Optimal Weight Formats

Code to reproduce the experiments of Optimal Formats for Weight Quantisation.

We recommend starting with Demo.ipynb for an introduction to the main results of the paper.

Index:

• weight_formats - module containing core implementation for quantisation and Fisher estimation
• weight_formats.experiments - module containing runners for various types of experiment (token_prediction (main), fisher, weight_stats)
• scripts - generic scripts for launching experiments and testing
• paper - scripts for reproducing experiments and notebooks for analysing results
• Demo.ipynb - demo of main results
• Usage.ipynb - tutorial for using the package directly

Development

python3 -m venv .venv
echo 'export PYTHONPATH="${PYTHONPATH}:$(dirname ${VIRTUAL_ENV})"' >> .venv/bin/activate
source .venv/bin/activate
pip install -r requirements.txt
./scripts/check.sh

License

Copyright (c) 2025 Graphcore Ltd. Licensed under the MIT License.