Why does stepwiseglm give me many chosen predictors with p value close to 1

Hi, I ran a stepwiseglm with around 32 predictors and used stepwiseglm with poisson as link function. After nearly 9 hours, it came up with a chosen model which includes several categorical variables whose p values are almost 1 at each level (of the categories). I used deviance as criterion and according to Algorithms in stepwiseglm, such predictors should be moved from the model. And it could have come up with a model with only predictors whose p- values are below 0.05 or 0.1. It does not have to use to those predictors whose p values are almost 1. How come ...?