f-dangel / curvlinops / 28679556886

92%

main: 92%

Build Type

Pull #308

github

Committed by f-dangel

Commit Message

[TEST] Cover reshape-wrapped bias detection in the FX layer-IO matcher

Add tests for the pattern the previous commit enables -- a Linear whose bias is
added after transparent, last-dim-preserving reshape ops, and a single weight
reused across matmuls of uneven reshape depth (as emitted by collapsed
Taylor-mode / jet graphs, one matmul per Taylor coefficient).

- test_param_io.py:
  - test_bias_add_through_transparent_reshapes: bias reached through
    `mm -> unsqueeze -> squeeze -> add(bias)` is linked to its layer.
  - test_weight_reused_across_reshape_wrapped_matmuls: one weight over three
    coefficient matmuls (only the forward one biased, via a reshape chain) of
    differing reshape depth -- exercises both the transparent-bias traversal and
    the prefix-coverage check in verification.py.
  - test_bias_after_last_dim_altering_reshape_not_matched: a reshape that alters
    the last (feature) dimension must NOT absorb the bias -- guards against
    over-matching through non-transparent shape ops.
- test_kfac_io.py:
  - test_kfac_io_weight_tying_through_reshaped_bias: end-to-end KFAC IO grouping
    for the tied-weight, reshape-wrapped-bias structure across all Fisher types.

All three positive tests fail on the parent commit and pass with the matcher
extension; the negative test passes on both.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>

Pull Request Pull Request #308: [ADD] Detect Linear weights/biases reused across reshape-wrapped matmuls

Coverage Stats

34 of 36 new or added lines in 1 file covered. (94.44%)

1 existing line in 1 file now uncovered.

2667 of 2888 relevant lines covered (92.35%)

0.92 hits per line