Srihari Reddy

Rengan Xu, Junjie Yang, Yifan Xu et al.

The integration of hardware accelerators has significantly advanced the capabilities of modern recommendation systems, enabling the exploration of complex ranking paradigms previously deemed impractical. However, the GPU-based computational costs present substantial challenges. In this paper, we demonstrate our development of an efficiency-driven approach to explore these paradigms, moving beyond traditional reliance on native PyTorch modules. We address the specific challenges posed by ranking models' dependence on categorical features, which vary in length and complicate GPU utilization. We introduce Jagged Feature Interaction Kernels, a novel method designed to extract fine-grained insights from long categorical features through efficient handling of dynamically sized tensors. We further enhance the performance of attention mechanisms by integrating Jagged tensors with Flash Attention. Our novel Jagged Flash Attention achieves up to 9x speedup and 22x memory reduction compared to dense attention. Notably, it also outperforms dense flash attention, with up to 3x speedup and 53% more memory efficiency. In production models, we observe 10% QPS improvement and 18% memory savings, enabling us to scale our recommendation systems with longer features and more complex architectures.

Sohini Roychowdhury, Adam Holeman, Mohammad Amin et al.

For online ad-recommendation systems, processing complete user-ad-engagement histories is both computationally intensive and noise-prone. We introduce Dynamix, a scalable, personalized sequence exploration framework that optimizes event history processing using maximum relevance principles and self-supervised learning through Event Based Features (EBFs). Dynamix categorizes users-engagements at session and surface-levels by leveraging correlations between dwell-times and ad-conversion events. This enables targeted, event-level feature removal and selective feature boosting for certain user-segments, thereby yielding training and inference efficiency wins without sacrificing engaging ad-prediction accuracy. While, dynamic resource removal increases training and inference throughput by 1.15% and 1.8%, respectively, dynamic feature boosting provides 0.033 NE gains while boosting inference QPS by 4.2% over baseline models. These results demonstrate that Dynamix achieves significant cost efficiency and performance improvements in online user-sequence based recommendation models. Self-supervised user-segmentation and resource exploration can further boost complex feature selection strategies while optimizing for workflow and compute resources.

Dinesh Ramasamy, Shakti Kumar, Chris Cadonic et al.

Sequence-based recommendations models are driving the state-of-the-art for industrial ad-recommendation systems. Such systems typically deal with user histories or sequence lengths ranging in the order of O(10^3) to O(10^4) events. While adding embeddings at this scale is manageable in pre-trained models, incorporating them into real-time prediction models is challenging due to both storage and inference costs. To address this scaling challenge, we propose a novel approach that leverages vector quantization (VQ) to inject a compact Semantic ID (SID) as input to the recommendation models instead of a collection of embeddings. Our method builds on recent works of SIDs by introducing three key innovations: (i) a multi-task VQ-VAE framework, called VQ fusion that fuses multiple content embeddings and categorical predictions into a single Semantic ID; (ii) a parameter-free, highly granular SID-to-embedding conversion technique, called SIDE, that is validated with two content embedding collections, thereby eliminating the need for a large parameterized lookup table; and (iii) a novel quantization method called Discrete-PCA (DPCA) which generalizes and enhances residual quantization techniques. The proposed enhancements when applied to a large-scale industrial ads-recommendation system achieves 2.4X improvement in normalized entropy (NE) gain and 3X reduction in data footprint compared to traditional SID methods.