Large W/L ratio for M1 and M2 such that gm could be larger
Small W/L and long-channel L for M3, M4 and M5 (as current mirrors)
For good matching and therefore smaller offset voltage, choose Vgs - Vth > 0.3 V (or even greater number) for the bias of M3, M4 and M5.
Choice of L:
The open loop gain (relevant to offset voltage in close-loop) is determined by gm Rout
Choose L > 10 um for all transistors if cascode configuration is not possible.
With power supply = 3.3 V +/- 10%, this structure can accomodate (high-swing) casecode for pMOS and for nMOS.
Without using cascode structure, the circuit can work quite well in 1.8 V system.
The information (sizing guidelines) I provided is not valid for academic research only.
It has been widely used industry.