Why do we require a training set in KNN? On top of that, why is k-fold cross-validation popular in KNN?

Question

Since there is no requirement of learning any weights, and the algorithm is standard.