same here last year. The modular exponentiation block alone pushed the circuit depth past 1000 gates even for factoring 15, and the readout gave random junk until I mapped the control qubits onto the lowest-error subset on the ibm backend.
what backend and coupling map did you end up with for your runs?