Gradients are not preserved in torch

[frazane]

When using asarray() in the torch backend, if called on a pre-existing tensor it will delete its autograd history by default (see https://pytorch.org/docs/stable/generated/torch.asarray.html). Consequently, one cannot compute gradients on functions that include asarray().

We need to use as_tensor() instead, or use asarray(*args, requires_grad=True).