You tune the inner (slave or secondary) loop first, just as if it were a stand alone loop not part of cascade. Then with the inner loop tuned and running, tune the outer (master or primary) loop, just as if it were a standalone loop.
There should be a big difference in speed of the two loops, with the inner loop at least four time as fast as the outer loop. If this is not the case, reduce the reset rate (make it slower) of the outer loop.