Fix meshgrid converter to handle non-1D inputs

[cdeil]

I am trying to use https://github.com/roboflow/rf-detr and convert the model to CoreML for use on iPad computer vision app. I'm using some of the fixes from https://github.com/landchenxuan/rf-detr-to-coreml and after upgrading to latest PyTorch 2.11 am getting good performance! :-)

There's one issue with torch.meshgrid that needs a small fix. For now I have a helper function to monkey-patch coremltools with this fix as workaround.

See description and code/test in this PR - done by coding agent.

Would be great if you could include it.

---

Summary

The meshgrid op converter raises ValueError("meshgrid received non-1d tensor.") when
inputs have rank > 1. This happens in practice when converting models that use
torch.linspace(...) / scalar before torch.meshgrid(), because the MIL real_div op
can broadcast the 1D linspace result with a 0D scalar divisor to produce a higher-rank
tensor.

This pattern is common in deformable attention modules (DINO, Deformable-DETR, RF-DETR)
where coordinate grids are created:

grid_y = torch.linspace(0.5, h - 0.5, steps=h) / h
grid_x = torch.linspace(0.5, w - 0.5, steps=w) / w
grid_y, grid_x = torch.meshgrid(grid_y, grid_x, indexing="ij")

The PyTorch JIT trace shows 1D tensors flowing into meshgrid, but during MIL conversion
the division by a 0D shape-derived scalar produces a result with rank > 1 in the MIL IR,
causing the meshgrid converter to reject it.

Changes

coremltools/converters/mil/frontend/torch/ops.py

• Removed the hard rejection of non-1D inputs (raise ValueError("meshgrid received non-1d tensor."))
• Added _flatten_inputs() helper that reshapes any non-1D inputs to 1D via mb.reshape(shape=[-1]) before meshgrid processing
• This is safe because meshgrid semantically requires 1D-like inputs, and the flatten simply recovers the intended shape

coremltools/converters/mil/frontend/torch/test/test_torch_ops.py

• Added test_meshgrid_non_1d_inputs to TestMeshgrid class
• Tests the deformable attention pattern (linspace / scalar → meshgrid) with both ij and xy indexing modes
• Parametrized across compute units and backends, consistent with existing test style

Testing

Red/green verified:

• RED: Without the fix, the new test fails with ValueError: meshgrid received non-1d tensor.
• GREEN: With the fix, all meshgrid tests pass (both existing and new)

4 passed (2 existing + 2 new) — no regressions

Related Issues

• Partially addresses #1674 (meshgrid conversion failures)
• Related to #1268 (previous meshgrid input fix)
• Support PyTorch 2.9 #2615

[TobyRoseman]

Code changes look good.

CI: https://gitlab.com/coremltools1/coremltools/-/pipelines/2418650256