How to use cache?

[liang23333]

How to use HSTU's cache? From the code, it seems that the cache is associated with delta_x_offsets, but I'm not sure what delta_x_offsets is supposed to look like.

[liang23333]

from hstu.py, the code
flattened_offsets = delta_x_offsets[1] + torch.arange(start=0, end=B * n, step=n, device=delta_x_offsets[1].device, dtype=delta_x_offsets[1].dtype),

This means that each row in the batch has only one element that needs to be updated? In other words, as long as each user has a new item-id, it will be sent to HSTU for an update?