Finetuning Large Language Models Efficiently on a Distributed Cluster

This repository is a boilerplate/fingerprint to fine tune any HuggingFace Large Language Model, such as FALCON-7B, using a distributed cluster. The purpose of this repo is to make it straightforward to fine tune any model efficiently by leveraging multi-GPU training. It uses Ray AIR to orchestrate the cluster on AWS, and DeepSpeed for parameter+optimizer sharding + offloading.

The following FALCON-7B model was fine-tuned using this repo: https://huggingface.co/AdrianBZG/falcon-7b-spanish-8bit

Setup

First, you need to clone the repo:

git clone https://github.com/AdrianBZG/LLM-distributed-finetune

Then, configure your aws credentials using the awscli package command aws configure. This will allow Ray to spawn the head node and provision workers with the auto-scaling mechanism. If you don't have awscli, you can install it using pip install awscli.

Working with the Ray cluster and submitting finetuning jobs

To spawn the cluster, simply run:

ray up ray_cluster.yaml

Once Ray has finished setting up the cluster, you can attach to the head node by doing:

ray attach ray_cluster.yaml

Now, to run a finetuning job, you can use the script finetune.py under /src.

An example usage is as below:

python finetune.py --model="tiiuae/falcon-7b" --num-workers 4 --data alpaca_data_cleaned.json

This will run a finetuning on the FALCON-7B model using 4 GPU workers, and the Alpaca instruction dataset. Feel free to adjust the arguments for your own purposes.

When you are finished, you can turn off the cluster with:

ray down ray_cluster.yaml

Changing DeepSpeed configuration

To tune the DeepSpeed configuration for your specific use case, edit the file on config/deepspeed.json. If you want to disable DeepSpeed, you can pass the --no-deepspeed parameter to the finetune.py script.

Datasets

I have successfully fine-tuned FALCON-7B on the following 2 datasets:

• Alpaca: https://huggingface.co/datasets/yahma/alpaca-cleaned
• Alpaca Spanish: https://huggingface.co/datasets/bertin-project/alpaca-spanish