How to pass big sparse tensors to a function to be used via dataset.map without hit the 2GB limit?

Question

I want to know what is the best way to have a constant SparseTensor along the same dataset.

I can think of two ways to do it: