training_step and forward

[9p15p]

Hi,
In tutorial,
why do we say "We recommend using forward for inference/predictions and keeping training_step independent."

It seems that unifing the code of the network in forward is more compact and easy to modify.