CFA is a way to establish the dimensionalities of your data. For instance, unidimensionality and local independence are the assumptions of (unidimensional) IRT.
Just learned it today...
The default ML estimation does not provide model fit statistics for MPlus if you are running Categorical CFA. However if we change the method of estimation to wlsmv, it will spit out CFI, TLI and RMSEA in the output.
ANALYSIS: estimator=wlsmv;
Following are the links: