torch.export with nn.Transformer creates a non-contiguous memory tensor for aten.view

[yushangdi]

There is an issue in pytorch/pytorch repo (pytorch/pytorch#159126) reporting an issue with using torch.export.export with torch_mlir on a transformer layer model.

After looking into it a bit more, the exported model itself appears to be fine on the PyTorch side. I'm re-posting here in case someone with more experience on the torch-mlir side might have thoughts on how to address this.

error message:

/lib/python3.11/site-packages/torch/_refs/__init__.py", line 3754, in _reshape_view_helper
    raise ValueError(msg)
ValueError: Cannot view a tensor with shape torch.Size([10, 4, 1, 64]) and strides (64, 640, 640, 1) as a tensor with shape (40, 64)!
While executing %view_12 : [num_users=1] = call_function[target=torch.ops.aten.view.default](args = (%permute, [40, 64]), kwargs = {})

Please see the original issue for more details, including repro steps.

[miladm]

Not sure who the right person would be for this issue... perhaps someone from the AMD staff?

cc my friend, @silvasean, in case he knows the right owner to point us to.