hf-finetune-sagemaker

This code allows you to fully finetune a model via Hugging Face using AWS's multi nodes and FSDP configuration for big models.

Requirements

• Python 3.10 or higher
• Sagemaker
• HF_TOKEN - Hugging Face's user access token with read and write permissions
• Your AWS RoleName to access AWS's gpus

Usage example for Jamba-1.5-Mini

1. Download or clone the repo to you local computer.
2. Create a new virtual environment based on Python >= 3.10.
3. Add HF_TOKEN and AWS_ROLE_NAME as environment variables.
4. pip install sagemaker (and pip install setuptools in case of python >= 3.12)
5. Run python sagemaker_training.py from the main dir.

Track your job in AWS training dashboard