Proposal: Add helper function to apply a SAE or SAEDict to a model about saelens HOT 2 CLOSED

#81 seems to implement the core functionality that would be needed here. I've defined an SAEPatcher class to create forward and backward hooks to inject a single SAE into the model and save its features and activations. The implementation matches this description:

always inject the SAE into model computation + error term by default, so the SAE gets gradients, tracks features, while not modifying model output

Possibly that could be re-written or extended to more closely match the usage you've defined here, e.g.:

• Defining a context manager method SAEPatcher.track_activations that returns an ActivationCache

from saelens.

HookedSAE Transformer provides this functionality and was merged as part of the 3.0 release.

from saelens.