What problem does it solve? Running mechanistic interpretability experiments on massive models (70B+) normally requires expensive local GPU clusters, and different interpretability libraries often force you to reimplement models or work only with specific architectures. This Skill provides guidance for using nnsight to access and modify the internals of any PyTorch model, with the same code running locally on small models or remotely on huge models through NDIF. ## Core Features & Use Cases - Activation Analysis and Patching: Trace model execution, save hidden states and attention patterns from any layer, and patch activations between clean and corrupted prompts to test causal relationships. - Remote Execution on Massive Models: Toggle remote=True to run identical interpretability code on Llama-3.1-70B/405B and other models hosted on NDIF without local GPUs. - Cross-Prompt Interventions and Steering: Share activations between prompts in a single trace, add steering vectors, and perform logit lens analysis across layers. - Use Case: A researcher wants to find which layer of Llama-3.1-70B encodes factual recall. They use this Skill to run an activation patching sweep remotely via NDIF, identifying the most causally important layer without owning any GPU hardware. ## Quick Start Ask the AI to help you write an nnsight script that traces a prompt through GPT-2, saves the layer 8 activations, and patches them into a corrupted prompt to compare predictions.
