AI_dev Europe 2024

Hugging Face Accelerate is an open-source library designed to make machine learning model training and inference simple and device-agnostic through an easy-to-use interface. It does so by keeping things at a low level, minimising the abstraction while maximising the freedom a user has over their code. Over the last two years, it has garnished new features to enhance usability in the ecosystem, such as; training agnostically between all of the latest ML acceleration hardware (CUDA, XLA, NPU, and XPU), training support in lower precisions for speed and memory efficiency, and easy-to-deploy large model inference that can scale with the amount of compute available through features like big model inference and pipeline parallelism. In this presentation, we will first cover some background into just what Accelerate is as a framework. Building on this overview, we will then detail how it has impacted the ML world at large. By the end of this talk, we will not only have identified the core parts of Accelerate but also introduce you to its easy-to-use device-agnostic API to help kickstart the journey into large-scale computing and local-first deployment of machine learning models.