Bagua - Lightweight Distributed Learning on Kubernetes

Explore a conference talk on Bagua, a lightweight distributed learning framework for Kubernetes developed by Kuaishou Technology and ETH Zürich. Discover how Bagua supports high-performance distributed deep learning without requiring special network devices or restrictive scheduling. Learn about its innovative communication algorithms and seamless integration with Kubernetes, enabling horizontal scaling of training with excellent speedup guarantees using ordinary ethernet connections. Examine Bagua's effectiveness across various scenarios and models, including ResNet on ImageNet, Bert Large, and large-scale industrial applications at Kuaishou. Gain insights into its performance advantages, outperforming PyTorch-DDP, Horovod, and BytePS in end-to-end training time by up to 1.95 times in production Kubernetes clusters. Understand how Bagua addresses challenges in recommendation model training with massive parameters, video/image understanding with billions of samples, and ASR with terabyte-level datasets.