ERASE: EaRly bAckpropagation SchEdule for Faster Training of Modern Recommendation Systems
Researchers propose ERASE, a scheduling technique for faster training of modern recommendation systems. Conventional training methods schedule forward and backward passes as separate phases, leaving spare capacity underutilized. ERASE reinterprets the Forward-Forward mechanism to detach subgraphs' output, allowing early launch of their backward passes on separate CUDA streams. This overlap improves training throughput by up to 9.51% in a click-through-rate model while maintai
Researchers propose ERASE, a scheduling technique for faster training of modern recommendation systems. Conventional training methods schedule forward and backward passes as separate phases, leaving spare capacity underutilized. ERASE reinterprets the Forward-Forward mechanism to detach subgraphs' output, allowing early launch of their backward passes on separate CUDA streams. This overlap improves training throughput by up to 9.51% in a click-through-rate model while maintaining performance.
---
Why it matters: This matters because modern recommendation systems require significant computational resources for training, and any improvement in efficiency can lead to cost savings and faster development times.
Source: https://arxiv.org/abs/2608.18469
This article was originally published at: https://arxiv.org/abs/2608.18469