RFdiffusion

5-30 minutesPaperCodeProduction Ready

Design or diversify proteins

Job Name

EXAMPLES

Task *PDB FileEnter PDB CodeContigsTarget ChainsBinder LengthBinder Hotspots

Hotspot ChainHotspotsBinder ChainDesigned ChainInterface ChainDesigned ResiduesDiffused ResiduesInterface ResiduesDesigned Lengths

Number of DesignsPartial TProvide SequenceFold Conditioning Target FileFold Conditioning Binder FileTarget SS FileTarget Adj File Verify with ProteinMPNN + AlphaFoldVerify SequencesNoise Scale CANoise Scale FramePotentialsSymmetry TypeOligomer LengthMotif ResiduesScaffold Length BeforeScaffold Length After Use Oligomer ContactsGuide ScaleGuide DecayModel

Readiness

Readiness: Production Ready
Smoke: passed
Artifacts: passed
Production: production_ready_shared_rfdiffusion_sif_cpu_slurm_verified
Storage: 5-20GB installed
Strategy: Production adapter uses shared RFdiffusion SIF and public checkpoint cache on /media/nik/seagate_nik/bio_server; CPU fallback is required on this RTX 5050 host because the current PyTorch CUDA stack lacks sm_120 kernels.
Next: Optional optimization: rebuild RFdiffusion SIF with PyTorch/CUDA supporting RTX 5050 sm_120 for GPU acceleration. Functional CPU Slurm/API path is production-ready for small designs.

generative structure/design model weights likely >5GB 2026-07-17T01:36:00+08:00 SOP Need >5GB next-10 audit: remains Need >5GB. Runner audit is not_configured and object-store evidence is generic-smoke only. Backend inventory found no real RFdiffusion source/checkpoints/SIF or overlay; adjacent RFdiffusion2/3 are also smoke-only. 2026-07-20T22:45:00+08:00 SOP Need >5GB continuation: classic RFdiffusion checkpoint cache is populated under /media/nik/seagate_nik/bio_server/cache/rfdiffusion/models (seven public checkpoints, 3.2GB). Started shared SIF pull from docker://rosettacommons/rfdiffusion:latest into /media/nik/seagate_nik/bio_server/containers/rfdiffusion/rfdiffusion.sif; process still running and no direct/Slurm benchmark yet. This same SIF/cache should serve rfdiffusion and rfpeptides. 2026-07-25T19:08:30+08:00 SOP Need >5GB batch: promoted to Production Ready. Shared SIF /media/nik/seagate_nik/bio_server/containers/rfdiffusion/rfdiffusion.sif (7.3GB) and checkpoint cache /media/nik/seagate_nik/bio_server/cache/rfdiffusion/models (seven public checkpoints, 3.2GB) are staged. Direct CPU smoke passed for contigmap.contigs=[50-50] with diffuser.T=15/final_step=14. Deployed real rfdiffusion adapter to /home/nik/bio_mega_runner and verified API/Slurm job e0100e0d2193 / Slurm 829 COMPLETED. Artifacts verified: rfdiffusion_smoke_0.pdb 13,400 bytes, rfdiffusion_smoke_0.trb 3,433 bytes, residue_count=50, summary.json, results.csv, empty slurm.err. GPU attempt failed on this RTX 5050 because the container PyTorch CUDA supports sm_37-sm_86 but the host is sm_120; production path uses CPU fallback until a newer CUDA image is built.