Why do `get_padding_and_inverse` in PTv3?

[yxchng]

Pointcept/pointcept/models/point_transformer_v3/point_transformer_v3m1_base.py

Line 182 in c759914

 pad, unpad, cu_seqlens = self.get_padding_and_inverse(point) 

I am learning PTv3 code and don't quite understand why do padding and inverse? Is it necessary?