e756c / Slurm 827 completed with summary.json, results.csv 315 bytes, stdout/stderr, and empty slurm.err.