Efficient way to sample per group with given number/proportions in R

I would like to know if there's an efficient way of sampling for groups, choosing an integer and/or proportion to sample from them. I am aware of the existence of sample_n and that it works with grouped dfs but as far as I know it samples the same number for each group.

A minimal description of the problem, on a simple case, would be to sample, from the dataframe mtg, 5 random rows (or vector of indexes of those rows) for cyl == 4, 7 for cyl == 6 and 3 for cyl == 8.

Thanks,