confusion about var_ema

in file model.py, _build_var_ema create and var_ema, and apply to all trainable variables, but it didn't use average to update the trainable variables. only in graph_handler.py, when load model from save would use var_ema。
is this a trick?
when training, create shadow variables for trainable variables, but dont update them, when predict, use this to initial the variables .
if it is, then why?