ponder-lab / Hybridize-Functions-Refactoring / #2488

83%

push

github

Exclude numpy-call results from tensor-computation detection (#774) (#775)

The benefit precondition (#709/#712) declines a function that performs no
tensor computation. `Util.performsTensorFlowOp` detected a tensor computation
in part by any instruction that defines a tensor-typed value. Because a numpy
array is tensor-convertible, the tensor-type analysis types the result of a
numpy call (`np.array`, `np.zeros`) as a tensor, so a pure-numpy body was
misreported as computing and hybridized despite containing no TensorFlow op.

Criterion (a) now excludes a tensor-typed def whose producing instruction is a
numpy invoke (a `Lnumpy/` callee), detected by the callee's points-to type,
mirroring the modeled-op check in `invokesTensorFlowOp`. Genuine TensorFlow
ops, tensor operators, and layer calls are unaffected. Other tensor-convertible
libraries (e.g. scipy) would be handled analogously, but numpy is the only one
the tensor-type analysis models.

Surfaces on Ariadne 0.52.26, which types the numpy-array parameters of four
`deep_recommenders` Cora data-prep methods, admitting them to the benefit
precondition where the false-positive let them hybridize. Adds
`testNumpyBodyTensorComputationFP`, a distilled repro: a numpy-array-fed
parameter with a pure-numpy body is declined `NO_TENSOR_COMPUTATION`, while a
`tf.reduce_sum` control still computes.


Claude-Session: https://claude.ai/code/session_01He1ULx8nRoHzGicMryry1A

Co-authored-by: Claude Opus 4.8 (1M context) <noreply@anthropic.com>

4 of 4 new or added lines in 1 file covered. (100.0%)

2138 of 2612 relevant lines covered (81.85%)

0.82 hits per line