Hey,
pretty sure a bug was introduced in this commit -
607d9b8
when passing uninitialized alpha torch.full throws an exception, becase we are trying to create a tensor from an uninitialized value.
I am not sure is this was intended or not, but you removed the option to pass uninitialized a.
Now if a is passed uninitialized torch.full returns an error.
Originally posted by @ofirDubi in 607d9b8