[ONNX] Deprecate operators.py

[justinchuby]

Deprecate torch.onnx.operators because it's only for backwards compatibility

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/87798

• 📄 Preview Python docs built from this PR
• 📄 Preview C++ docs built from this PR
• ❓ Need help or want to give feedback on the CI? Visit our office hours

Note: Links to docs will display an error until the docs builds have been completed.

❌ 2 Failures

As of commit 7f12f2e:

The following jobs have failed:

• linux-bionic-cuda11.6-py3.10-gcc7 / test (functorch, 1, 1, linux.4xlarge.nvidia.gpu)
• linux-bionic-cuda11.7-py3.10-gcc7 / test (slow, 1, 2, linux.4xlarge.nvidia.gpu)

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[BowenBao]

hmm is this not bc breaking?

[justinchuby]

Looks like its used somewhere else. Taking a closer look

[justinchuby]

Hmm. Looks like its still used in caffe2 models. I will add a deprecation message instead?

[BowenBao]

lgtm

[justinchuby]

@pytorchbot merge -g

[pytorchmergebot]

Merge started

Your change will be merged once all checks on your PR pass since you used the green (-g) flag (ETA: 0-4 Hours).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

[pytorchmergebot]

Merge failed

Reason: The following mandatory check(s) failed (Rule ONNX exporter):

• pull

Dig deeper by viewing the failures on hud

Details for Dev Infra team

Raised by workflow job

[pytorch-bot]

❌ 🤖 pytorchbot command failed:

@pytorchbot: error: argument command: invalid choice: 'All pull tests passed; trunk flaky' (choose from 'merge', 'revert', 'rebase', 'label')

usage: @pytorchbot [-h] {merge,revert,rebase,label} ...

Try @pytorchbot --help for more info.

[justinchuby]

@pytorchbot merge -f "All pull tests passed; trunk flaky"

[pytorchmergebot]

Merge started

Your change will be merged immediately since you used the force (-f) flag, bypassing any CI checks (ETA: 1-5 minutes).

Learn more about merging in the wiki.

Questions? Feedback? Please reach out to the PyTorch DevX Team

Advanced Debugging

Check the merge workflow status
here

[weiwangmeta]

@pytorchbot revert -m "breaking internal builds see D40797126" -c ghfirst

[pytorchmergebot]

@pytorchbot successfully started a revert job. Check the current status here.
Questions? Feedback? Please reach out to the PyTorch DevX Team

[pytorchmergebot]

@justinchuby your PR has been successfully reverted.

[justinchuby]

@weiwangmeta could you post the errors from D40797126? Thanks!

[weiwangmeta]

@weiwangmeta could you post the errors from D40797126? Thanks!

Sorry for the delayed response. The errors stemmed from
https://github.com/facebookresearch/fairseq/blob/main/fairseq/modules/sinusoidal_positional_embedding.py#L68
and it was like

RuntimeError: 
undefined value torch:
...
@_deprecation.deprecated("1.14", "1.16", "use torch._shape_as_tensor")
def shape_as_tensor(x):
    return torch._shape_as_tensor(x)
           ~~~ <--- HERE
'shape_as_tensor' is being compiled since it was called from 'SinusoidalPositionalEmbedding.forward'
  File "....../fairseq/modules/sinusoidal_positional_embedding.py", line 68
    ):
        """Input is expected to be of size [bsz x seqlen]."""
        bspair = torch.onnx.operators.shape_as_tensor(input)
        ~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~ <--- HERE
        bsz, seq_len = bspair[0], bspair[1]
        max_pos = self.padding_idx + 1 + seq_len

Stack trace:
Exception type: torch::jit::ErrorReport

[weiwangmeta]

I realize this PR is all about BC breaking and it is not surprising to see the above failures. But looks like we have internal usage that is still relying on the old API.

cc @osalpekar @mehtanirav @seemethere How should we embrace this deprecation?

[justinchuby]

I wonder why torch would be an undefined value? Is this running jit.script?

[weiwangmeta]

Looks like so. Something like this JIT test: https://github.com/facebookresearch/fairseq/blob/acd9a53607d1e5c64604e88fc9601d0ee56fd6f1/tests/test_sequence_generator.py#L68 and https://github.com/facebookresearch/fairseq/blob/acd9a53607d1e5c64604e88fc9601d0ee56fd6f1/tests/test_sequence_generator.py#L117-L120

[justinchuby]

Thanks. It could be that jit script doesn't like the deprecation logic. I'd be curious to hear suggestions on the deprecation path, as the underlying aten ops seem to be non-public.

[weiwangmeta]

Thanks @justinchuby. Let's involve more folks. cc @malfet @kit1980

[malfet]

I see there are two slightly different problem:

• torch.jit.script fails to recognize the deprecation decorator
• We should not deprecate public function in favor of private function. May be creating a public wrapper in torch. module is a way to go, but then it needs further discussion (i.e. why was it private in the first place?)