[DimaMolod]

Draft until an AlphaPulldown release contains KosinskiLab/AlphaPulldown#655, #656 and #657. Merge it together with the bump of the default images in config/config.yaml to that release.

Why

AlphaPulldown now resumes AlphaFold 3:

• a rerun skips folds that are already finished;
• with --num_seeds, it keeps the seeds an interrupted fold had already written out.

--allow_resume defaults to True there, so with the new images a crashed or timed-out batch, or an attempt 2 retry, continues where it stopped instead of predicting everything again. The workflow does not need to pass anything for that.

Change

• _AF3_INFERENCE_FLAGS (and the pinned copy in test_inference_flag_tables.py) accept allow_resume, so --allow_resume: false passes the parse-time check. config/config.yaml shows that opt-out in a comment.
• batch_inference_args still adds --allow_resume for AlphaFold 2 batches only. AlphaFold 3 resumes by default, and its images up to 2.10.0 reject the flag.
• Comments, docstrings and test examples that called --allow_resume AlphaFold2-only now use other AF2-only flags (--models_to_relax, --num_cycle).
• README: two in-place edits and no added lines.
  • The batching section said AlphaFold 3 batches recompute unfinished folds.
  • The flag warning used --allow_resume as its AF2-only example.

Tests

• pytest test: 167 passed, 48 skipped; the one new test checks that allow_resume is valid on both backends.
• test_tables_match_alphapulldown_when_importable (skipped in CI, which doesn't install AlphaPulldown):
  • passes against AlphaPulldown with #655–#657;
  • fails against 2.10.0 because of the new AF3 allow_resume, as expected. That is why this waits for the release.

🤖 Generated with Claude Code