Why does the queries are normalized while keys not when using multihead_attention?

[feynmanma7]

 # Self-attention
self.seq = multihead_attention(queries=normalize(self.seq),
                               keys=self.seq,
                               num_units=args.hidden_units,
                               num_heads=args.num_heads,
                               dropout_rate=args.dropout_rate,
                               is_training=self.is_training,
                               causality=True,
                               scope="self_attention")

https://github.com/kang205/SASRec/blob/e3738967fddab206d6eeb4fda433e7a7034dd8b1/model.py#L54

Thank you!