You could pad them all to be the same length, but there's no real reason to run a sorting algorithm to determine that a length 5 number in binary is larger than a length 2 one. You would likely get better performance by grouping the numbers by length and running your radix sort within each group. Of course, that's dependent upon how you group them and then on how you sort your groups.
An example of how you might do this would be to run through all the items once and throw them all into a hash table (length --> numbers of that length). This takes linear time, and then let's say nlogn time to access them in order. A radix sort runs in O(nk) time where n is the number of items and k is their average length. If you've got a large k, then the difference between O(nk) and O(nlogn) would be acceptable.