💊 CI failures summary and remediations

As of commit 681be55 (more details on the Dr. CI page):

✅ None of the CI failures appear to be your fault 💚

• 1/1 broken upstream at merge base 8ecfa9d on Aug 25 from 9:36am to 12:34pm PDT (8 commits; 6a2d7a0 - 9420c77)

🚧 1 fixed upstream failure:

These were probably caused by upstream breakages that were already fixed.

git fetch https://github.com/pytorch/pytorch viable/strict
git rebase FETCH_HEAD

• pytorch_linux_xenial_cuda10_2_cudnn7_py3_gcc7_test on Aug 25 from 9:36am to 12:34pm PDT (8 commits; 6a2d7a0 - 9420c77)
  • 🔁 rerun

Hmm, still some inscrutable JIT errors that need fixing. Part of the issue is that get_type_line() in torch/jit/annotations.py seems to get confused by # type: ignore comments. I'll see if I can improve that.

Okay, isolated the issue and a proposed way of fixing it in gh-43454. This PR is blocked until that's resolved.

Hmm, there's one CI failure (in caffe2_onnx_ort2_py3_6_clang7_ubuntu16_04_test) that seems to be caused by this PR:

self = <test_operators.TestOperators testMethod=test_meshgrid>

    def test_meshgrid(self):
        x = torch.ones(3, requires_grad=True)
        y = torch.zeros(4, requires_grad=True)
        z = torch.ones(5, requires_grad=True)
>       self.assertONNX(lambda x, y, z: torch.meshgrid(x, y, z), (x, y, z))

test/onnx/test_operators.py:849: 
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ 
test/onnx/test_operators.py:66: in assertONNX
    self.assertExpected(onnx_model_pbtxt, subname)
../.local/lib/python3.6/site-packages/torch/testing/_internal/common_utils.py:1305: in assertExpected
    self.assertMultiLineEqual(expected, s)
E   AssertionError: 'ir_v[355 chars]ut: "x"\n    input: "3"\n    output: "4"\n    [4894 chars]n}\n' != 'ir_v[355 chars]ut: "0"\n    input: "3"\n    output: "4"\n    [4894 chars]n}\n'
E   Diff is 5863 characters long. Set self.maxDiff to None to see it.

Likely a matter of updating the expected result, I'll look into it.

EDIT: yes, I can reproduce locally. The tweak to call a _meshgrid seems to cause input name changes like:

    node {
-     input: "x"
?             ^
+     input: "0"
?             ^

This looks good to go now, the one CI failure is unrelated.

@malfet merged this pull request in 573940f.