NNsight and NDIF: Democratizing Access to Foundation Model Internals
Jaden Fiotto-Kaufman, Alexander R Loftus, Eric Todd, Jannik Brinkmann, Caden Juang, Koyena Pal, Can Rager, Aaron Mueller, Samuel Marks, Arnab Sen Sharma, Francesca Lucchetti, Michael Ripa, Adam Belfki, Nikhil Prakash, Sumeet Multani, Carla Brodley, Arjun Guha, Jonathan Bell, Byron Wallace, David Bau
2024-07-23
Summary
This paper introduces two tools, NNsight and NDIF, designed to make it easier for researchers to access and experiment with large language models (LLMs) without needing expensive hardware or complex setups.
What's the problem?
Many advanced AI models are very large and require a lot of computing power, making them difficult for most researchers to use. Customizing experiments with these models can be costly and complicated, which limits the ability of scientists to explore and improve these technologies. Additionally, many commercial models are not transparent, meaning researchers can't see how they work internally.
What's the solution?
To solve these problems, the authors developed NNsight, an open-source Python package that allows users to interact with any PyTorch model by building computation graphs. This makes it easier to modify and experiment with models. They also created NDIF, a platform that gives researchers access to large-scale LLMs through the NNsight API. This combination allows scientists to run experiments on powerful models hosted remotely without needing their own expensive hardware.
Why it matters?
This research is important because it democratizes access to advanced AI technologies, allowing more researchers to work with large language models. By providing tools that simplify the experimentation process and reduce costs, NNsight and NDIF can help accelerate innovation in AI research and make it more inclusive.
Abstract
The enormous scale of state-of-the-art foundation models has limited their accessibility to scientists, because customized experiments at large model sizes require costly hardware and complex engineering that is impractical for most researchers. To alleviate these problems, we introduce NNsight, an open-source Python package with a simple, flexible API that can express interventions on any PyTorch model by building computation graphs. We also introduce NDIF, a collaborative research platform providing researchers access to foundation-scale LLMs via the NNsight API. Code, documentation, and tutorials are available at https://www.nnsight.net.