ponder-lab / Hybridize-Functions-Refactoring / #2309

80%

main: 80%

Build Type

Pull #712

github

Committed by khatchad

Commit Message

Walk through TensorFlow summary nodes in the tensor-op detector

`performsTensorFlowOp` pruned the entire subtree at a TensorFlow library node; a higher-order summary (`strategy.run`, `dataset.map`) invokes user callbacks whose ops live behind it, so the prune hid them. The walk now skips only the library node's own body scan and continues through its successors, which are per-node and hence reach exactly the callbacks this function passes in.

Corpus-verified on the 0.52.15 overlay regeneration: the fix eliminates the one P6 false de-hybridization (`distributed_train_step` in TensorFlow2.0-Examples, a developer-written decorator previously stripped), leaving that subject's overlay byte-identical to the main-line pristine; the function now correctly fails only the de-hybridization preconditions and stays hybrid. The `__dist_train_step` pair (MusicTransformer) is out of this fix's reach: its receiver is the Keras-internal `_distribution_strategy`, for which no invoke edge exists to walk.

`testTensorComputationThroughStrategyRun` guards the `strategy.run` reach. The guard is green even without the walk change when the call-graph edge exists (the summary's result typing fires the body criterion first), so the corpus gate is the regression backstop for the walk itself.

Refs #714.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01S1U9TQRo3Wmhfu86DRcBNW

Pull Request Pull Request #712: Add benefit preconditions for barren functions

Coverage Stats

121 of 129 new or added lines in 5 files covered. (93.8%)

10 existing lines in 1 file now uncovered.

1756 of 2199 relevant lines covered (79.85%)

0.8 hits per line