[DimaMolod]

Plan item 2 of 3: share the prediction execution code between the two adapters.

Problem

PreparedPredictionAdapter (the single-fold command) and AlphaPulldownPredictionAdapter (the batch command) each had their own copy of:

• configuration validation;
• backend selection and setup;
• seed choice;
• the predict and post-process loop.

A fix had to reach both copies.

Change

Two private helpers in prediction_batch.py, used by both adapters:

• _setup_session validates the configuration, selects the backend, sets it up and returns the existing session object.
• _predict_and_postprocess chooses the seed with the existing policy, predicts, and post-processes each yielded fold.

Parsing, feature loading, output-directory resolution and fold composition stay in the adapters. The classes, method signatures, session object and backend result dictionaries are all unchanged.

Behaviour kept

• One setup per resident batch.
• The seed is chosen once per command or batch job, as before.
• A complex's AF3 inputs stay one prediction; independent manifest jobs stay separate.
• AF2 monomer and multimer grouping is unchanged.
• A resumed fold that yields nothing still counts as done.
• Setup errors propagate; a job's own failure leaves later jobs running.

Tests

• Parity: both adapters, given the same prepared AF3 complex, with the seed unset and set. Every backend call is compared: backend selection, setup, prediction and each post-processing call.
• Failures: a job failing in prediction, and one failing in post-processing. The next job still completes, and setup ran once.
• Existing tests: setup reuse, the AF3 seed policy for both adapters, resumed folds that yield nothing, CLI exit codes and naming. All pass unchanged.
• Unit + integration, both release images: identical to main, plus the new tests. AF2 image: 978 passed. AF3 image: 1049 passed.
• GPU: the full AF2 and AF3 prediction and resume checks run on main after this round is merged.

🤖 Generated with Claude Code

[chatgpt-codex-connector]

You have reached your Codex usage limits for code reviews. You can see your limits in the Codex usage dashboard.