Explore a cutting-edge approach to optimizing deep learning recommendation models (DLRMs) in this conference talk from OSDI '23. Dive into AdaEmbed, an innovative system designed to reduce embedding size while maintaining model accuracy through in-training embedding pruning. Learn how this technique leverages heterogeneous access patterns and weights across embedding rows to dynamically identify and prune less important embeddings at scale. Discover the potential of AdaEmbed to significantly reduce deployment costs and improve model execution speed in large-scale recommendation systems. Gain insights into the challenges of working with DLRMs containing billions of embeddings and how AdaEmbed addresses these issues in industrial settings. Understand the impact of this approach on embedding size reduction, model execution speed, and accuracy gains in real-world applications.
Overview
Syllabus
OSDI '23 - AdaEmbed: Adaptive Embedding for Large-Scale Recommendation Models
Taught by
USENIX