[ONNX] Fix graph position to insert clone node for inplace op removal

[BowenBao]

Previous insertBefore approach might end-up inserting clone node in inner sub-blocks, while then the node being used later at other outside call sites.

[facebook-github-bot]

💊 CI failures summary and remediations

As of commit 8aeda1f (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_backward_compatibility_check_test (1/1)

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

Jan 30 02:53:55 The PR is introducing backward incompatible changes to the operator library. Please contact PyTorch team to confirm whether this change is wanted or not.

Jan 30 02:53:55 processing existing schema:  __getstate__(__torch__.torch.classes.xnnpack.LinearOpContext _0) -> ((Tensor, Tensor?, Scalar?, Scalar?) _0)
Jan 30 02:53:55 processing existing schema:  __setstate__(__torch__.torch.classes.xnnpack.LinearOpContext _0, (Tensor, Tensor?, Scalar?, Scalar?) _1) -> (None _0)
Jan 30 02:53:55 processing existing schema:  __getstate__(__torch__.torch.classes.xnnpack.Conv2dOpContext _0) -> ((Tensor, Tensor?, int[], int[], int[], int, Scalar?, Scalar?) _0)
Jan 30 02:53:55 processing existing schema:  __setstate__(__torch__.torch.classes.xnnpack.Conv2dOpContext _0, (Tensor, Tensor?, int[], int[], int[], int, Scalar?, Scalar?) _1) -> (None _0)
Jan 30 02:53:55 processing existing schema:  __getstate__(__torch__.torch.classes.xnnpack.TransposeConv2dOpContext _0) -> ((Tensor, Tensor?, int[], int[], int[], int[], int, Scalar?, Scalar?) _0)
Jan 30 02:53:55 processing existing schema:  __setstate__(__torch__.torch.classes.xnnpack.TransposeConv2dOpContext _0, (Tensor, Tensor?, int[], int[], int[], int[], int, Scalar?, Scalar?) _1) -> (None _0)
Jan 30 02:53:55 processing existing schema:  __init__(__torch__.torch.classes._nnapi.Compilation _0) -> (None _0)
Jan 30 02:53:55 processing existing schema:  init(__torch__.torch.classes._nnapi.Compilation _0, Tensor _1, Tensor[] _2) -> (None _0)
Jan 30 02:53:55 processing existing schema:  run(__torch__.torch.classes._nnapi.Compilation _0, Tensor[] _1, Tensor[] _2) -> (None _0)
Jan 30 02:53:55 processing existing schema:  __init__(__torch__.torch.classes.dist_rpc.WorkerInfo _0, str _1, int _2) -> (None _0)
Jan 30 02:53:55 The PR is introducing backward incompatible changes to the operator library. Please contact PyTorch team to confirm whether this change is wanted or not. 
Jan 30 02:53:55 
Jan 30 02:53:55 Broken ops: [
Jan 30 02:53:55 	aten::fake_quantize_per_channel_affine_cachemask_backward(Tensor grad, Tensor mask) -> (Tensor)
Jan 30 02:53:55 	aten::fake_quantize_per_channel_affine_cachemask(Tensor self, Tensor scale, Tensor zero_point, int axis, int quant_min, int quant_max) -> (Tensor output, Tensor mask)
Jan 30 02:53:55 ]
Jan 30 02:53:55 + cleanup
Jan 30 02:53:55 + retcode=1
Jan 30 02:53:55 + set +x
Jan 30 02:53:55 =================== sccache compilation log ===================
Jan 30 02:53:55 =========== If your build fails, please take a look at the log above for possible reasons ===========

---

This comment was automatically generated by Dr. CI (expand for details).

Follow this link to opt-out of these comments for your Pull Requests.

Please report bugs/suggestions to the (internal) Dr. CI Users group.

[neginraoof]

Thanks for fixing this.
Would you please show the graph for the test you added in comments?
Is this testing insertion into subblocks?

[BowenBao]

Thanks for fixing this.
Would you please show the graph for the test you added in comments?
Is this testing insertion into subblocks?

Showing before and after graph comparison.

# before
graph(%x.1 : Float(3, 16, strides=[16, 1], requires_grad=0, device=cpu),
      %fc.weight : Float(16, 16, strides=[16, 1], requires_grad=0, device=cpu),
      %fc.bias.1 : Float(16, strides=[1], requires_grad=0, device=cpu)):
  %3 : int = prim::Constant[value=2]() # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:22
  %4 : int = prim::Constant[value=1]()
  %5 : int = aten::dim(%x.1) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %6 : bool = aten::eq(%5, %3) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %out.1 : Tensor = prim::If(%6) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:4
    block0():
      %8 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:39
      %ret.2 : Tensor = aten::addmm(%fc.bias.1, %x.1, %8, %4, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:14
      -> (%ret.2)
    block1():
      %10 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:30
      %22 : None = prim::Constant()
      ### <----- %fc.bias clone inserted here
      %fc.bias : Float(16, strides=[1], requires_grad=0, device=cpu) = aten::clone(%fc.bias.1, %22)
      %output.2 : Tensor = aten::matmul(%x.1, %10) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:17
      %23 : Tensor = aten::add(%output.2, %fc.bias, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1669:12
      -> (%23)
  %13 : int = aten::dim(%out.1) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %14 : bool = aten::eq(%13, %3) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %out.3 : Tensor = prim::If(%14) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:4
    block0():
      %16 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:39
      ### <----- %fc.bias clone used in another block
      %ret.1 : Tensor = aten::addmm(%fc.bias, %out.1, %16, %4, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:14
      -> (%ret.1)
    block1():
      %18 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:30
      %output.1 : Tensor = aten::matmul(%out.1, %18) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:17
      %24 : Tensor = aten::add(%output.1, %fc.bias, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1669:12
      -> (%24)
  return (%out.3)

# after
graph(%x.1 : Float(3, 16, strides=[16, 1], requires_grad=0, device=cpu),
      %fc.weight : Float(16, 16, strides=[16, 1], requires_grad=0, device=cpu),
      %fc.bias.1 : Float(16, strides=[1], requires_grad=0, device=cpu)):
  %22 : None = prim::Constant()
  ### <----- %fc.bias clone inserted at the earliest after %fc.bias.1
  %fc.bias : Float(16, strides=[1], requires_grad=0, device=cpu) = aten::clone(%fc.bias.1, %22)
  %3 : int = prim::Constant[value=2]() # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:22
  %4 : int = prim::Constant[value=1]()
  %5 : int = aten::dim(%x.1) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %6 : bool = aten::eq(%5, %3) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %out.1 : Tensor = prim::If(%6) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:4
    block0():
      %8 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:39
      %ret.2 : Tensor = aten::addmm(%fc.bias.1, %x.1, %8, %4, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:14
      -> (%ret.2)
    block1():
      %10 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:30
      %output.2 : Tensor = aten::matmul(%x.1, %10) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:17
      %23 : Tensor = aten::add(%output.2, %fc.bias, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1669:12
      -> (%23)
  %13 : int = aten::dim(%out.1) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %14 : bool = aten::eq(%13, %3) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:7
  %out.3 : Tensor = prim::If(%14) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1663:4
    block0():
      %16 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:39
      ### <----- %fc.bias clone used here
      %ret.1 : Tensor = aten::addmm(%fc.bias, %out.1, %16, %4, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1665:14
      -> (%ret.1)
    block1():
      %18 : Tensor = aten::t(%fc.weight) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:30
      %output.1 : Tensor = aten::matmul(%out.1, %18) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1667:17
      %24 : Tensor = aten::add(%output.1, %fc.bias, %4) # /home/bowbao/repos/pytorch_test/torch/nn/functional.py:1669:12
      -> (%24)
  return (%out.3)