Skip to content

External activation capture runtime

The prototype uses ordinary forward hooks on Hugging Face transformers decoder blocks. This is the narrowest useful integration: it exposes per-layer, per-token residual tensors while keeping PyTorch and model weights out of MAC's base dependency set. src/mac/activation_probe/runtime.py therefore accepts any hookable model and imports neither package itself.

Compared with TransformerLens, this covers more current open-weight model architectures without conversion. Compared with nnsight or baukit, it adds no runtime server or tracing dependency. The cost is that layer paths remain model-specific and operators must validate them when registering a model.

Hooks copy only configured layers and are removed in a finally block. Memory and latency overhead are proportional to selected layers, sequence length, and hidden width; production pilots should capture one or two late residual layers rather than every block. Model licensing remains the license of the selected open-weight checkpoint. The mechanism is compatible with OpenShell because all model code and weights stay inside the worker sandbox.

These hooks work only when the caller owns the forward pass and can reference the model modules directly. They cannot be attached to a hosted-model API or to MAC's router. Instrumenting a local observer model would capture the observer's states only and must not be described as observing the hosted actor.