raftbioworks/boltz2_quotient_validation_20260918_060803

Boltz 2 quotient-validation study

Inference-only experiments with group-indexed pair features and tiled attention, compared with the supplied Anthropic optimization kit's actual big mode on H100 GPUs.

Memory savings were measured; improved or preserved folding accuracy was not established. This is not a newly trained model. The checkpoint is the unchanged upstream Boltz 2 structure checkpoint, SHA-256 090e82ac8c92f5e943fa1b39e7410a44027bea7243c0bbb3caa67a77fc1428e1 (2,286,561,469 bytes). No training or fine-tuning occurred.

• —Final report and limitations
• —Length / memory / runtime / OOM figure
• —Length / designability figure
• —Validation workflow and API usage
• —Unchanged checkpoint
• —Reproduction archive: source snapshots, dependency adapter, inputs, references, raw predictions/logs, tests and plotting scripts; read REPRODUCE.md after extraction.
• —W&B evaluation logs

At 3,200 synthetic C8 residues, peak allocation is 16.93 GiB versus 51.54 GiB for one-GPU big, with slower forward execution. One H100 completes 7,336-residue protein-only GroEL; 9,552-residue phosphoketolase still OOMs, while four-GPU big completes both. Capacity tests are not evidence of correct folding. Both methods have 0/6 strict assembly-designability successes in the tiny saved FrameDiff cohort (two designs per backbone).

Upstream weights/code: Boltz 2.2.1. Baseline: optimization kit commit f4f62fa. The supplied ZeUS manuscript is not redistributed. ESMFold2/OpenFold3 adapters were schema-tested only; only Boltz2 was GPU-benchmarked.