🔗 Helpful links

• 🧪  See artifacts and rendered test results at hud.pytorch.org/pr/63218
• 📄  Preview docs built from this PR
• 📄  Preview C++ docs built from this PR
• 🔧  Opt-in to CIFlow to control what jobs run on your PRs

💊 CI failures summary and remediations

As of commit 84998cc (more details on the Dr. CI page):

• 1/1 failures introduced in this PR

🕵️ 1 new failure recognized by patterns

The following CI failures do not appear to be due to upstream breakages:

pytorch_linux_xenial_cuda10_2_cudnn7_py3_gcc7_old_gradcheck_test1 (1/1)

Step: "Run tests" (full log | diagnosis details | 🔁 rerun)

Sep 21 23:58:23 Generating XML reports...
Sep 21 23:58:23 Generated XML report: test-reports/python-unittest/test_ops/TEST-TestCommonCUDA-20210921202126.xml
Sep 21 23:58:23 Generated XML report: test-reports/python-unittest/test_ops/TEST-TestGradientsCUDA-20210921202126.xml
Sep 21 23:58:23 Generated XML report: test-reports/python-unittest/test_ops/TEST-TestJitCUDA-20210921202126.xml
Sep 21 23:58:23 Generated XML report: test-reports/python-unittest/test_ops/TEST-TestMathBitsCUDA-20210921202126.xml
Sep 21 23:58:26 Traceback (most recent call last):
Sep 21 23:58:26   File "test/run_test.py", line 1027, in <module>
Sep 21 23:58:26     main()
Sep 21 23:58:26   File "test/run_test.py", line 1005, in main
Sep 21 23:58:26     raise RuntimeError(err_message)
Sep 21 23:58:26 RuntimeError: test_ops failed!
Sep 21 23:58:27 + cleanup
Sep 21 23:58:27 + retcode=1
Sep 21 23:58:27 + set +x
Sep 21 23:58:27 =================== sccache compilation log ===================
Sep 21 23:58:27 =========== If your build fails, please take a look at the log above for possible reasons ===========
Sep 21 23:58:27 Compile requests                      0
Sep 21 23:58:27 Compile requests executed             0
Sep 21 23:58:27 Cache hits                            0
Sep 21 23:58:27 Cache misses                          0
Sep 21 23:58:27 Cache timeouts                        0

There exists torch.batch_norm but it takes an extra arg: cudnn_enabled (not there in functional variant). This is passed from the functional variant to torch.batch_norm here: https://github.com/pytorch/pytorch/blob/master/torch/nn/functional.py#L2282. test_variant_consistency_jit fails with an error: (when passed an alias)

In that case, wouldn't torch.batch_norm be the main implementation and nn.functional.batch_norm be an alias for it? So OpInfo should be for torch.batch_norm and nn.functional.batch_norm should be in the alias list?

There exists torch.batch_norm but it takes an extra arg: cudnn_enabled (not there in functional variant). This is passed from the functional variant to torch.batch_norm here: https://github.com/pytorch/pytorch/blob/master/torch/nn/functional.py#L2282. test_variant_consistency_jit fails with an error: (when passed an alias)

In that case, wouldn't torch.batch_norm be the main implementation and nn.functional.batch_norm be an alias for it? So OpInfo should be for torch.batch_norm and nn.functional.batch_norm should be in the alias list?

I don't think that will also work since the functional variant doesn't accept an arg of cudnn_enabled.

ERROR: test_variant_consistency_eager_batch_norm_cpu_float32 (__main__.TestCommonCPU)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/home/krshrimali/Documents/Projects/Quansight/pytorch/torch/testing/_internal/common_device_type.py", line 373, in instantiated_test
    result = test(self, **param_kwargs)
  File "/home/krshrimali/Documents/Projects/Quansight/pytorch/torch/testing/_internal/common_device_type.py", line 780, in test_wrapper
    return test(*args, **kwargs)
  File "/home/krshrimali/Documents/Projects/Quansight/pytorch/test/test_ops.py", line 468, in test_variant_consistency_eager
    _test_consistency_helper(samples, variants)
  File "/home/krshrimali/Documents/Projects/Quansight/pytorch/test/test_ops.py", line 457, in _test_consistency_helper
    variant_forward = variant(cloned,
TypeError: batch_norm() got an unexpected keyword argument 'cudnn_enabled'

There exists torch.batch_norm

IMO torch.batch_norm should not exist and we just forgot to exclude it and now we can't because of a deprecation cycle... but I wouldn't worry about the additional "cudnn_enabled" argument

Have fixed the merge conflicts, @zou3519 - could you please take a look whenever you find the time? Thanks! :)

Have fixed the merge conflicts, @zou3519 - could you please take a look whenever you find the time? Thanks! :)

yes, will do! Sorry for the delayed response.

@zou3519 has imported this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

Test failures are unrelated; Attempting to merge

@zou3519 merged this pull request in 28bfdbb.