ONNXRuntimeError with a model that passes onnx.checker and was supported on previous ort versions

[cestpasphoto]

Describe the issue

I have a model exported by pytorch which was supported until ort 1.16.3 but now fails with ort 1.17.0: onnxruntime.capi.onnxruntime_pybind11_state.Fail: [ONNXRuntimeError] : 1 : FAIL : Node (MatMulBnFusion_Gemm) Op (Gemm) [ShapeInferenceError] First input does not have rank 2

I have tried several versions of pytorch, several config of torch.onnx.export but always fail. As soon as I downgrade to 1.16.3, it works whatever the pytorch version is, so that makes me think that ort is the culprit. Plus I've run onnx.checker.check_model(mymodel, full_check=True) and raised no issue. It happens even with pure CPU environment.

The following code passes with ort 1.16.3 but returns an error with 1.17.0

To reproduce

import onnx
import onnxruntime as ort

file = 'example_onnx_file.onnx'
mymodel = onnx.load(file)
onnx.checker.check_model(mymodel, full_check=True) # No error

ort.InferenceSession(file, providers=['CPUExecutionProvider']) # Raise an error

Urgency

No response

Platform

Linux

OS Version

Debian stable

ONNX Runtime Installation

Released Package

ONNX Runtime Version or Commit ID

1.17.0

ONNX Runtime API

Python

Architecture

X64

Execution Provider

Default CPU

Execution Provider Library Version

No response

[cestpasphoto]

Please use the following file: https://github.com/cestpasphoto/alpha-zero-general/blob/master/splendor/example_onnx_file.onnx

To answer the auto-label, CUDA is NOT involved here, it happens with CPU-only environment.