Explore the open-source library Hugging Face Accelerate, designed to simplify machine learning model training and inference across various devices. Learn how this framework maintains a low-level approach, minimizing abstraction while maximizing code flexibility. Discover its evolution over the past two years, including support for training on diverse ML acceleration hardware (CUDA, XLA, NPU, and XPU), lower precision training for improved speed and memory efficiency, and scalable large model inference. Gain insights into Accelerate's impact on the ML landscape and get introduced to its user-friendly, device-agnostic API. By the end of this 23-minute conference talk, acquire the knowledge needed to begin your journey into large-scale computing and local-first deployment of machine learning models using Hugging Face Accelerate.