Dayhoff

1-5 minutesPaperCodeNeed >5GB

Generate sequences conditioned on existing sequence

Readiness

Readiness: Need >5GB
Smoke: passed
Artifacts: passed
Production: deferred_dayhoff_cuda_mamba_flashattention_models
Storage: 5-20GB
Strategy: shared PyTorch SIF + tool ext3 overlay; avoid per-tool torch duplication
Next: Build shared CUDA PyTorch/Mamba/FlashAttention container, mirror at least Dayhoff-170m-UR50, run minimal GPU generation to non-empty canonical FASTA, then add runner adapter/artifacts.

SOP agent checked upstream microsoft/dayhoff main revision df35b0131a3a2fb1757c6e54df8d3ad684b46b34. Public/ungated HF model refs include microsoft/Dayhoff-170m-UR50 and microsoft/Dayhoff-3b-GR-HM-c. Backend has no Dayhoff install/model cache/CUDA PyTorch-Mamba-FlashAttention shared container; no Slurm benchmark artifacts or runner adapter. Evidence log: /media/nik/seagate_nik/bio_server/reports/dayhoff_sop_evidence_20260701.md.