Adding topic steering on layers

[benxh1995]

Prerequisites

Please answer the following questions for yourself before submitting an issue.

• [ x ] I am running the latest code. Development is very rapid so there are no tagged versions as of now.
• [ x ] I carefully followed the README.md.
• [ x ] I searched using keywords relevant to my issue to make sure that I am creating a new issue that is not already open (or closed).
• [ x ] I reviewed the Discussions, and have a new bug or useful enhancement to share.

Feature Description

Add the ability to set/apply steers on a layer by layer basis in order to ensure alignment with specific concepts, tokens or words.

Motivation

Based on recent papers and research on LLM steerability through applying steers on specific layers, I believe that llama.cpp would greatly benefit from incorporating a feature like this.

The colab in the section below should provide a very interesting application which results in far higher reasoning abilities, for example from a 3B.

Possible Implementation

Links to known implementations:
https://github.com/Mihaiii/llm_steer
https://colab.research.google.com/github/Mihaiii/llm_steer/blob/main/demo/llm_steer_demo.ipynb

And

https://github.com/nrimsky/LM-exp/tree/main/steering

[Engininja2]

Fyi there was draft PR experimenting with steering vectors #1472.

[peerschuett]

I took a look at #1472 , but the structure of llama.cpp changed in the meantime too much. In #1472 the steering was done in this function https://github.com/ggerganov/llama.cpp/blob/95dc4d7270e04bef3792a61be9bb045cb74a9fc4/llama.cpp#L1132 , but the whole function doesn't exist in this form any more.

[github-actions]

This issue is stale because it has been open for 30 days with no activity.

[github-actions]

This issue was closed because it has been inactive for 14 days since being marked as stale.