ponder-lab / Hybridize-Functions-Refactoring / #2371

83%

push

github

Add an eager-only-call safety precondition (#741)

A function whose body (transitively) invokes an eager-only API such as `Tensor.numpy()` crashes on first call under `@tf.function` tracing, so hybridizing it does not preserve semantics. `Util.callsEagerOnlyApi` mirrors the `performsTensorFlowOp` walk (#712): user bodies only, transitive over callees, walking through TensorFlow library nodes to reach callbacks. Detection is by callee attribute name (`numpy`) rather than receiver typing: the receiver's tensor typing is frequently unavailable (e.g. the result of a user-defined callable), and a missed call hybridizes a crasher while an over-match only declines an optimization.

The failure (`HAS_EAGER_ONLY_CALLS`, code 15) is checked ahead of the barren benefit signal: safety precedes benefit. An undetermined result (no call-graph node) neither blocks nor passes, mirroring the recursion and tensor-computation checks. The computation shares the tensor-parameter-candidate gate; `alwaysCheckEagerOnlyCalls` forces it on every function for corpus-wide measurement, mirroring `alwaysCheckTensorComputation`, wired through `EvaluationOption`, the evaluator handler, `run-headless-evaluator.sh` (`CHECK_EAGER_ONLY_CALLS`), and new "eager-only calls" columns in `functions.csv` and the settings record.

`testEagerOnlyCallsBlockHybridization` pins the direct and transitive cases, the still-hybridizable clean sibling (P1), and that an eager-only call on a function's result is attributed to the caller, not the function.


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

Co-authored-by: Claude Fable 5 <noreply@anthropic.com>

39 of 45 new or added lines in 4 files covered. (86.67%)

1827 of 2278 relevant lines covered (80.2%)

0.8 hits per line