[CoreML EP] Support Gather with scalar indices on the MLProgram path

[john-rocky]

Fixes #28180.

PyTorch's ONNX exporter routinely emits Gather with a rank-0 (scalar) indices input — every per-layer style-code lookup in StyleGAN / StyleGAN2 derivatives is one. In the GFPGAN 1024×1024 generator that's 16 such Gathers, each splitting the CoreML subgraph and forcing a partition boundary. GatherOpBuilder::IsOpSupportedImpl historically rejected this shape outright.

The actual blocker isn't MIL gather

MIL gather does accept rank-0 indices and produces the expected output (gathered axis dropped). Verified by building a small MIL program end-to-end against iOS 15:

@mb.program(input_specs=[mb.TensorSpec(shape=(1, 16, 512), dtype=types.fp32)])
def prog(x):
    return mb.gather(x=x, indices=mb.const(val=np.int32(3)), axis=1)

main[CoreML3](%x: (1, 16, 512, fp32)) {
  %gather_0: (1, 512, fp32) = gather(x=%x, indices=3, axis=1)
}

The reason the EP rejected the case was ORT-side boundary handling, not MIL:

• ModelBuilder::RegisterModelInputOutput rewrites any rank-0 boundary tensor to {1} because MLMultiArray has no rank-0 representation.
• On the NeuralNetwork path, RegisterInitializers does the same for rank-0 initializers (LoadConstantND requires rank ≥ 1).

But constant initializers on the MLProgram path flow through OnnxTensorToCoreMLTensor, which preserves rank — so MIL gather can consume them directly.

Change

IsOpSupportedImpl: allow rank-0 indices on the MLProgram path only when it is a constant initializer. NeuralNetwork path and non-initializer scalar indices remain rejected.

AddToModelBuilderImpl: unchanged from main — the existing code already passes indices straight through to MIL gather.

Test

Existing CPU OpTester cases already cover the shape and run against every registered EP:

• GatherOpTest.Gather_axis0_scalar_indices (data [2,2,2], scalar index, axis 0)
• GatherOpTest.Gather_axis1_scalar_indices (data [2,2,2], scalar index, axis 1)

Both were skipped on CoreML before (IsOpSupportedImpl returned false); with this change they stay on CoreML EP and the OpTester output matches the reference.

[Copilot]

Pull request overview

Note

Copilot was unable to run its full agentic suite in this review.

Adds CoreML MLProgram-path support for ONNX Gather with scalar (rank-0) indices by rewriting it inside the CoreML builder to avoid CPU fallback and partition splits (fixes #28180).

Changes:

• MLProgram path: detect scalar indices and emit reshape([1]) -> gather -> squeeze(axes=[axis]).
• Support check: allow scalar indices only on MLProgram and adjust effective output-rank calculation accordingly.
• Minor refactor: use input_defs / output_defs locals and pass logger through.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[john-rocky]

Apologies — I missed @maxwbuckley's earlier #28278 when looking at this issue. Their PR covers both the NN and MLProgram paths and is already in review with @yuslepukhin, so closing this in favor of theirs.