Almost safe wrapper for BitNet-cpp based on llama-cpp
Add bitnet submodule:
cd bitnet-cpp-sys
git submodule add --name bitnet --depth 10 -- https://github.com/microsoft/BitNet.git bitnet
Pull recursive updates:
cd bitnet-cpp-sys/bitnet
git submodule update --init --recursive
Configure using python:
cd bitnet-cpp-sys/bitnet
pip install -r requirements.txt
python setup_env.py --hf-repo HF1BitLLM/Llama3-8B-1.58-100B-tokens -q i2_s # only needs gen_code()
- minimize crate size
- rewrite sampler PR 9294
- look into Metal implementation of MatMul kernerls for BitNet
- add more examples
- replace all remaining python from BitNet repo
- add better division between
arm64
andx86_64
Heavily inspired by llama-cpp-rs
© 2024, Eugene Hauptmann