The attention matrix is computed based on all tokens in the context, so it kind of functions non-parametrically (but over the batch instead of over the whole tr

The attention matrix is computed based on all tokens in the context, so it kind of functions non-parametrically (but over the batch instead of over the whole training dataset)