When to use VIF to detect multicollinearity for the variables in a GLM and how to handle transformed variables?

Question

This is my first project creating my own models. I have 12 possible variables for a habitat model. I am using glms (binominal, logit). I want to check for multicollinearity using the VIF. I have variables on which I will use a log transformation, some that will need quadratic terms and some that will be used with an interaction term with the sex of the animal. I will select the best combination and transformation of variables for a summer season and a winter season model separate by making candidate models for my hypotheses.

Now I wonder what's the smartest/standard way to use the VIF in the process:

Is it a preliminary analysis where I just put all my variables in and kick out the ones with a value over my thresholds (VIF:3, Tolerance:0,2) until all values are below these thresholds?

OR

Do I do it for the complete sets of variables for my 3 hypothesis groups and kick out the ones with a value over my thresholds until all values are below the thresholds?

OR

Do I do it after I found the best candidate models?

Furthermore I am not sure how to include transformations, interaction terms and quadratic terms ? My variables are standardized. Should I include these alterations of the variables or do I use the pure variables (If I do it as a preliminary analysis I probably don't know for sure which variable alterations I will be using in the end)?

Thanks for any help.