whittle 0.3.0

Two-stage neural architecture search for large language models

whittle

Whittle is a library for Neural Architecture Search (NAS) aimed at compressing large language models (LLMs). The core idea is to treat the network as a super-network and select sub-networks that optimally balance performance and efficiency. Whittle provides the following functionalities:

• Support for a variety of LLMs to define super-networks
• Checkpoint generation for sub-networks, enabling deployment
• Downstream evaluation of sub-networks using LM-Eval-Harness
• A simple interface for various super-network training strategies and multi-objective search methods

Installation

Whittle supports and is tested for python 3.9 to 3.11.

You can install whittle with:

pip install whittle

Install from source

Install whittle from source to get the most recent version:

git clone [e-mail]:whittle-org/whittle.git
cd whittle
pip install -e .

Getting started with whittle

To explore and understand different functionalities of whittle checkout this colab notebook and examples/

Projects that use whittle

• Structural Pruning of Pre-trained Language Models via Neural Architecture Search
• HW-GPT Bench

How to get involved

We more than happy for any code contribution. If you are interested in contribution to whittle, please read our contribution guide.