offload most used layers to gpu to get better performance

[PinkD]

I found that llama.cpp always offload the first n layers into gpu.
Is there any way to offload the most used layers to gpu to get better performance?
For example, we can run the model with cpu to count how many times each layers have been accessed and write the result into a file. Then we use this file to tell llama.cpp which layer should stay in gpu to get the theoretical best performance.

[PinkD]

I forgot that every layer in the model are equally traversed. So the count of every layer should be the same.