When writing a wrapper tensor you typically want the wrapper to "look like" the wrapped tensor as much as possible. Some things are obvious and directly accepted like _make_wrapper_subclass such as size, dtype, device, layout, requires grad. But there are some things that are not so obvious:

• What storage() returns
• Whether or not the tensor reports as a view()
• What the strides are
• What the storage offset is

It's not so easy to get these to report correctly, since you have to create a whole "phantom" universe of storages at your wrapper level to make everything look right. @bdhirsh had to do some of this in his functionalization pass. Need to make it easier for people in general to do.

cc @albanD