Row-wise softmax normalizes each row so its entries sum to 1, turning raw similarity scores into attention weights.