This link points to the current documentation.
And here is the updated rendered documentation. Inconsistent use of italics and code highlight is fixed (in favor of italics), inline math sections are replaced with italics as well.

💊 CI failures summary and remediations

As of commit 7fc28c5 (more details on the Dr. CI page):

• 1/1 failures possibly* introduced in this PR
  • 1/1 non-CircleCI failure(s)

XLA's torch.pinverse test fails. This test compares the output of normal PyTorch with XLA's PyTorch https://github.com/pytorch/xla/blob/master/test/cpp/test_aten_xla_tensor.cpp#L3341-L3353
XLA's SVD is implemented here https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/xla/client/lib/svd.cc#L808-L878
A possible reason for the failure could be that here I changed at::pinverse to use at::linalg_svd instead of at::svd.
A quick fix for this PR would be not to use linalg_svd for now until relevant bindings are implemented in pytorch/xla.

XLA's torch.pinverse test fails. This test compares the output of normal PyTorch with XLA's PyTorch https://github.com/pytorch/xla/blob/master/test/cpp/test_aten_xla_tensor.cpp#L3341-L3353
XLA's SVD is implemented here https://github.com/tensorflow/tensorflow/blob/master/tensorflow/compiler/xla/client/lib/svd.cc#L808-L878
A possible reason for the failure could be that here I changed at::pinverse to use at::linalg_svd instead of at::svd.
A quick fix for this PR would be not to use linalg_svd for now until relevant bindings are implemented in pytorch/xla.

cc @ailzhang @JackCaoG

Continuing to use at::svd to implement pinverse (for now) sounds good.

This PR resolves #45821.

The problem was that when introducing the support of complex inputs for torch.svd it was overlooked that LAPACK/MAGMA returns the conjugate transpose of V matrix, not just the transpose of V. So torch.svd was silently returning U, S, V.conj() instead of U, S, V.

Behavior of torch.linalg.pinv, torch.pinverse and torch.linalg.svd (they depend on torch.svd) is not changed in this PR.

Just so I'm confident I understand what's going on here:

• torch.svd has this weird thing where it returns V^T for real-valued inputs, not V
• the generalization of that for complex inputs is that it would return the conjugate transpose of V, V^H
• when complex support was added to torch.svd, it was mistakenly added so that V was returned, not V^H
• this PR address that inconsistency

Is that correct?

Follow-up question: how long has torch.svd been returning V and not V^H for complex inputs?

LAPACK and MAGMA return V.transpose() for real inputs and V.conj().transpose() for complex inputs.
Then in PyTorch output of LAPACK and MAGMA is transposed, so on current master branch torch.svd returns V.transpose().transpose() -> V for real inputs and V.conj().transpose().transpose() -> V.conj() for complex inputs.

torch.svd has this weird thing where it returns V^T for real-valued inputs, not V
the generalization of that for complex inputs is that it would return the conjugate transpose of V, V^H
when complex support was added to torch.svd, it was mistakenly added so that V was returned, not V^H

• No, torch.svd returns V for real-valued inputs
• The generalization to complex inputs is to return V as well. But it was mistakenly added so that V.conj() is returned, not V
• this PR address that inconsistency

torch.svd had been returning V.conj(), not V since the 3rd of October. Introduced in this PR #45795

LAPACK and MAGMA return V.transpose() for real inputs and V.conj().transpose() for complex inputs.
Then in PyTorch output of LAPACK and MAGMA is transposed, so on current master branch torch.svd returns V.transpose().transpose() -> V for real inputs and V.conj().transpose().transpose() -> V.conj() for complex inputs.

torch.svd has this weird thing wh

torch.svd has this weird thing where it returns V^T for real-valued inputs, not V
the generalization of that for complex inputs is that it would return the conjugate transpose of V, V^H
when complex support was added to torch.svd, it was mistakenly added so that V was returned, not V^H

• No, torch.svd returns V for real-valued inputs
• The generalization to complex inputs is to return V as well. But it was mistakenly added so that V.conj() is returned, not V
• this PR address that inconsistency

torch.svd had been returning V.conj(), not V since the 3rd of October. Introduced in this PR #45795

Got it, thank you for clarifying. This all makes sense now. Looks like this support is part of PyTorch 1.7.1, so this will be a BC-breaking change. Worse, it's a silent correctness issue. I expanded the PR summary with a BC-breaking note reflecting the current state of this PR (please correct me if I'm still missing something).

It would be safer if we could disallow complex inputs to torch.svd, even if internally we rely on at::svd with complex inputs. I can't think of any reasonable way to do that, however. Luckily torch.svd is being deprecated and this functionality was never documented, so it may be OK to change this.

Codecov Report

Merging #51012 (7fc28c5) into master (627a331) will increase coverage by 0.21%.
The diff coverage is 100.00%.

@@            Coverage Diff             @@
##           master   #51012      +/-   ##
==========================================
+ Coverage   80.70%   80.91%   +0.21%     
==========================================
  Files        1926     1926              
  Lines      210012   210020       +8     
==========================================
+ Hits       169485   169933     +448     
+ Misses      40527    40087     -440     

@albanD merged this pull request in ddf2681.