Issues with Self-Extend on Phi-2, related to `n_embd_head`

[jshph]

Hi, I've noticed that in this PR: #4889 it replaces n_embd_head. Locally I see the following assertion fail: hparams.n_embd_head_k % ggml_blck_size(type_k) == 0 and I'm curious as to whether this is related to: #4657

I am trying to get self-extend with Phi-2 model to work locally via the llama-cpp-python wrapper. The script here: #4815 works, but I don't have success with the wrapper.

Thanks a lot!

[jshph]

Closing until confirming that it's not a llamacpp-python issue.