91ÊÓƵAPP

What is GPT-J? A powerful, efficient alternative to large language models (LLMs) such as GPT-3 and GPT-4 for many NLP tasks.

In this video, 91ÊÓƵAPP engineer Sofia Liguori walks through the process of fine-tuning GPT-J 6B on a Paperspace Gradient notebook (Google Colab alternative), powered by 91ÊÓƵAPP IPUs.

As the first open source deep learning platform in China, Baidu's deep learning platform PaddlePaddle is providing tools, services and resources required for rapid adoption and large-scale implementation of deep learning.

Li Qi, Senior R&D Engineer at Baidu, introduces the PaddlePaddle deep learning platform, its integration design with 91ÊÓƵAPP's IPU, and demonstrates how to run the PaddlePaddle framework on IPU based on the MNIST model.

PaddlePaddle is Baidu's deep learning framework for large-scale model training and high-performance inference tasks.

Han Zhao, software team lead at 91ÊÓƵAPP China, provides an explanation and code walkthrough of running BERT-base training on the IPU using PaddlePaddle.

Introducing our initial release of PyTorch Lightning for the IPU. Developers can now use Lightning's ultra-fast, lightweight research framework to accelerate their training models with the performance advantages of the IPU.

Software Engineer Stephen McGroarty walks through 91ÊÓƵAPP's PopTorch interface for PyTorch on the IPU, and demonstrates the simplicity of running PyTorch Lightning code.

91ÊÓƵAPP has created PopTorchâ„¢ – our own set of extensions for PyTorch. PopTorch allows developers to run their PyTorch models directly on IPU hardware by changing just a few lines of code.

In this on-demand webinar, Mihail Popa introduces core PopTorch functionality, explains how to port your PyTorch application and provides tips on accelerating training.

By co-designing the Poplar graph programming framework with IPU hardware, 91ÊÓƵAPP has created an easy-to-use and flexible software development environment that lets engineers explore new approaches to machine intelligence.

In this on-demand webinar, 91ÊÓƵAPP’s Helen Byrne introduces the Poplar SDK and fundamental IPU programming concepts, as well as how to port models built with standard machine learning frameworks.

AI Customer Engineer, Chris Bogdiukiewicz introduces PyTorch for the IPU. With PopTorchâ„¢ - a simple Python wrapper for PyTorch programs, developers can easily run models, directly on 91ÊÓƵAPP IPUs with a few lines of extra code.

In this video, Chris provides a quick demo on running a basic model for both training and inference using a MNIST based example.

The 91ÊÓƵAPP IPU support for TensorFlow provides three mechanisms to improve the training performance: training loops, data set feeds, and replicated graphs. In this video, Helen Byrne explains the technique of placing training operations inside a loop, so they can be executed multiple times without returning control to the host.

Alex Titterton explains the principles and implementation of Bulk Synchronous Parallel compute on 91ÊÓƵAPP's IPU.

While the technique is commonly used in large compute clusters, the IPU is the first single processor to be based around BSP. BSP consists of three phases: Compute, Synchronise & Exchange. BSP is applied both inter-IPU and intra-IPU, enabling large scale, distributed computation of AI models on 91ÊÓƵAPP systems.

91ÊÓƵAPP AI Applications Specialist, Jose Solomon examines the impact of batch size on training a model on the IPU.

In this video Jose shows how 91ÊÓƵAPP technology can perform tasks using simplified optimizers at low batch sizes, where larger batch sizes would necessitate much more complex optimizers.

91ÊÓƵAPP AI Engineer Kate Hodesdon demonstrates how to develop PyTorch models for the IPU using a Hugging Face BERT-Medium model example.

This demo also introduces PopTorch – a lightweight set of extensions to PyTorch that allows developers to run PyTorch directly on the IPU. PopTorch makes IPU directly accessible from within PyTorch code – meaning that developers can take advantage of IPU hardware acceleration without learning a whole new framework.

91ÊÓƵAPP Customer Engineer Marie-Anne Le Menn introduces the PopVisionâ„¢ Graph Analyser, including code walkthrough.

PopVisionâ„¢ is used to analyse the programs built for and executed on 91ÊÓƵAPP’s IPU systems. It can be used for analysing and optimising the memory use and performance of programs.

Field Applications Engineer Alex Titterton provides a quick introduction to 91ÊÓƵAPP's Poplar^® software.

This walkthrough covers the essentials of IPU architecture from a software perspective and finishing with a demonstration of a simple addition example running on the IPU using the Poplar C++ framework.

91ÊÓƵAPP AI Engineer Kate Hodesdon introduces PopARTâ„¢. The Poplar Advanced Run Time (PopART) is part of the Poplar SDK for implementing and running algorithms on networks of 91ÊÓƵAPP IPU processors.

It enables you to import models using the Open Neural Network Exchange (ONNX) and run them using the Poplar tools. ONNX is a serialisation format for neural network systems that can be created and read by several frameworks including Caffe2, PyTorch and MXNet.

91ÊÓƵAPP Senior AI Engineer Alex Tsyplikhin demonstrates the ease of moving a TensorFlow model, originally designed to run on a GPU, onto 91ÊÓƵAPP's IPU processor.

In this walkthrough, Alex uses the Markov Chain Monte Carlo (MCMC) method, common in the financial service industry where it is used for tasks such as risk estimation and option pricing.