ponder-lab / Hybridize-Functions-Refactoring / #2657

84%

main: 84%

Build Type

Pull #831

github

Committed by khatchad

Commit Message

Add a symbolic-iteration safety precondition

Iterating a tensor parameter with a Python loop works eagerly, since the elements are tensors, but raises `OperatorNotAllowedInGraphError` under `tf.function` tracing once the parameter is symbolic, even with AutoGraph converting the loop (runtime-verified on TensorFlow 2.9.3). MusicTransformer's `EventSeq.from_array` is the live case: through Ariadne 0.52.36 its barren verdict accidentally shielded the conversion, and 0.52.76's element-typing improvement removed the shield without adding the guard, which the 0.52.76 verification audit surfaced (#830).

`TensorIterationAnalysis` scans each body: parameter value numbers seed a forward def-use taint, and an `EachElementGetInstruction` whose iterated object is tainted and tensor-typed fires `HAS_TENSOR_PARAMETER_ITERATION` (code 20), the sixth safety failure in the family. The tensor-typed gate keeps ordinary collection parameters out, and the parameter-derived anchor keeps AutoGraph-supported in-body `tf.range` loops convertible. `alwaysCheckTensorIteration` mirrors the sibling knobs end-to-end.

`testTensorParameterIterationBlocksHybridization` pins the `from_array` distillation declining and the `tf.range` control converting; the fixture runs to completion eagerly, which is exactly the shape that made the hazard invisible without tracing.

Co-Authored-By: Claude Fable 5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01GcqABsog1rtb4m2CDZdcRD

Pull Request Pull Request #831: Add a symbolic-iteration safety precondition

Coverage Stats

49 of 57 new or added lines in 4 files covered. (85.96%)

36 existing lines in 1 file now uncovered.

2968 of 3554 relevant lines covered (83.51%)

0.84 hits per line