[john-rocky]

Summary

x.squeeze(1) on a [1, 4, 8] tensor is a no-op in PyTorch, but the exported aten.squeeze.dims node still lists dim 1. replace_squeeze_dims passed every listed dim to coreai.shrink_dims, which rejects a dim of size 4 (ValueError: Operation creation failed). With torch 2.13.0, x.squeeze() decomposes to squeeze.dims over every dim and fails the same way. A listed dynamic dim, which PyTorch keeps, was shrunk too, and conversion failed (squeeze[0]: rank 2 vs 3).

The lowering now shrinks only the listed dims whose size is a static 1 in the exported input, and returns the input when there are none.

Testing

test_squeeze_dims_not_size_one runs torch.squeeze(x, [1]), torch.squeeze(x, [0, 1]) and x.squeeze() on [1, 4, 8], static and with dims 1 and 2 dynamic, against PyTorch. All six fail before the fix and pass after it.

uv run pytest tests/ops/test_ops.py -k squeeze   # 50 passed
uv run pytest tests/ -n 4                        # 6560 passed; 3 failed in tests/debugging/, as on main
uv run --extra dev ruff check .                  # All checks passed
uv run --extra dev ruff format --check .         # all files formatted

Fixes #143