From profiling my Vasprun initialisations, I found that one of the main contributions to runtime was repeated Incar.get() calls:

This can be slow when called many times because of the UserDict.get() implementation from collections.abc and formatting code. Because incar.get("ML_LMFF") was being used in the Vasprun parsing loop, this method was called e.g. 1e4 - 1e6+ times in my vasprun parsing. Changing this so that incar.get("ML_LMFF") is only called once to determine if the vasp run is an ML run, means incar.get only gets called ~20 times and reduces the vasprun parsing time by ~20-50% (depending on choices of parse_dos and parse_projected_eigen).