[Q] Where is the rollout AttCAT method implemented?

[NightMachinery]

Where is AttCAT-aggregated-by-rollout implemented?

Looking through the code, I could only find the sum AttCAT method.

Mathematically speaking, AttCAT produces attributions of shape (batch, from_token) for each layer. So we cannot apply the rollout method on it at all, as rollout needs attributions with the shape (batch, to_token, from_token).

Here is the part of the paper that discusses these two methods: