[torch.jit.trace] Inplace Index Put Silent Error

[YifanShenSZ]

🐞Describing the bug

This simple model can be converted but gives incorrect result

        class IndexPutModel(torch.nn.Module):
            def forward(self, x, position, val):
                y = x.clone()
                y[:, position] = val
                return y

The issue is in how we parse the TorchScript

graph(
    %x : Tensor(2, 2, "<class 'coremltools.converters.mil.mil.types.type_int.make_int.<locals>.int'>"),
    %position : Tensor(1, "<class 'coremltools.converters.mil.mil.types.type_int.make_int.<locals>.int'>"),
    %val : Tensor(1, "<class 'coremltools.converters.mil.mil.types.type_int.make_int.<locals>.int'>"),
):
  %4 = constant[]()
  %y = clone[](%x, %4)
  %6 = constant[value=0]()
  %7 = constant[value=0]()
  %8 = constant[value=9223372036854775807]()
  %9 = constant[value=1]()
  %10 = slice[](%y, %6, %7, %8, %9)
  %11 = listconstruct[]()
  %12 = view[](%val, %11)
  %13 = constant[]()
  %14 = listconstruct[](%13, %position)
  %15 = constant[value=False]()
  %16 = index_put_[](%10, %14, %12, %15)
return (%y)

That we did not realize %10 is a view of %y and %16 is a reference of %10, so we early terminate translation when we see the output %y has been created

main[CoreML5](%x: (2, 2, int32)(Tensor),
              %position: (1,int32)(Tensor),
              %val: (1,int32)(Tensor)) {
  block0() {
    %x_tmp: (2, 2, int32)(Tensor) = identity(x=%x, name="x_tmp")
    %position_tmp: (1,int32)(Tensor) = identity(x=%position, name="position_tmp")
    %val_tmp: (1,int32)(Tensor) = identity(x=%val, name="val_tmp")
  } -> (%x)
}

Potential Solution

We probably won't be able to fully fix this, since the solution is equivalent to torch's own functionalization 😂 But we do should find some way to error out, rather than silently produces a wrong Core ML model