Does inline_local_functions preserve ir_version?

[xadupre]

Bug Report

onnxruntime complains about a wrong ir_version. Maybe the one from the orginal model is not preserved?

System information

• OS Platform and Distribution (e.g. Linux Ubuntu 20.04): linux
• ONNX version: 1.16
• Python version: 3.10

[gramalingam]

Based on a quick look, the code does not touch fields like ir_version when inlining. It seems unlikely to be happening during inlining. Is there an onnx model that causes a problem for reproduction?

[xadupre]

Every model used in unit tests usuaaly specifies the ir_version to avoid onnxruntime to fail because it does not support the new one yet. Any model already used in the existing unit tests is fine. This is how I found it: https://github.com/sdpython/onnx-extended/blob/main/_unittests/ut_tools/test_onnx_inline.py#L58.

[gramalingam]

@xadupre : I looked at the example in the above link. It is an invalid ONNX model: subgraphs cannot use outer-scope names as the output-names (and expect those values to be returned). Instead, it is expected that the subgraph will have at least one node such as subgraph_output = Identity(outer_scope_name).

I realize you do run the checker. I am guessing it doesn't catch it. If so, we should fix it in the checker and produce an error message.

I don't think this is connected to IR version.