[DimaMolod]

Stacked on #660. Fixes mistyped backend options being dropped silently.

The problem

The adapters call every backend's setup and predict with the whole configuration, so each one took the rest as **kwargs and read its own options with kwargs.get(name, default). A name mistyped on either side, in the configuration table (inference_flags._MODEL_FLAG_SOURCES) or in a backend, fell back to the default without a word. validate_model_configuration could not catch it: it checks the configuration against a key set built from that same table.

The fix

• Named options. setup and predict of AlphaFold 2, AlphaFold 3 and AlphaLink now name every option they use, as does AF2's predict_individual_job. Options meant for other backends arrive as **other_backends_options and go only to validate_model_configuration, which rejects any that no backend takes.
• AF3 resume settings. AF3 hands its resume code an explicit dict of the settings it records (num_seeds, num_diffusion_samples, num_recycles, the embedding and distogram switches) instead of its kwargs.
• Dead option removed. AF3 no longer forwards af3_input_json: nothing supplied it, and prepare_input ignores it.

Tests

• test_inference_flags reads the three backends' signatures from source, since AF2 and AF3 cannot be imported together. It checks three things:

  • every configuration key is named by some backend;
  • every option a backend names is a configuration key;
  • the unnamed options are only validated, never read.

  Mutation-checked: a mistyped key in the table, a mistyped AF3 predict parameter, and an option read back with .get each fail it.

• predict of both AlphaFold backends accepts other backends' options and reports a mistyped one.

• Unit + integration, both release images: the same results as main plus the new tests. AF2 image: 946 passed, 26 skipped. AF3 image: 1017 passed, 15 skipped. The same pre-existing environmental failures.

• GPU: full AF2 and AF3 suites, which go through these signatures with the real configuration. AF2: 18 passed, 3 skipped. AF3: 38 passed, 1 skipped. None failed. The skips are the expensive tests that the gate in test/cluster/code_changes.py left out because none of their covered paths changed: AF2's two fast-kernel tests and its LocalMmseqs comparison, and AF3's metadata end-to-end test, which passed on refactor: move AF3 model running out of the backend #660.

🤖 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.