[DimaMolod]

Last of three behaviour-preserving PRs that split alphafold3_backend.py; stacked on the input-preparation PR (#659). The backend goes from 2,455 lines on main to 300, holding only AlphaFold3Backend (setup, prepare_input, predict, postprocess).

Layout after the three PRs

Module	Holds
alphafold3_backend.py	AlphaFold3Backend; re-exports process_fold_input, predict_structure, ResultsForSeed
alphafold3_inputs.py	AlphaPulldown objects and AF3 JSON → one fold input (FoldInputBuilder, PreparedFold)
alphafold3_runner.py	Model running, adapted from upstream run_alphafold.py
alphafold3_outputs.py	Output directory: per-seed and per-fold writers, ModelCIF metadata, resume
alphafold3_debug.py	Debug artifacts: translated MSAs and their summary, final complex MSA
utils/output_paths.py	String naming rules, shared with the AF2 path, no AF3 import

No extracted module imports the backend. alphafold3_outputs names ResultsForSeed only for type checking.

Commits

1. The move. ConfigurableModel, ModelRunner, ResultsForSeed, predict_structure, process_fold_input, _record_inference_kernels and _initialise_tokamax_trace_context go unchanged to alphafold3_runner.py, so the code adapted from upstream can be diffed against it on its own. The backend re-exports what callers import from it. Unit tests of runner code patch and call it in the runner.
2. setup tidied:
   • Its nested make_model_config becomes a runner function, without the leftover "new code approach" comments.
   • Its docstring says once why every field is set behind hasattr: AF3 releases differ in which fields their config has. return_distogram came to the AF3 history later than the others (2025-05).
   • setup looks up model.Model once per call and passes it to both the config and the runner, as the nested import did.
   • predict's buckets is typed as what it accepts: an int or a sequence.

Verification

• CPU equivalence harness (the 15 inputs of the previous PR), after each commit: byte-identical to main.
• Unit + integration, both release images: the same results as main plus the 6 builder tests of the previous PR. AF2 image: 937 passed, 26 skipped. AF3 image: 1008 passed, 15 skipped. The same pre-existing environmental failures.
• GPU, full AF3 suite with RUN_EXPENSIVE_GPU_TESTS=1: 39/39 passed, including TestAlphaFold3MetadataEndToEnd (real feature generation through inference and ModelCIF), resume, fast kernels and the MMseqs/regression checks.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

Codex Review Summary

This comment shows the latest Codex review activity on this pull request.

Review	Status	Commit	Review trigger
📝 Code Review	✅ Completed 2026-10-13T18:22:53.901459Z	11cde8f	PR opened

ℹ️ About Codex in GitHub

Your team has set up Codex to review pull requests in this repo. Reviews are triggered when you

• Open a pull request for review
• Mark a draft as ready
• Comment "@codex review" or "@codex security review".

Codex reacts with 👀 while any review is running, comments if it has suggestions, and reacts with 👍 once all reviews finish with no findings.