[Feature Request] Support dtype arg in torch._foreach_norm

[awgu]

torch.linalg.vector_norm() and torch.norm support the dtype argument. Can we get support for it in torch._foreach_norm?

cc @crcrpar @mcarilli @janeyx99

[crcrpar]

it should be doable.
just to confirm, with dtype we specify the dtype of output tensors, right?

ref: https://pytorch.org/docs/stable/generated/torch.norm.html

[awgu]

I think semantically, it should equivalent to casting inputs to dtype first:

https://pytorch.org/docs/stable/generated/torch.linalg.vector_norm.html

I am mostly interested in this behavior. It is not a high priority though because in most cases, the tensors are already fp32 (e.g. gradients for gradient norm clipping).