Chai-1

5-30 minutesPaperCodeProduction Blocked

AlphaFold3 reproduction, predict protein/nucleic acid/small molecule complexes

Job Name

EXAMPLES

Use Chai serverNumber of Samples *

Use MSA serverInput Format Use MSAMSA Database Use PDB100 templatesTemplate CIF FilesNumber of Trunk SamplesNumber of SeedsNumber of RecyclesSeedPocket RestraintsContact RestraintsCovalent RestraintsRestraints Min DistanceRestraints Max DistanceModifications Use BFVD Logan MSA

All sequences must be non empty.

Readiness

Readiness: Production Blocked
Smoke: passed
Artifacts: passed
Production: production_blocked_chai_current_backend_gpu_oom_needs_larger_vram
Storage: runtime/cache staged; current 8GB GPU is insufficient
Strategy: Shared Chai layer over python312_slim.sif and boltz_py312_site; may require separate newer PyTorch SIF or larger-GPU node for validation.
Next: Provision or route Chai-1 to a larger-VRAM GPU node, or implement a proven Chai-specific lower-memory recipe; rerun the tiny benchmark until CIF artifacts are produced, then wire and validate the real Slurm/API adapter.

2026-07-18T17:33:47+08:00 SOP large-tool batch: remains Need >5GB. Official chai-lab docs say weights download automatically and can be redirected with CHAI_DOWNLOADS_DIR. Backend has no chai-lab source/runtime/weights under the shared root; current runner is chai-1-smoke placeholder. 2026-07-18T18:39:16+08:00 SOP Need >5GB active batch: Read-only audit confirmed chai-lab code/weights are public Apache-2.0; model files are about 1.1GiB but GPU runtime/cache likely exceeds 5GB. Backend has no chai runtime/cache yet. 2026-07-18T19:02:14+08:00 SOP Need >5GB active install batch: Staged Chai runtime layer at /media/nik/seagate_nik/bio_server/envs/chai_py312_site (about 504MB) with CHAI_DOWNLOADS_DIR=/media/nik/seagate_nik/bio_server/cache/chai. `from chai_lab.chai1 import run_inference` imports successfully using shared Torch 2.13.0+cu130 and NumPy 1.26.4 after avoiding duplicate NumPy. Direct tiny no-MSA/no-ESM GPU run on 8GB RTX 5050 reached model execution but failed CUDA OOM while allocating 1GB; no production artifacts generated. 2026-07-18T20:49:48+08:00 SOP targeted make-ready/reclassify batch: Reclassified from Need >5GB to Production Blocked. Chai runtime layer and imports are staged, but the bounded no-MSA/no-ESM tiny GPU benchmark on the RTX 5050 8GB reached model execution and failed CUDA OOM while allocating about 1GB. This is no longer a storage-install blocker; it needs a larger-VRAM backend GPU or a proven lower-memory Chai execution recipe before production artifacts and adapter validation are possible.