For classification problems, should model always have logits output for loss calculation?

[Ahnaf19]

I am a bit confused: previous of starting this course, I saw models with softmax, sigmoid or torch.round() at the end of the layer for classification tasks, which is understandable to me. But, here (I am at the end of 03. PyTorch Computer Vision) I understood about logits. It seems like we calculate logits --> prob_preds --> pred_labels.

Should my model always have logits output for loss calculation and apply softmax/sigmoid separately for getting pred_labels?