Graphcore and Hugging Face Expand IPU-Ready Transformer Models

Graphcore and Hugging Face have expanded the range of machine learning modalities and tasks available in Hugging Face Optimum, providing developers with 10 off-the-shelf transformer models optimized for Graphcore's Intelligence Processing Unit (IPU). This expansion lowers the barrier for innovators to deploy high-performance transformer models across natural language processing, speech, and computer vision tasks.

Optimized Model Lineup by Modality

Graphcore and Hugging Face now provide IPU configuration files and ready-to-use pre-trained and fine-tuned weights for the following models:

Natural Language Processing (NLP)

  • GPT-2: A generative pre-trained transformer for text generation, trained in a self-supervised fashion on a large English corpus.
  • RoBERTa: A robustly optimized BERT approach using a masked language modeling (MLM) objective, primarily intended for fine-tuning on downstream tasks.
  • DeBERTa: An improved version of BERT and RoBERTa utilizing a disentangled attention mechanism and an enhanced mask decoder to increase pretraining efficiency and downstream performance.
  • BART: A sequence-to-sequence (seq2seq) model with a bidirectional encoder and autoregressive decoder, effective for text generation tasks like summarization and translation.
  • T5: A text-to-text transfer transformer that converts all NLP problems into a unified text-to-text format for transfer learning.
  • LXMERT: A multimodal transformer that learns vision and language representations using three encoders (object relationship, language, and cross-modality) to achieve state-of-the-art results on VQA and GQA datasets.

Computer Vision

  • ViT (Vision Transformer): An image recognition model that processes images by dividing them into small patches, which are then encoded by a transformer embedding.

Speech

  • HuBERT (Hidden-Unit BERT): A self-supervised speech recognition model that learns combined acoustic and language models over continuous inputs.
  • Wav2Vec2: A self-supervised model for automatic speech recognition that learns speech representations from unlabeled data via a contrastive pretraining objective.

Hardware and Software Integration

Performance gains are driven by the integration of Hugging Face Optimum with Graphcore's latest hardware and software stack.

Bow IPU Hardware

The Bow IPU utilizes Wafer-on-Wafer (WoW) 3D stacking technology. Compared to the previous generation, the Bow IPU delivers:

  • Up to 350 teraFLOPS of AI compute (a 40% performance increase).
  • Up to 16% improvement in power efficiency.

Users of Hugging Face Optimum can transition to Bow processors without requiring any code changes.

Poplar SDK

Optimum provides a plug-and-play experience with the Poplar SDK (version 2.5), which integrates with standard machine learning frameworks including PyTorch, PyTorch Lightning, and TensorFlow. This compatibility allows developers to port models from other compute platforms to the IPU using orchestration tools like Docker and Kubernetes.

Ecosystem and Accessibility

By integrating IPUs with Hugging Face, developers can leverage both optimized models and the extensive datasets available on the Hugging Face Hub. This partnership aims to reduce the time and specialized skill required to optimize transformer performance for real-world applications in fields such as cybersecurity, drug discovery, and voice call automation.

Sources