torch_addmm does not broadcast addition.

[ralmond]

I'm getting the following error when running torch_addmm()

he expanded size of the tensor (1) must match the existing size (2) at non-singleton dimension 1.  Target sizes: [4, 1].  Tensor sizes: [1, 2]
Exception raised from inferExpandGeometryImpl at ../aten/src/ATen/ExpandUtils.cpp:95 (most recent call first):

This seems to be happening at the add step. Shouldn't the addition of a [1,2] and a [4,1] tensor broadcast to a [4,2] tensor?

Here is a short script that reveals the problem.

pt <- torch_tensor(matrix(c(0,0,1,0,01,11,4,2,byrow=TRUE)))
at <- torch_tensor(matrix(1:2,2,1))
bt <- torch_tensor(matrix(c(0,-10),1,2))
torch_addmm(bt,pt,at)

This produces an error, but

torch_add(bt,torch_matmul(pt,at))

Produces the expected [4,2] tensor.