NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals

Gespeichert in:
Bibliographische Detailangaben
Veröffentlicht in:arXiv.org (Dec 8, 2024), p. n/a
1. Verfasser: Fiotto-Kaufman, Jaden
Weitere Verfasser: Loftus, Alexander R, Todd, Eric, Brinkmann, Jannik, Pal, Koyena, Troitskii, Dmitrii, Ripa, Michael, Belfki, Adam, Rager, Can, Juang, Caden, Mueller, Aaron, Marks, Samuel, Arnab Sen Sharma, Lucchetti, Francesca, Prakash, Nikhil, Brodley, Carla, Guha, Arjun, Bell, Jonathan, Wallace, Byron C, Bau, David
Veröffentlicht:
Cornell University Library, arXiv.org
Schlagworte:
Online-Zugang:Citation/Abstract
Full text outside of ProQuest
Tags: Tag hinzufügen
Keine Tags, Fügen Sie das erste Tag hinzu!

MARC

LEADER 00000nab a2200000uu 4500
001 3083764303
003 UK-CbPIL
022 |a 2331-8422 
035 |a 3083764303 
045 0 |b d20241208 
100 1 |a Fiotto-Kaufman, Jaden 
245 1 |a NNsight and NDIF: Democratizing Access to Open-Weight Foundation Model Internals 
260 |b Cornell University Library, arXiv.org  |c Dec 8, 2024 
513 |a Working Paper 
520 3 |a We introduce NNsight and NDIF, technologies that work in tandem to enable scientific study of very large neural networks. NNsight is an open-source system that extends PyTorch to introduce deferred remote execution. NDIF is a scalable inference service that executes NNsight requests, allowing users to share GPU resources and pretrained models. These technologies are enabled by the intervention graph, an architecture developed to decouple experiment design from model runtime. Together, this framework provides transparent and efficient access to the internals of deep neural networks such as very large language models (LLMs) without imposing the cost or complexity of hosting customized models individually. We conduct a quantitative survey of the machine learning literature that reveals a growing gap in the study of the internals of large-scale AI. We demonstrate the design and use of our framework to address this gap by enabling a range of research methods on huge models. Finally, we conduct benchmarks to compare performance with previous approaches. Code documentation, and materials are available at https://nnsight.net/. 
653 |a Application programming interface 
653 |a Python 
653 |a Source code 
653 |a Neural networks 
700 1 |a Loftus, Alexander R 
700 1 |a Todd, Eric 
700 1 |a Brinkmann, Jannik 
700 1 |a Pal, Koyena 
700 1 |a Troitskii, Dmitrii 
700 1 |a Ripa, Michael 
700 1 |a Belfki, Adam 
700 1 |a Rager, Can 
700 1 |a Juang, Caden 
700 1 |a Mueller, Aaron 
700 1 |a Marks, Samuel 
700 1 |a Arnab Sen Sharma 
700 1 |a Lucchetti, Francesca 
700 1 |a Prakash, Nikhil 
700 1 |a Brodley, Carla 
700 1 |a Guha, Arjun 
700 1 |a Bell, Jonathan 
700 1 |a Wallace, Byron C 
700 1 |a Bau, David 
773 0 |t arXiv.org  |g (Dec 8, 2024), p. n/a 
786 0 |d ProQuest  |t Engineering Database 
856 4 1 |3 Citation/Abstract  |u https://www.proquest.com/docview/3083764303/abstract/embedded/ZKJTFFSVAI7CB62C?source=fedsrch 
856 4 0 |3 Full text outside of ProQuest  |u http://arxiv.org/abs/2407.14561