Large language models (LLMs) made easy, EasyLM is a one stop solution for pre-training, finetuning, evaluating and serving LLMs in JAX/Flax. EasyLM can scale up LLM training to hundreds of TPU/GPU accelerators by leveraging JAX's pjit functionality.
Building on top of Hugginface's transformers and datasets, this repo provides an easy to use and easy to customize codebase for training large language models without the complexity in many other frameworks.
EasyLM is built with JAX/Flax. By leveraging JAX's pjit utility, EasyLM is able to train large models that don't fit on a single accelerator by sharding the model weights and training data across multiple accelerators. Currently, EasyLM supports multiple TPU/GPU training in a single host as well as multi-host training on Google Cloud TPU Pods.
Currently, the following models are supported:
We are running an unofficial Discord community (unaffiliated with Google) for discussion related to training LLMs in JAX. Follow this link to join the Discord server. We have dedicated channels for several JAX based LLM frameworks, include EasyLM, JaxSeq, Alpa and Levanter.
OpenLLaMA is our permissively licensed reproduction of LLaMA which can be used for commercial purposes. Check out the project main page here. The OpenLLaMA can serve as drop in replacement for the LLaMA weights in EasyLM. Please refer to the LLaMA documentation for more details.
Koala is our new chatbot fine-tuned on top of LLaMA. If you are interested in our Koala chatbot, you can check out the blogpost and documentation for running it locally.
The installation method differs between GPU hosts and Cloud TPU hosts. The first step is to pull from GitHub.
git clone https://github.com/young-geng/EasyLM.git
cd EasyLM
export PYTHONPATH="${PWD}:$PYTHONPATH"The GPU environment can be installed via Anaconda.
conda env create -f scripts/gpu_environment.yml
conda activate EasyLMThe TPU host VM comes with Python and PIP pre-installed. Simply run the following script to set up the TPU host.
./scripts/tpu_vm_setup.shThe EasyLM documentations can be found in the docs directory.
If you found EasyLM useful in your research or applications, please cite using the following BibTeX:
@software{geng2023easylm,
author = {Geng, Xinyang},
title = {EasyLM: A Simple And Scalable Training Framework for Large Language Models},
month = March,
year = 2023,
url = {https://github.com/young-geng/EasyLM}
}
- The LLaMA implementation is from JAX_llama
- The JAX/Flax GPT-J and RoBERTa implementation are from transformers
- Most of the JAX utilities are from mlxu
- The codebase is heavily inspired by JAXSeq