Out-variant kernels are lacking BC protection for new default args being added.

🐛 Describe the bug

In general adding a new argument with a default value to an operator in PyTorch is not considered a BC breaking change.

Backwards Compatible (BC) here meaning new runtime old model.

Today though since out tensors are at the end of the list of args we lose the nice default value behavior of c++ headers, and adding new default args is a BC breaking change because of this.

The solution is to generate wrappers in code gen that place out at the front so that we can still get the default arg behavior. If we inline these functions there should be no regression either.

Today:

boxed_kernel(Evalue* args) {
return out_at_back(/* scale???? */, args[0].toTensor());
}

Solution:

boxed(Evalue* args) {
return out_at_front(args[0].toTensor());
}

inline Tensor out_at_front(Tensor out, int scale = 1) {
return out_at_back(scale, out);
}

Versions

.

cc @larryliu0820 @lucylq