grad doesn't work with _VF.frobenius_norm about functorch HOT 4 CLOSED

Tracked this issue down to https://github.com/pytorch/pytorch/blob/master/aten/src/ATen/native/LinearAlgebra.cpp#L2049

from functorch.

For this issue I think we should just rewrite all of the norm operations to not do what they're doing. It should not be standard for a CompositeImplicitAutograd operator to call into an out variant. In particular, frobenius_norm should not call frobenius_norm_out and instead they should have independent implementations. This gives the following benefits:

• It solves our problem
• frobenius_norm is faster due to not having to materialize an extra tensor and copying into it
• frobenius_norm_out is faster due to not having to materialize an extra tensor and copying into it
  There is a downside to maintainability: if someone makes a change to frobenius_norm then the same change need to be applied to frobenius_norm_out. But I'm not too worried about that.

from functorch.

It seems like this issue has been fixed? from testing on master.

from functorch.

I modified resize in PyTorch to not skip dispatch if it was operating on functorch wrappers as a short term solution a while back

from functorch.