ONNX CI worflow is broken

[pmeier]

Failure is visible since this morning: https://app.circleci.com/pipelines/github/pytorch/vision/17824/workflows/63ed118f-e479-47e5-91c1-2f0805ce2e5a/jobs/1443424

RuntimeError: Unsupported: ONNX export of operator prim::device, output type is not `_C.DeviceObjType`.. Please feel free to request support or submit a pull request on PyTorch GitHub.

The last commit before the weekend b969cca did not have this failure. Thus, I'm guessing we only have look for the nightlies of the 21th and 22th for possible offenders.

cc @neginraoof @seemethere

[pmeier]

The offender should be between pytorch/pytorch@d1bb420 and pytorch/pytorch@fbc417b inclusive.

The only PR that references ONNX is pytorch/pytorch#77423, but that seems unrelated. The only mention of DeviceObjType is in the symbolic_optset9.py module, but IIUC the only thing that changed was instead of importing directly from torch._C import DeviceObjType now we import torch._C and use it as _C.DeviceObjType.

[pmeier]

I bisected and it turns out pytorch/pytorch#77423 is the culprit after all @justinchuby @BowenBao

[pmeier]

pytorch/pytorch@0d76299 made it into the branch cut and thus we need to resolve before the release.

[justinchuby]

Updated:

• [ONNX] Fix typo when comparing DeviceObjType pytorch#78085

[pmeier]

Fix worked! https://app.circleci.com/pipelines/github/pytorch/vision/17866/workflows/49b2a6cf-c54e-4c21-8a0d-a477e70c884f/jobs/1446667

Thanks @justinchuby!