Comment by ifz

It's really simple, basically just the magnitude of the value vector, weighted by QK dot product, summed across all attention heads and layers.

When I started, I expected I'd have to experiment a lot to find something comprehensible. But this simple computation can already show some patterns.