patch atexit_serde_create_op_test

[rdspring1]

• Failed to run correctness tests because the scope member was not initialized correctly.

[rdspring1]

!build

[jjsjann123]

LGTM with minor nitpick

[rdspring1]

!build

[rdspring1]

@jjsjann123 I removed the per-op serialization check because it was making the CI take too long.

Currently, the python tests take about an hour. With the serialization check, it is over 1 hour, 45 minutes.

We'll just run the serialization check at exit for the opinfo tests in the CI.

[jjsjann123]

Currently, the python tests take about an hour. With the serialization check, it is over 1 hour, 45 minutes.

We'll just run the serialization check at exit for the opinfo tests in the CI.

naive question, are the two checks identical?
I thought the atexit stuff only checks for serialization. It doesn't attempt to deserialize and verify the test like serde_check_ops does.

If so, should we at least run serde_check_ops in our nightly/weekly runs to ensure coverage?

[rdspring1]

You're correct. atexit only does the serialization part in the CI. serde_check_ops does serialization, deserialization, and verify. We have the serde_check_ops in test_python_frontend, so we have daily coverage there.

The opinfo tests covers each op individually and handles more inputs, so it is slower. The opinfos are useful for development or debugging because every op is isolated.

If so, should we at least run serde_check_ops in our nightly/weekly runs to ensure coverage?

That is a good idea. If you have unlimited time it doesn't hurt to run them, but test_python_frontend provides good quality assurance.