ponder-lab / Hybridize-Functions-Refactoring / #2716

84%

push

github

Block hybridizing calls to the Keras training surface (#848)

* Block hybridizing calls to the Keras training surface

The training-surface members (fit, predict, evaluate, the on-batch
family, and the deprecated generator wrappers) are guarded by Keras's
_disallow_inside_tf_function: each raises RuntimeError inside a
tf.function trace, runtime-verified per member, because they are
training-loop orchestrators managing their own traces rather than
tensor operations. The eager-only check previously knew only numpy, so
hybridization could convert such calls into guaranteed-crashing code.

The names are generic ML verbs with in-corpus collisions (a
test-framework base class's evaluate; user-defined predict overrides),
so blocking requires dispatch evidence: a resolved call-graph target in
the TensorFlow namespace blocks; a resolved user-defined override does
not, since the transitive walk analyzes its body on its own merits; an
unresolved call site blocks only when the receiver's points-to holds a
summarized framework instance, and the remaining residue allows
(wala/ML#571 keeps a class-hierarchy walk from identifying user model
subclasses).

The whole eager-only walk moves from Util into a dedicated
EagerOnlyCallAnalysis, joining the sibling analysis classes; the
member-name vocabulary is staged for the framework profile of #844.

Closes #836.

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

* Document the by-design tf.io and tf.train benefit exclusion

The whole-namespace exclusion is deliberate although both namespaces
mix builders with tensor-producing ops: the benefit signal asks whether
graph execution accelerates the function's work, and parsing or
decoding input is preprocessing, typically per-element work living in
tf.data pipelines that trace their map functions independently. The
corpus tally backs the premise (roughly 120 builder and checkpoint
references against 6... (continued)

41 of 42 new or added lines in 2 files covered. (97.62%)

3050 of 3647 relevant lines covered (83.63%)

0.84 hits per line