Does this skill support gradient-based analysis?

Yes, nnsight supports gradient access during backward passes for local execution, though this is not currently supported for remote NDIF or vLLM execution.

nnsight is a specialized library for interpreting and manipulating PyTorch model internals, supporting both local execution for small models and remote execution for massive models via NDIF.

Can I run experiments on Llama-3.1 405B using this skill?

Yes, by setting the remote parameter to true in your trace context, nnsight allows you to run interpretability tasks on massive models hosted on the NDIF infrastructure.

Do I need a local GPU to use this Claude Code skill?

No, if you utilize the remote NDIF execution feature, you can perform deep model analysis using only an API key without needing local GPU resources.

How does nnsight differ from TransformerLens?

While TransformerLens offers a consistent API for specific transformer models locally, nnsight supports any PyTorch architecture and offers remote execution capabilities for models that don't fit on local GPUs.

nnsight Remote Interpretability

Name: nnsight Remote Interpretability
Author: Orchestra-Research

byOrchestra-Research

•

3,983

•

Data Science & ML

Interprets and manipulates neural network internals across local and remote models using the nnsight library and NDIF execution.

nnsight provides a unified interface for mechanistic interpretability, allowing researchers to trace, analyze, and intervene in the internal computations of any PyTorch model. Its standout feature is the seamless integration with the National Deep Inference Fabric (NDIF), which enables the execution of complex interpretability experiments on massive foundation models like Llama-3.1-405B without requiring local high-end GPU clusters. By abstracting model internals into a deferred execution graph, nnsight simplifies advanced research tasks such as activation patching, gradient-based analysis, and cross-prompt activation sharing across diverse architectures.

Key Features

01Unified tracing API compatible with any PyTorch architecture including Transformers and Mamba

02Real-time activation manipulation and patching for causal analysis

033,983 GitHub stars

04Deferred execution model for efficient processing and reduced memory overhead

05Remote execution on massive models (up to 405B) via NDIF integration

06Cross-prompt activation sharing within a single tracing context

Use Cases

01Performing mechanistic interpretability on foundation models too large for local hardware

02Developing and testing custom interventions during the model generation process

03Causal mediation analysis through activation patching to locate specific model knowledge

What are Skills?·How to Install

Install with 🐟 Skill.Fish

npx skillfish add orchestra-research/ai-research-skills nnsight

For use in Claude.ai and ChatGPT

Download Skill