For a 100 MHz 5G carrier on a GaN Doherty, I would start much simpler than most people expect and add complexity only where the ACLR or EVM actually fails. A memory polynomial or GMP with moderate nonlinearity order — typically 7th to 9th order — and a short memory depth of 3 to 5 taps is usually enough to get within a few dB of the target, provided the training signal matches the real waveform in PAPR and bandwidth. Training on a CW tone or a narrowband signal and then switching to a 100 MHz carrier is the classic trap: the ACLR looks good in training and falls apart on the real signal.
The Doherty-specific problem is that the main and peaking devices turn on at different drive levels, so one global polynomial has to fit two different nonlinear shapes. I have had better results with a segmented or dual-input model that treats the low-power and high-power regions separately, rather than pushing a single model to very high order. Extra order mostly fits measurement noise and the observation receiver's own distortion, and it makes the coefficients unstable when the signal changes.
Memory depth should follow the envelope bandwidth and the thermal time constant, as LunaWei noted. Fast memory from bias and matching covers a few samples; slow thermal drift needs a separate slow adaptation or a temperature table, not deeper memory. My acceptance check is always the same: report carrier configuration, PAPR, model order and memory depth, iteration count, and ACLR, EVM and power consumption together. If adding memory buys less than 1–2 dB of ACLR, I stop there — the extra coefficients are not worth the risk on the bench.