CI break: ATenlib matmul is flaky

[justinchuby]

Test for matmul is flaky: sometimes the difference is NaN: https://github.com/microsoft/onnx-script/actions/runs/4470190371/jobs/7853362946#step:5:3924

This starts to happen when we use PyTorch 2.0

[BowenBao]

Flakiness introduced by empty tensor. Looks like undefined behavior for outputs where nan is possible. If values are all zeros the test passed.

onnx inputs:  [array([], shape=(5, 5, 0), dtype=float32), array([], shape=(0, 5), dtype=float32)] {}
onnx outputs:  [array([[[ 0.,  0., -0.,  0., -0.],
        [ 0., -0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.]],

       [[ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.]],

       [[ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.]],

       [[ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.]],

       [[ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.],
        [ 0.,  0.,  0.,  0.,  0.]]], dtype=float32)]

@justinchuby thought about this again. I think we don't need to care about this particular test case. Feel free to skip.

[titaiwangms]

pytorch/pytorch#117718 aten_mm might might have the same issue in torchlib.