how to compute the cosine similarity between task vectors?

[cht619]

I have a question regarding the computation of cosine similarity between task vectors. Should the cosine similarity be computed by flattening all parameters into a single vector (high memory cost), or should it be calculated separately for different parts of the model (e.g., MLP, CNN, RNN...) and then aggregated by a weighted average (where the weight are based on the dimensionality of each layer)?
Thank you for your time and assistance.

[DiSheng555]

I also encountered the same problem. By calculating different parts of the model separately and then summarizing them through weighted averages, the similarity obtained is very high, which contradicts the content of the article. I don't know why