It depends on the type of virus, and what you mean by "evolve separately" etc.

Some viruses recombine very often, so that the evolution of one gene (or even gene regions within each gene) can evolve very independently of others.  Other viruses almost never recombine.  Even within a single virus group, such as the HIV-1 M group of viruses, different lineages experience different levels of recombination due to factors such as the density of infections in a given location (high density, higher chance for dual infection leading to recombination).

Nearly all phylogenetic "models of evolution" really come down to models of DNA base substitution or amino acid substitution, and not models of selection pressures and other influences on evolution.  However, even the most simplistic models are highly useful for many purposes, and more complex models provide a gain in accuracy that is quite often of some use.

So, the answer really is that you don't have to guess whether you should use different models and partions etc, you can do tests on your virus of interest to see if there is recombination, differential evolution, etc.   In some cases, simply "eyeballing" the multiple sequence alignment of a set of carefully chosen genomes can give you some answers.  But tools like SimPlot and the NCBI Genotyping tool can help a lot, and they are pretty easy to learn to use.

https://www.ncbi.nlm.nih.gov/projects/genotyping/formpagex.cgi

http://sray.med.som.jhmi.edu/SCRoftware/simplot/

You may find that the different proteins of your virus are best fit by different models. This can be related to their different structural properties, despite being code by a single gene. The same may apply even to different regions of the same protein. Besides, each protein may have different evolutionary constraints and be best modelled by different gamma distributions of rate variation across sites. You can try to model each protein separately or try a quick approach considering all the proteins to be evolving under the same model. I would not expect large differences between the two, although the first approach is probably more elegant.

Thank you both! I used PartitionFinder2, split in 5 partitions (5UTR-NS_Polyprotein-Promoter-Structural_Polyprotein-3UTR) for an alignment with 340 genomes (same virus). Unfortunately, RefSeq sequence for this particular genome just shows the structure as 5 partitions and some of the genomes (about 50) that have the proteins within each polyprotein annotated, have variation regarding to the actual real size and boundaries of each protein. Based on that, I decided to go with 5 partitions instead of more and assume the boundaries based on the RefSeq. Better safe than sorry? Asking here, critics are more than welcome, better here than when paper is under peer-review.