arXiv:2605.23066v2 Announce Type: replace-cross 
Abstract: In a landscape of high-performance distributed ML systems, JAX has emerged as a framework of choice. However, JAX's modular design philosophy leaves it without a standardized checkpointing solution. In this paper, we introduce Orbax, a modular, JAX-native checkpointing library that abstracts the complexities of distributed accelerator systems while also providing flexibility for user-friendly checkpoint manipulations throughout the ML model lifecycle. We demonstrate performance exceeding comparable PyTorch competitors by up to 3.5$\times$ for saving and 2$\times$ for loading. The library is available at https://github.com/google/orbax.

تم تقديم Orbax كمكتبة نقاط تحقق معيارية ومخصصة لـ JAX، مصممة لتبسيط تعقيدات أنظمة التعلم الآلي الموزعة، متفوقة على الحلول المماثلة في PyTorch من حيث أداء الحفظ والتحميل.

Orbax se ha presentado como una biblioteca de puntos de control modular y nativa de JAX, diseñada para simplificar las complejidades de los sistemas de aprendizaje automático distribuidos, superando a soluciones similares en PyTorch en rendimiento de guardado y carga.

Orbax a été introduit comme une bibliothèque de point de contrôle modulaire et native à JAX, conçue pour simplifier les complexités des systèmes d'apprentissage automatique distribués, surpassant les solutions similaires en PyTorch en termes de performance de sauvegarde et de chargement.

Orbax has been introduced as a modular, JAX-native checkpointing library designed to simplify the complexities of distributed machine learning systems, outperforming similar solutions in PyTorch for saving and loading performance.