Habana Labs and Hugging Face Partnership for Transformer Model Training

Habana Labs and Hugging Face have partnered to integrate Habana's SynapseAI software suite with the Hugging Face Optimum open-source library. This collaboration allows data scientists and machine learning engineers to accelerate the training of transformer models on Habana Gaudi processors with minimal code changes, reducing training costs and increasing productivity.

Hardware Acceleration via Habana Gaudi

Habana Gaudi training solutions are designed to provide high-efficiency, purpose-built deep learning processing. These solutions power Amazon EC2 DL1 instances and Supermicro's X12 Gaudi AI Training Server.

Key technical specifications and benefits of the Gaudi processor include:

  • Price/Performance: Gaudi delivers price/performance up to 40% lower than comparable training solutions.
  • Scalability: Every Gaudi processor integrates ten 100 Gigabit Ethernet ports, enabling the system to scale from a single processor to thousands of Gaudis efficiently.
  • Framework Support: The SynapseAI software suite is optimized for Gaudi performance and supports both TensorFlow and PyTorch frameworks, specifically targeting computer vision and natural language processing (NLP) applications.

Integration with Hugging Face Optimum

Through the Hugging Face Hardware Partner Program, the Habana SynapseAI suite is integrated into the Optimum library. This integration enables users to leverage Habana Gaudi's hardware acceleration for transformer models with just a few lines of code.

This partnership aims to expand the library of transformer models optimized for Habana Gaudi, facilitating high-quality training across various domains, including:

  • Natural Language Processing (NLP)
  • Computer Vision
  • Speech

Executive Perspectives

Industry leaders from both organizations emphasize the efficiency and accessibility of this partnership:

"We’re excited to partner with Hugging Face and its many open-source developers to address the growing demand for transformer models that benefit from the efficiency, usability, and scalability of the Gaudi training platform", said Sree Ganesan, head of software product management, Habana Labs.

"Habana Gaudi brings a new level of efficiency to deep learning model training, and we’re super excited to make this performance easily accessible to Transformer users with minimal code changes through Optimum", said Jeff Boudier, product director at Hugging Face.

Sources