xgboost multiclass classification problem

Now the xgboost training if the objective function is multi:softmax, it will training each class one tree, so in this case if i want use 50 trees to boost or parallel, in deed i use 50 * num_class trees, that may cost a lot in inference.
So my question is whether xgboost support real multi classification, use one tree to predict all the class's probility?