0.5.1

OpenFold3 v0.5.1

Highlights

• Large memory reduction for inference and validation. MSA and template features are now kept as compact indices and coordinates, and are expanded inside the model only when needed. On a 2,022-token / 20,003-atom sample, persistent feature memory dropped from about 8.2 GB to about 430 MB (#462, @christinaflo).
• Deterministic inference. Feature creation (e.g. ligand conformers) is now seeded per datapoint. Atom→token aggregation at inference uses an order-stable segmented reduction instead of CUDA scatter_add_. Together these make repeated runs reproducible. Chunk-size tuning results are also cached across targets (#320, @heathcliff233).
• Numerics and kernel improvements from @GMNGeoffrey:
  • fp32 boolean mask sums (#282)
  • Triton tl.dot now respects torch fp32 matmul precision (#457)
  • TF32 is no longer enabled globally in tests (#470), and tests fail on leaked precision changes (#471)
  • attention kernel backward test fixes (#258)
  • GPU tests now run on ROCm (#433)
• Lower peak GPU memory. PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True is now the default, which cuts reserved memory by about 3 GB at token budget 384, with larger savings at higher budgets (#414, @chelsea0x3b). Metrics computation is also more memory-efficient (#393).
• Biotite 2 support. Biotite ≥2.0 now works (#464, #465). This release also drops the temporary biotite<2 pin (#460) and the rdkit<2026 pin (#485).

⚠️ Breaking and behavior changes

• Python ≥3.12 is now required (#480). The floor was raised from 3.10, following [SPEC 0](https://scientific-python.org/specs/spec-0000/).
• Conda has been removed from installation and CI paths (#401). Pixi is now the documented development setup (#443).
• Seeding behavior changed:
  • --num_model_seeds now starts from the configured seed instead of a fixed 42 (#436).
  • Feature RNG is scoped to each datapoint’s seed (#320).Outputs for the same inputs may differ from v0.5.0.
• MSA deletion_value fix (#347). The coefficient was 4× too large (8/π instead of 2/π). This changes model inputs for nearly every query that has an MSA.
• Config changes for training users (#462):
  • Template distogram bin settings moved to model.template.template_pair_embedder.distogram.
  • The feature dict now carries msa as uint8 indices and template_pseudo_beta_coords / template_frame_atom_coords instead of the distogram and unit-vector features.

What's Changed - full list

• Update readme with OpenBind-0 information by @etowahadams in #377
• Fix naming of openbind parameter download by @jnwei in #380
• chore(typing): annotate inference configurations by @ahmedtaha100 in #384
• Specify file destination path of AMD parameter by @jnwei in #387
• Allow custom residue names for SMILES ligands by @ahmedtaha100 in #386
• fixes #390: citation file updates by @jandom in #391
• chore(deps): bump prefix-dev/setup-pixi from 0.10.1 to 0.10.2 in the github-actions group by @dependabot[bot] in #392
• Update parameter documentation by @jnwei in #396
• Add CIF template-based generating example notebook by @sukritsingh in #255
• Add round robin retry of AWS AZs for GPU CI tests by @jnwei in #400
• Removes conda from installation and CI paths by @jnwei in #401
• In CI / CD tests: Fix OPENFOLD_CACHE env variable leakage between tests by @jnwei in #402
• Fix documentation dependency and pocket configuration example by @peter-s-obi in #398
• More memory efficient metrics by @chelsea0x3b in #393
• fix: keep inference intermediates under the output directory by @ahmedtaha100 in #363
• chore: update pixi to 0.8.0 by @jandom in #403
• [CI/CD] Stop the AMD and EC2 runners sharing a parameter cache by @jnwei in #404
• ci: fix broken integration tests by @jandom in #405
• Improve template fetching by ColabFoldQueryRunner from RSCB server by @jnwei in #417
• chore(deps): bump docker/build-push-action from 7.3.0 to 7.4.0 in the github-actions group by @dependabot[bot] in #424
• doc update for #172: updating the docs to correctly reflect that msa … by @Pana-TsK in #337
• Fix cached template reuse in inference query sets by @ahmedtaha100 in #425
• test: fix flaky BF16 Triton template stack by @ahmedtaha100 in #429
• Use upstream AZ fallback from start-aws-gha-runner v1.4.0 by @jnwei in #426
• Don't skip GPU tests on ROCm by @GMNGeoffrey in #433
• fix: make register_dataset_config return the decorated class by @alexbinnie-apheris in #434
• Do boolean mask sums directly in fp32 by @GMNGeoffrey in #282
• chore: update mypy-baseline by @jandom in #437
• Start --num_model_seeds from the configured seed instead of a fixed 42 by @OtoYuki in #436
• Inference determinism: atom aggregation, feature seeding, and chunk-size reuse by @heathcliff233 in #320
• Sets PYTORCH_CUDA_ALLOC_CONF=expandable_segments:True by default by @chelsea0x3b in #414
• ci: fix daily tests after expandable_segments by @jandom in #438
• [Feat] Template preprocessing for training dataset cache jsons by @jnwei in #439
• [Feature] Template preprocessing entry point by @jnwei in #440
• docs: document Pixi development setup for contributors by @ethanholz in #443
• Remove unnecessary script init files by @sdvillal in #446
• Workflow for nightly CI on Brev runners by @jnwei in #448
• Fix cyclic relpos offsets: shortest ring distance, handle tokenized residues, only cyclize residue offsets by @GMNGeoffrey in #445
• Replace assert_sole_holder with check for freed storage by @GMNGeoffrey in #451
• chore(deps): bump prefix-dev/setup-pixi from 0.10.2 to 0.11.0 in the github-actions group by @dependabot[bot] in #452
• refactor: remove monkeypatch from tests by @jandom in #449
• fix: correct deletion_value coefficient (was 4x too large) by @Pana-TsK in #347
• fix: use bin centers, not left edges, for pLDDT/PAE/PDE classificatio… by @Pana-TsK in #355
• Parse a query entry from an mmcif file by @jnwei in #381
• initial draft of SKILLS.md by @jnwei in #382
• chore: strict mypy on everything but the tests by @jandom in #454
• Use torch fp32 matmul precision setting for tl.dot by @GMNGeoffrey in #457
• chore: side-step issue with dot_precision.py typing by @jandom in #461
• fix: pin biotite version to below 2.0 by @rchauhan9 in #460
• fix: configurable cluster environment; stop SLURM autodetection hangs by @jandom in #459
• Keep instance defaults for partial dataset config updates by @GMNGeoffrey in #458
• Fix degeneracy in attention kernel backward tests by @GMNGeoffrey in #258
• chore(deps): bump prefix-dev/setup-pixi from 0.11.0 to 0.11.2 in the github-actions group by @dependabot[bot] in #467
• adds g6.4xlarge to allowed instance types for brev integration tests by @jnwei in #469
• Avoid enabling TF32 globally in tests by @GMNGeoffrey in #470
• Biotite v2 compatibility changes by @jnwei in #464
• Allow biotite 2 by @jnwei in #465
• Feature dict refactor to reduce memory for validation / inference by @christinaflo in #462
• Mark test_shape_large_bf16_train test as slow by @GMNGeoffrey in #478
• chore: bump to 3.12 by @jandom in #480
• Fail tests on leaked changes to fp32 matmul precision by @GMNGeoffrey in #471
• ci: python floor checks by @jandom in #481
• Fix/skip biotite ccd update by @jnwei in #474
• Ci/gpu test marker by @jnwei in #472
• fix: two bugs in extract_alignments_to_pair (crash + off-by-one) by @Pana-TsK in #350
• fix: conformer generating timeout escape by @jandom in #483
• test: add msa_mask/padding tests for create_msa_feature_precursor_of3 by @Pana-TsK in #346
• perf: restrict frame-atom distance matrix to start-atom rows by @heathcliff233 in #455
• [chore] update openfold3 release to v0.5.1 by @jnwei in #484
• Remove rdkit<2026 pin by @jnwei in #485
• Pin ruff version by @GMNGeoffrey in #489

New Contributors

• @sukritsingh made their first contribution in #255
• @chelsea0x3b made their first contribution in #393
• @alexbinnie-apheris made their first contribution in #434
• @OtoYuki made their first contribution in #436
• @heathcliff233 made their first contribution in #320
• @ethanholz made their first contribution in #443
• @rchauhan9 made their first contribution in #460

Full Changelog: 0.5.0...0.5.1