CI is toasted

[NicolasHug]

From bf1914c on main branch:

$SRC_DIR/torchvision/csrc/ops/quantized/cpu/qnms_kernel.cpp:2:10: fatal error: ATen/native/quantized/affine_quantizer.h: No such file or directory
    2 | #include <ATen/native/quantized/affine_quantizer.h>
      |          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~

Because torch core renamed affine_quantizer.h into AffineQuantizer.h.

This will eventually be fixed by #6133. The irony is that some jobs are relying on the nightly from today while some other jobs are still relying on nightlies from yesterday where the changes from core haven't propagated. So these jobs are red on #6133.

I guess we'll have to wait till tomorrow to merge #6133 and get our green CI back.

cc @seemethere

[malfet]

We can add a script to create a symlink from one to another, if header is not available (Or #include one from another), but this is bad. Wonder if change on core were marked as BC breaking?

[seemethere]

Doesn't appear so, pytorch/pytorch#77037

cc @dzdang

[malfet]

To copy-paste a comment from #6133 : jobs that are broken depend on torchdata, which nightly build failed yesterday, so it forces torch nightly installation prior to that change.

[dzdang]

@malfet Sorry, just saw this. yeah, I think in retrospect, I should've marked the PR as bc-breaking. sorry about that (wasn't aware of existent projects that depended on the header file that wasn't part of the pytorch and fbcode codebase). do we need to revert the PR, or proceed with the solution above?

[malfet]

Also, looks like M1 CI is somehow affected by this as well:

2022-06-08T19:28:17.4266160Z /Users/ec2-user/runner/_work/vision/vision/torchvision/csrc/ops/quantized/cpu/qnms_kernel.cpp:2:10: fatal error: 'ATen/native/quantized/AffineQuantizer.h' file not found
2022-06-08T19:28:17.4267120Z #include <ATen/native/quantized/AffineQuantizer.h>
2022-06-08T19:28:17.4267400Z          ^~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~~
2022-06-08T19:28:17.4267580Z 1 error generated.

[malfet]

@malfet Sorry, just saw this. yeah, I think in retrospect, I should've marked the PR as bc-breaking. sorry about that (wasn't aware of existent projects that depended on the header file that wasn't part of the pytorch and fbcode codebase). do we need to revert the PR, or proceed with the solution above?

@dzdang I guess it's a question of whether we have any public facing tutorials/docs talking about AffineQuantizer, etc

[dzdang]

@malfet We don't. I took a look at qnms_kernel.cpp, and I don't think I saw any dependencies in it that would require #include that header file. maybe we can just remove the #include?

[vfdev-5]

maybe we can just remove the #include?

Removed locally the include and can confirm that torchvision compiles and test_ops.py::test_qnms are passing.

[NicolasHug]

#6141 merged, closing. Thanks everyone.