Question about Attention Sharing

In your code, you seem to be only manipulating only parts of the keys and values

```py
            all_store_kv['key'].append(self.scale * key[:, :, 512:, :][:, :, dim_neg2_indices, :])
            all_store_kv['value'].append(value[:, :, 512:, :][:, :, dim_neg2_indices, :])
``` 

Is there a reason you do this? Why don't you use the all of the key or value?