ponder-lab / Hybridize-Functions-Refactoring / #2269

80%

main: 80%

Build Type

Pull #712

github

Committed by khatchad

Commit Message

Recognize TensorFlow ops via import alias when module global points-to is empty

The NO_TENSOR_COMPUTATION detector (#709) resolved a callee's fully-qualified
name by walking its attribute chain back to the TensorFlow module via
points-to. In whole-program analysis the `tf` module global frequently has an
empty points-to set, so every `tf.<op>` call went unresolved and genuine
forward passes (e.g. `OutputLayer.call`, whose body is `tf.matmul`/`tf.reshape`/
`tf.shape`) were misreported as performing no tensor computation and blocked
from hybridizing.

`Util.isTensorFlowModule` now roots the attribute chain at TensorFlow by
points-to or, when the module global's points-to set is unavailable, by the
module import alias on the global read. The fallback is incompleteness-safe:
over-recognizing a tensor op only lets an eager function hybridize (the
pre-signal default). The existing `tensorflow.train.`/`tensorflow.io.`
exclusion still keeps proto and spec builders barren.
`Function.computeTensorComputation` unions over all call-graph nodes rather
than sampling the first.

`testGpt2GetLossVendored` gains a guard that `OutputLayer.call` performs a
tensor computation, reproducing the empty-points-to condition on the vendored
subject; the assertion fails on the pre-fix resolver.

The residual sub-layer-dispatch false positives are tracked in #714, which
blocks #709.

Refs #709.

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01FuhM8SsTRB6BraLt3Ph7ZC

Pull Request Pull Request #712: Add a benefit precondition for barren eager functions

Coverage Stats

81 of 87 new or added lines in 4 files covered. (93.1%)

114 existing lines in 3 files now uncovered.

1741 of 2182 relevant lines covered (79.79%)

0.8 hits per line