Transformer models, powered by deep learning, achieve top-tier results across various machine learning tasks including natural language processing, computer vision, and speech. Yet, training these models at scale often demands significant computational resources, making the process slow, complex, and expensive.
Today, Habana Labs, a leader in high-efficiency deep learning processors, and Hugging Face, the hub for transformer models, announced a partnership to simplify and accelerate transformer model training. By integrating Habana's SynapseAI software suite with Hugging Face's Optimum library, data scientists and engineers can now speed up their training jobs on Habana processors with minimal code changes, boosting productivity and cutting costs.
Habana's Gaudi training solutions, used in Amazon EC2 DL1 instances and Supermicro servers, offer up to 40% lower price-performance compared to alternatives. Each Gaudi processor includes ten 100 Gigabit Ethernet ports for easy scaling from single units to thousands. SynapseAI, optimized for Gaudi, supports TensorFlow and PyTorch, focusing on computer vision and NLP.
Hugging Face, with over 60,000 GitHub stars, 30,000 models, and millions of monthly visits, is a cornerstone of the machine learning community. Through its Hardware Partner Program, Hugging Face provides Gaudi hardware integrated with the Transformer toolset, expanding the library of ready-to-use models for diverse applications.
"We're excited to partner with Hugging Face to meet the growing demand for efficient, scalable transformer training on Gaudi," said Sree Ganesan, head of software product management at Habana Labs.
Jeff Boudier, product director at Hugging Face, added, "Habana Gaudi brings new efficiency to deep learning training, and we're thrilled to make this performance accessible with minimal code changes via Optimum."
For more details, visit Habana Developer or the Hugging Face Habana page.