aten::normal_ not handled as a special op in RemoveTensorMutation pass.

[neginraoof]

🐛 Bug

aten::normal_ should be considered as a special op in remove mutation pass.
The op does not have an equivalent aten::normal

To Reproduce

Steps to reproduce the behavior:

Apply remove mutation on the model below:

class Noise(nn.Module):
    def forward(self, image):
        return image.new_empty(2, 1, 3, 4).normal_()

Expected behavior

Error is:
RuntimeError: 0 INTERNAL ASSERT FAILED at "/pytorch/torch/csrc/jit/ir/alias_analysis.cpp":532, please report a bug to PyTorch. We don't have an op for aten::normal but it isn't a special case. Argument types: Tensor, float, float, None,

cc @ezyang @bhosmer @smessmer @ljk53 @bdhirsh @ailzhang @gmagogsfm

[SplitInfinity]

@eellison You wrote this pass, right? Do you have any insight to offer?

[ezyang]

The easiest fix is to probably just implement normal

[eellison]

Yes, but there is logic to validate that the replaced schema exists and there are operators for it, so i'm not sure what's going on: https://github.com/pytorch/pytorch/blob/master/torch/csrc/jit/passes/remove_mutation.h#L66. Edit: I guess normal is a weird case where the signatures of the inplace dont match up with the inplace variant..

[SplitInfinity]

@eellison Would it be okay if I assigned this to you for further investigation?

[neginraoof]

@eellison
There's a list of special ops like normal:

pytorch/torch/csrc/jit/passes/remove_mutation.h

Line 22 in 425a5dc

bool isSpecialMappedOp(Node* n) {

[eellison]

Yea, i'll add a fix.

[eellison]

really, we should add a normal_like constructor... im going to do a quick fix than open an issue for that.