🔗 Helpful links

• 🧪  See artifacts and rendered test results at hud.pytorch.org/pr/73637
• 📄  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 7cc8ed3 (more details on the Dr. CI page):

💚 💚 Looks good so far! There are no failures yet. 💚 💚

Fix ##73639

The issue is described in ##73639. Thanks.

The purpose is to enable decomposition pass only for GPU device, here two reasons listed ,

1. On our device backend, we hope to have an operator level optimization for BatchNorm and LayerNorm (mkldnn support) on JIT graph. So we don't want to decompose these two operators.
2. According to the current implementation and my understanding, decomposition pass only benefits GPU device, where we can get elementwise kernels fusion by TensorExpr lowering. But the condition is to enable decomposition for all non-CPU devices,

 44   auto device = input->type()->expectRef<TensorType>().device();
 45   // As of now, we do the decomposition for batchnorm/layernorm on GPU device
 46   // only
 47   if (!device || (*device).is_cpu()) {
 48     return false;
 49   }

@davidberard98 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Hi, @eellison , any comments?

Hey @arthuryuan1987.
You've committed this PR, but it does not have both a 'release notes: ...' and 'topics: ...' label. Please add one of each to the PR. The 'release notes: ...' label should represent the part of PyTorch that this PR changes (fx, autograd, distributed, etc) and the 'topics: ...' label should represent the kind of PR it is (not user facing, new feature, bug fix, perf improvement, etc). The list of valid labels can be found here for the 'release notes: ...' and here for the 'topics: ...'.
For changes that are 'topic: not user facing' there is no need for a release notes label.

@davidberard98 @eellison Can you help to add a release notes and topics for the PR? release notes: jit

Hi, @davidberard98 @eellison Could you please add a release_notes and topics to the PR? and reopen the PR?

@arthuryuan1987 added labels. Why did you want to reopen the PR?

@davidberard98 has imported this pull request. If you are a Meta employee, you can view this diff on Phabricator.

Hi, @davidberard98, I am not familiar with the PR process of PyTorch. The PR has not been merged. Does it mean the commit will be merged by you?

Ah, the way it works (or at least how it used to work, the process changed a little bit recently) is 1) import PR to internal codebase, 2) merge internally, and 3) a bot rebases the commit onto master.

You can see that the changes got merged here:
https://github.com/pytorch/pytorch/blob/master/torch/csrc/jit/passes/decompose_ops.cpp#L47

Ah, the way it works (or at least how it used to work, the process changed a little bit recently) is 1) import PR to internal codebase, 2) merge internally, and 3) a bot rebases the commit onto master.

You can see that the changes got merged here: https://github.com/pytorch/pytorch/blob/master/torch/csrc/jit/passes/decompose_ops.cpp#L47

Thanks for your patient explanation.