ponder-lab / Hybridize-Functions-Refactoring / #2544

82%

main: 83%

Build Type

Pull #792

github

Committed by khatchad

Commit Message

Reduce sequence containers to nested input-signature entries

A parameter receiving a Python list or tuple of tensors previously blocked
input-signature inference wholesale (`TENSOR_CONTAINER_UNSUPPORTED`): Phase 3
classified it as a tensor container while `getTensorContainers` discarded the
element types the analysis holds (`pair.snd`). TensorFlow accepts and enforces a
nested signature for exactly this shape (verified on 2.9.3: a list of specs
inside the outer list; a bare tensor against it raises `TypeError`, a length
mismatch raises `ValueError`, and retracing across distinct singleton-list calls
collapses to one trace), so the elements are now surfaced and reduced.

`InputSignature` models a parameter's contribution as a sealed `SpecEntry`:
`Single` (one `TensorType`, the pre-existing shape) or `Sequence` (per-element
types, rendering as a nested spec list). Rendering, required-name collection,
and `relate()` dispatch over both; a flat entry against a nested one is
incomparable in either direction, since the two admit disjoint call structures,
and two sequences of different arity likewise (no wildcard arity exists).
`ofSingles` keeps the flat construction path (a constructor overload is
impossible: `List<TensorType>` and `List<SpecEntry>` erase identically), and
`singleTypes()` is the flat projection consumers that predate nesting use. The
supplied-signature parser still produces flat signatures only, so a nested
inferred signature against a developer-written flat one reconfigures via the
incomparable path.

`Parameter` now caches container element evidence per analysis (container
instance key to field name to element types) instead of the bare key set, and
`extractContainerElements` surfaces a parameter's per-position element types
when the form is modeled: every reaching value a list or tuple known to the
analysis, catalogs a contiguous constant run 0..n-1, arities agreeing, every
position evidenced. Disagreeing arities set their own signal; an... (continued)

Pull Request Pull Request #792: Reduce sequence containers to nested input-signature entries

Coverage Stats

114 of 138 new or added lines in 5 files covered. (82.61%)

1 existing line in 1 file now uncovered.

2308 of 2816 relevant lines covered (81.96%)

0.82 hits per line