Remove upsample_*.vec ops

[wonjoolee95]

Fixes https://github.com/pytorch/pytorch/pull/85638

---

The upstream PR makes upsample_*.vec ops CompositeImplicit, so the backward does not need an explicit implementation anymore.

[wonjoolee95]

The upstream PR's base branch is a bit outdated. Left a comment to rebase the PR.

[wonjoolee95]

PyTorch python op tests are failing:

======================================================================
ERROR: test_upsamplingNearest2d_xla (__main__.TestNNDeviceTypeXLA)
----------------------------------------------------------------------
Traceback (most recent call last):
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_device_type.py", line 391, in instantiated_test
    raise rte
  File "/opt/conda/lib/python3.7/site-packages/torch/testing/_internal/common_device_type.py", line 378, in instantiated_test
    result = test(self, **param_kwargs)
  File "/tmp/pytorch/xla/test/../../test/test_nn.py", line 14537, in test_upsamplingNearest2d
    helper(torch.contiguous_format, "nearest")
  File "/tmp/pytorch/xla/test/../../test/test_nn.py", line 14500, in helper
    out_t.backward(torch.randn_like(out_t))
  File "/opt/conda/lib/python3.7/site-packages/torch/_tensor.py", line 484, in backward
    self, gradient, retain_graph, create_graph, inputs=inputs
  File "/opt/conda/lib/python3.7/site-packages/torch/autograd/__init__.py", line 199, in backward
    allow_unreachable=True, accumulate_grad=True)  # Calls into the C++ engine to run the backward pass
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

I was able to see similar pt/xla cpp unit tests fail with the same error. I was in the assumption that the upstream PR makes the backward.vec op's explicit implementation not required. @bdhirsh, do you have any idea why this error is caused in PyTorch?

cc @JackCaoG

[JackCaoG]

@bdhirsh This message is too vaugue, do you spot anything that's obviously wrong here?

[lezcano]

At risk of stating the obvious, that message happens when you try to run backward on an element that does not have a requires_grad=True or comes from a computation with gradient. E.g.

>>> import torch
>>> torch.randn(1).backward()
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/usr/lib/python3.10/site-packages/torch/_tensor.py", line 396, in backward
    torch.autograd.backward(self, gradient, retain_graph, create_graph, inputs=inputs)
  File "/usr/lib/python3.10/site-packages/torch/autograd/__init__.py", line 173, in backward
    Variable._execution_engine.run_backward(  # Calls into the C++ engine to run the backward pass
RuntimeError: element 0 of tensors does not require grad and does not have a grad_fn

[lezcano]

Ok, so I think what's happening here is that you are removing the whole backward function, while in the other PR, they are removing only the backward function for the .vec overload.

[wonjoolee95]

Ok, after removing the upsample_nearest2d.vec and updating the existing GetOutputSizeWithScale function to accept scale_h and scale_w as such: https://github.com/pytorch/xla/blob/1d0c3393fb48cb8740379e4fea9c37a0e131a7dd/torch_xla/csrc/aten_xla_type.cpp#L165-L175 I can confirm the related cpp tests are passing: UpsampleNearest2D, UpsampleNearest2DWithScale, UpsampleNearest2DBackward, and UpsampleNearest2DBackwardWithScale.

The CI will fail for now because the pinned PyTorch PR is not rebased with latest master. Once the upstream rebases and updates the commit pin, I'll re-run the CI.

[wonjoolee95]

Upstream PR has merged, deleting torch_pin and merging this.