nnsight-remote-interpretability

What problem does it solve?

This Skill provides a unified interface for interpreting and manipulating the internal workings of any PyTorch neural network, including massive models that cannot be run locally.

Core Features & Use Cases

• Remote Execution: Run interpretability experiments on models up to 405B parameters without local GPU resources via NDIF.
• Universal PyTorch Support: Works with any PyTorch architecture (Transformers, Mamba, custom models).
• Mechanistic Interpretability: Access and modify activations, gradients, and attention patterns for deep model understanding.
• Use Case: Analyze the internal representations of a 70B parameter model to understand how it processes specific linguistic phenomena, without needing a supercomputer.

Quick Start

Use the nnsight skill to trace the model 'gpt2' with the prompt 'Hello world' and save the hidden states from layer 5.