nnsight-remote-interpretability

You can analyze neural network internals on large PyTorch models without local GPUs by using remote execution via NDIF. This approach enables interpretability experiments on 70B+ parameter models directly from a standard laptop.

Yes, you can inspect and modify activations for any PyTorch architecture using nnsight. It provides universal support for transformers, Mamba, and custom models to help you understand model behavior through activation patching.

Mechanistic interpretability involves inspecting and manipulating internal model activations to understand behavior. Remote execution handles massive models by running your interpretability analysis code on external powerful infrastructure instead of local hardware.

This remote interpretability approach works with any PyTorch architecture, including standard transformers, Mamba, and custom models. It leverages the nnsight library to provide universal support across different model designs.

To run interpretability experiments on a 70B parameter model from your laptop, you use the provided Python scripts to execute your analysis code remotely on NDIF infrastructure, bypassing local hardware limitations.