ponder-lab / Hybridize-Functions-Refactoring / 27223192097

79%

push

github

Parse the keyword-form `input_signature` argument of `@tf.function` (#593)

* Parse the keyword-form `input_signature` argument of `@tf.function` (#557)

Read the content of a keyword-form `input_signature=[tf.TensorSpec(...)]`
decorator argument into an `InputSignature` (ordered `TensorType`s) and expose
it via `HybridizationParameters.getSuppliedInputSignature()`. Previously only
the argument's presence was detected (`hasInputSignatureParam()`); its content
was never read.

The parse is all-or-nothing: any element that cannot be fully modeled — an
unsupported `TensorSpec` subtype (`RaggedTensorSpec`/`SparseTensorSpec`, tracked
by #524 and #533), a non-`TensorSpec` callee, or malformed content — drops the
whole signature to empty. The getter Javadoc documents the resulting
presence/parse contract so downstream reconfiguration distinguishes "no
signature supplied" (safe to infer and write) from "supplied but unmodelable"
(leave as-is), keying off `hasInputSignatureParam()` rather than the `Optional`
alone.

Scope is the keyword form only; positional binding is #573, and wiring the
parsed content into inference is #563. No inference behavior changes here.

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

* Parse a well-formed empty `input_signature=[]` as a present, empty signature

`input_signature=[]` (a no-arg function) is well-formed and fully modeled, so
it parses to a present, empty `InputSignature` rather than dropping to empty.
Empty is reserved for a supplied signature that cannot be modeled, per the
getter's presence/parse contract.

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

---------

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

54 of 63 new or added lines in 1 file covered. (85.71%)

1383 of 1777 relevant lines covered (77.83%)

0.78 hits per line