ponder-lab / Hybridize-Functions-Refactoring / #3274

87%

main: 88%

push

github

Decline hybridizing a custom-gradient method with a bound argument (#1006)

* Decline hybridizing a custom-gradient method with a bound argument

Under `tf.function`, `tf.custom_gradient` runs in graph mode and
converts every argument to a tensor, so an instance method's `self` or a
classmethod's `cls` raises. With `tf.function` beneath
`tf.custom_gradient` instead, the gradient function's return is
rejected. No decorator order works, so CONVERT_TO_HYBRID is declined
with the new IS_BOUND_CUSTOM_GRADIENT_METHOD (32), and an already-hybrid
such method is reported rather than reconfigured. A staticmethod and a
plain function have no bound argument and are still hybridized above
`tf.custom_gradient` (#996).

The test is structural, so it holds whether or not the analysis reaches
the method.

Fixes #1000.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

* Report a bound custom-gradient method in place of the optimal verdict

With inference on, an already-hybrid bound custom-gradient method was
reported with both IS_BOUND_CUSTOM_GRADIENT_METHOD and
HAS_NO_PRIMITIVE_PARAMETERS. The latter is the already-optimal verdict,
which is false for a method whose decorator fails on every call, so the
bound method now reports in its place, as other operative failures in
that block do.

The predicate's Javadoc now names its known limits, and the fixture
comment no longer names an exception type the program doesn't raise.

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
Claude-Session: https://claude.ai/code/session_01NTnpjPjJj2aZy1ZkUD1eAn

---------

Co-authored-by: Claude Opus 5.5 <noreply@anthropic.com>

25 of 25 new or added lines in 2 files covered. (100.0%)

4306 of 4942 relevant lines covered (87.13%)

0.87 hits per line