llama-70b-chat-4-shards

This repository contains a script to transform the weights of Llama v2 70B (chat)/Llama v3 70B (chat) from an 8-shard configuration to a 4-shard configuration, making it more accessible for users with machines that have only 4 GPUs.

For convenience, you can directly download weights from https://huggingface.co/Jinawei/llama-v2-70b-chat-4-shards

Introduction

Meta recently released the weights for Llama v2 70B/Llama v3 70B, distributed across 8 shards. However, some users may have hardware constraints, such as having only 4 GPUs on their machine, making it difficult to load and utilize the model directly. This repository provides a solution to this problem by offering a script that can transform the 8-shard weight distribution of LLAMA v2 70B/Llama v3 70B or into a 4-shard configuration, facilitating easier usage on machines with fewer GPUs.

Usage

You can convert weights of Llama v2 using the command below:

python convert.py \
    --input_llama_path ~/llama-2-70b-chat \
    --input_shards 8 \
    --output_llama_path ~/llama-2-70b-chat-4-shards \
    --output_shards 4
    --llama_version 2

You can convert weights of Llama v3 using the command below:

python convert.py \
    --input_llama_path ~/Meta-Llama-3-70B-Instruct \
    --input_shards 8 \
    --output_llama_path ~/Meta-Llama-3-70B-Instruct-1-shards \
    --output_shards 4
    --llama_version 3

Star as Activation

If this script proves useful in your work or projects, please consider giving it a star on GitHub. Your support helps to make the project more visible, encourages future development, and is greatly appreciated

Acknowledgements

• Thanks to Meta for releasing the Llama v2 70B/ Llama v3 weights.
• Any other acknowledgements or credits you'd like to give.

Contact

For any inquiries or to report issues, please open an issue on this repository.