Sorry, I forgot to include in my comment this part:
If you include multiple version of llama, and each of those llama version depends on different GPU libraries, that could balloon the download size.
If these GPU libraries change rarely, then yes, you are correct, it might not be a problem.