Notice that there is no activation function on the output layer for our forward propagation logic here. So the output is “logits”, meaning linear outputs, not activation outputs. That is why you need “True” as the value of from_logits to tell the cost function it is getting logits and needs to internally apply the activation function.
The other thread that I linked earlier also talks about this in some detail and explains the mathematical reasons why it is better to do it this way.