Failed to legalize torch.operator "onnx.Expand"

[vitreo12]

What happened?

Hello!

I have been trying to iree-compile a converted onnx model, but I hit this error:

error: failed to legalize operation 'torch.operator' that was explicitly marked illegal
    %1806 = torch.operator "onnx.Expand"(%1804, %1805) : (!torch.vtensor<[],f32>, !torch.vtensor<[?],si64>) -> !torch.vtensor<[],f32> 

Looking at nod-ai/SHARK-ModelDev#215, it looks like Expand should be supported, since it's marked as completed. Is this the case? I tried with both the latest stable release and the most recent release candidate.

Steps to reproduce your issue

1. Go to '...'
2. Click on '....'
3. Scroll down to '....'
4. See error

What component(s) does this issue relate to?

No response

Version information

No response

Additional context

No response

[ScottTodd]

We have two tests (from the upstream ONNX test suite) for onnx.Expand:

• https://github.com/nod-ai/SHARK-TestSuite/blob/main/iree_tests/onnx/node/generated/test_expand_dim_changed/model.mlir
• https://github.com/nod-ai/SHARK-TestSuite/blob/main/iree_tests/onnx/node/generated/test_expand_dim_unchanged/model.mlir

Both are passing on all in-tree backends. If that specific op is the problem, the dynamic dimension (the ? in the shape of the second arg, %1805) might be unsupported.

More context about the compilation would help:

• The source program as a .mlir file (if you can share it)
• Which pass failed (either by looking at the --mlir-print-ir-before-all output, --mlir-print-ir-after-failure output, or some other mechanism)

[vitreo12]

The entire mlir file is 1.5gb, including the weights. Perhaps it's enough that I share the graph though? It's still quite a big graph, generated from an onnx file. Here it is: https://pastebin.com/FqJWwhaA