coreai-torch v0.4.1

Note:

• .aimodel artifacts converted with coreai-torch v0.4.0 will fail to load/specialize on-device starting with OS 27 second beta onwards. Reconvert your model using coreai-torch v0.4.1 or later to produce a compatible artifact.

Changes:

• Added support for masked_scatter (writing values into a tensor at masked positions)
• repeat op lowering now supports dynamic/symbolic repeat counts (not just fixed integers)
• Mixed SymInt/int shape lowering across pow, round, upsample, cat, and repeat
• Support concat (cat) lowering when a non-concat axis stays dynamic after shape promotion
• For arange op lowering, preserve static output shape for float dtypes with integer bounds
• Skip externalization for submodules never invoked in the exported graph
• Reworked debug_info and inspector modules for improved debug location tracking and intermediate value inspection
• Bumped coreai-core dependency to 1.0.0b2

coreai-core v1.0.0b2 release notes:

• NDArray Python bindings now support strided views.
• AIProgram.save_asset now records the producer in asset metadata, overwrites an existing serialized model instead of failing, and validates that the destination has a .aimodel extension.
• Added string representations for SpecializationOptions, ComputeUnitKind, and InferenceFunction.
• Fixed an import cycle in the Python package.
• NDArray.View/MutableView.subscript(scalarAt:) now a public API.
• Added image type support to InferenceValue.Kind.
• Documentation polish for InferenceFunction.encode.
• Removed the unused ml_asset module and the legacy Torch importer..