Skip to content

Add type hints + unit tests across foundry.utils and rf3#321

Merged
woodsh17 merged 12 commits into
RosettaCommons:productionfrom
lyskov-ai:0038-foundry-rotation-augmentation-mypy-strict
Jun 10, 2026
Merged

Add type hints + unit tests across foundry.utils and rf3#321
woodsh17 merged 12 commits into
RosettaCommons:productionfrom
lyskov-ai:0038-foundry-rotation-augmentation-mypy-strict

Conversation

@lyskov-ai

Copy link
Copy Markdown
Contributor

Brings the shared foundry.utils layer and several rf3 modules under stricter type checking and fills unit-test gaps. No behaviour changes — the source edits are type annotations and test additions only, plus two small documented exceptions noted below.

Type hints (shared layer)

Annotated and opted into strict mypy checking (untyped defs disallowed, function bodies checked) for six foundry.utils modules: rigid, components, torch, alignment, weights, and rotation_augmentation. Any new untyped function in these modules now fails the type-check gate.

Unit tests

  • Shared layer: new tests/test_rigid.py and tests/test_components.py, plus extensions to tests/test_torch_utils.py and tests/test_weight_loading.py, pinning the geometry, contig-parsing, tensor, and weight-loading helpers.
  • rf3: CPU-only, fixture-backed tests for pure logic — loss gradients, distance encoding / init helpers, layer shape helpers, the diffusion sampler's noise schedule, triangle attention/multiplication, the conformer-averaging layer, the outer-product layers, the predicted-error / chiral metrics, and the symmetry-resolution geometry helpers. All run without GPU, checkpoints, or cluster data.

Other

  • rf3.data.paired_msa: moved its type-check suppression from the central pyproject.toml reject-list into a file-level directive in the module itself. The module is currently broken against the installed atomworks (it subclasses a symbol that became a factory function) and needs a separate refactor; the in-file directive keeps the suppression visible next to the broken code.
  • rf3/utils/io.py: one documented cast where annotating weighted_rigid_align surfaced a loosely-typed caller (dump_trajectories, whose parameter is wider than its tensor-only body). Behaviour is preserved; narrowing the caller is filed as follow-up.

lyskov and others added 12 commits June 9, 2026 17:49
…_module

Cover the pure numeric helpers in rf3.loss.loss and rf3.util_module:

- calc_ddihedralmse_dxyz: the hand-derived closed-form dihedral-loss
  gradient, pinned against torch.autograd of a mirrored forward, plus
  zero-at-truth and leading-dim shape preservation.
- calc_chiral_grads_flat_impl: empty-centres, scatter routing onto only
  the centre atoms, index_add_ accumulation for shared atoms, and the
  no_grad_on_chiral_center flag.
- rbf: Gaussian distance-encoding (feature dim, exact values, [0,1]).
- init_lecun_normal: module returned, Parameter weight, +/-2*stddev
  bound, and Lecun std sqrt(scale/fan_in).

Test-only; both modules are already mypy-clean. Tests run in float32
(the production coordinate dtype).

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
…duct layer

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
…helpers

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
Replace the central pyproject ignore_errors override for rf3.data.paired_msa with a file-level '# mypy: ignore-errors' directive in the module itself. The module is broken against the installed atomworks (subclasses a now-function) and needs a PandasDataset-API refactor to clear honestly; keeping it in mypy's files scope with an in-file directive makes the suppression visible at the point of breakage and re-enables checking the moment the directive is removed.

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
…it tests

Bring src/foundry/utils/rigid.py under a per-module disallow_untyped_defs +
check_untyped_defs override (the first Track 1 direction-(b) strictness slice) by
annotating its ~47 previously-untyped Rotation/Rigid methods and quaternion/matrix
helpers. Annotation-only, no behaviour change; self-references use string forward-refs
and the identity helpers' shape params are widened to Tuple[int, ...].

Add tests/test_rigid.py: 33 fixture-backed CPU unit tests pinning the pure geometry
(quat<->matrix round-trips, compose/invert, apply, the 4x4/7-vector encodings, and
from_3_points) against independent references.

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
…g tests

Bring src/foundry/utils/components.py under the per-module disallow_untyped_defs +
check_untyped_defs override by annotating the contig parsers and mask getters.
Annotation-only apart from two behaviour-preserving refactors (collapse a str->int
variable reuse in split_contig; store fixed_parts as tuples to avoid object-typed
heterogeneous lists). Also widen get_name_mask's query_names to str | list[str] to
match the isinstance branch and docstring.

Add tests/test_components.py: 25 unit tests pinning the contig/component parsing
grammar (split_contig, extract_pn_unit_info, get_motif_components_and_breaks,
get_design_pattern_with_constraints) and the get_name_mask atom selector.

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
Bring src/foundry/utils/torch.py under the per-module disallow_untyped_defs +
check_untyped_defs override by annotating map_to's kwargs, the tracer-warning
contextmanager, assert_shape, and the Timer/Timers methods. Annotation-only; the
two warnings.filters mutations carry a documented type: ignore[attr-defined]
(typeshed types it immutable, but it is a list at runtime).

Extend tests/test_torch_utils.py with 12 tests for the untested pure helpers:
scatter_mean, assert_shape/assert_same_shape, device_of, and Timer/Timers.

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
…d unit tests

Add foundry.utils.alignment and foundry.utils.weights to the per-module
strict-mypy override (disallow_untyped_defs + check_untyped_defs) and
annotate their untyped definitions:

- alignment.py: weighted_rigid_align / get_rmsd / the dead superimpose
  stub. get_rmsd's torch.sum(..., axis=) -> dim= (torch numpy-compat
  alias; the stubs only declare dim, behaviour-identical).
- weights.py: -> None on the two __post_init__ methods.

Annotating weighted_rigid_align's params honestly as torch.Tensor
surfaced a real arg-type mismatch in rf3's dump_trajectories (param
loosely typed list[torch.Tensor | np.ndarray], body tensor-only);
resolved with documented cast(torch.Tensor, ...) at the alignment
call site.

Extend tests/test_weight_loading.py with 3 tests for documented but
untested behaviour: the '?' and '[...]' glob wildcards plus literal-dot
escaping, and the COPY-policy missing-parameter fallback chain.

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
Add foundry.utils.rotation_augmentation to the per-module strict mypy
override (disallow_untyped_defs + check_untyped_defs) and annotate its four
tensor-geometry functions. Annotation-only; already covered by
tests/test_rotation_augmentation.py.

Co-authored-by: lyskov-ai <277346777+lyskov-ai@users.noreply.github.com>
@lyskov-ai lyskov-ai requested a review from woodsh17 June 10, 2026 18:38
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

3 participants