Using Hugging Face*

 

 

 

 

Overview

Hugging Face* is an open source repository of Transformer and Diffuser-based models and datasets for generative AI (GenAI) and LLMs. The repository sorts training (such as DeepSpeed) and inference into tasks.

Intel collaborated with Hugging Face to create the Optimum for Intel® Gaudi® AI accelerators to allow Hugging Face models to run on Intel Gaudi AI accelerators. The library replaces generic Hugging Face code with Intel Gaudi accelerator configurations and is preconfigured for use on many popular training and inference tasks. Tasks and model examples in the library are fully validated and documented, making it easier to start training and inference.

Watch the video to get started using Hugging Face on Intel Gaudi accelerators.

Install Hugging Face Models​​​

To install Optimum for Intel® Gaudi® and run examples:

  1. Get access to an Intel® Gaudi® accelerator node.
  2. Using Intel® Gaudi® software v1.21.0,† run the Docker* image for PyTorch*. ​
  3. Install the Optimum for Intel® Gaudi® library​: ​
pip install optimum-habana==1.16.0

Install the Examples repository​​:

cd ~
git clone -b v1.16.0 https://github.com/huggingface/optimum-habana

Select Hugging Face Tasks from the following location, and then follow the README directions included in the examples:​​

cd ~/optimum-Habana/examples​

The installation is the most recent version of the optimum-habana library from Python* Package Index (PyPI) and checking out the same tag in GitHub for the optimum-habana model examples.​

To use the latest working version, install the optimum-habana library from source:

pip install git+https://github.com/huggingface/optimum-habana.git

Use the main branch from Github for all model examples.​

†If you are unsure what version of Intel® Gaudi® software you are running, see Software Verification and Support Matrix.