The weights replace the original word embeddings permanently.

The weights are used to compute a weighted sum of the value vectors.

The weights are sorted alphabetically.

The weights are directly used as the final predicted token IDs.