My implementation of Sparse Autoencoders for personal use. Since I work with networks that are not LLMs, and whose architectures can vary a lot, the code is as general as possible.
These implementations are by no means the most efficient, nor the most elegant code, and probably missing a lot of useful features (and SAE architectures) present in other packages. They could also contain errors / bugs.
Yet, if it is useful for you, the code is publicly available.
For feature-rich packages, or nice implementations, albeit tailored for LLMs, you may find useful to visit the following: