AdaEmbed: Adaptive Embedding for Large-Scale Recommendation Models
summary
AdaEmbed is a system that performs in-training embedding pruning for DLRMs, ranking billions of embedding rows by runtime access frequency and a moving average of aggregate gradient norms and then pruning low-importance rows across workers. It adds a lookup table shim (virtual-memory style indirection) to avoid expensive reallocations and a coordinator that samples importance distributions to limit coordination overhead. Deployed on 100+ GPUs across production models, it reduces deployment embedding size 35–60% and speeds execution ~11–34% while improving accuracy.
AI-written summary · about
chapters (9)
- 0:00Introduction to the problem
- 1:04Recommendation model mechanics
- 3:05The case for adaptive pruning
- 5:10AdaEmbed system architecture
- 6:12Identifying embedding importance
- 8:01Overcoming coordination overhead
- 9:48Virtual memory for embeddings
- 11:28Performance evaluation
- 13:14Conclusion and results
related talks
15mAccelerating Neural Recommendation Training with Embedding Scheduling
20mOPER: Optimality-Guided Embedding Table Parallelization for Large-scale...
12mQuickUpdate: a Real-Time Personalization System for Large-Scale Recommendation Models
38mEmbeddings! Embeddings everywhere!
38mHave you met Ada? Word Embeddings with Laravel
20mDataEngBytes 2023 - PER-T1-06 - Kamal Hossain
13mTraining Embedding Model Resiliently for Multimodal M...
37mDeep Recommender Systems