Across industries, recommendation systems are evolving from pointwise architectures (focusing on per user activity) toward generative sequential architectures (building on a user's past activity history).

One popular architecture in the generative sequential space is the Generative Recommender (GR), which models user behavior as item-activity token sequences. This approach improves reasoning over time and provides better long-context personalization compared with traditional Deep Learning Recommendation Models (DLRM). However, these increasingly capable GRs also introduce new system-level challenges at scale, including variable-length sequences, memory pressure from large embedding tables, and I/O bottlenecks during data ingestion.

In this blog post, we share engineering techniques - validated by LinkedIn’s large-scale GR deployments - that address common challenges and significantly improve training efficiency without impacting model quality, including data pipeline, compute, and training lifecycle optimizations. When applied together, these system optimizations reduced end-to-end GPU hours by up to 65% in internal production training workloads.

Generative and Sequential Recommenders at industrial scale

Traditionally, LinkedIn’s recommendation systems for Feed and Ads ranking have relied on DLRM-based architectures, which are effective at modeling sparse feature interactions but are limited in capturing rich sequential user behavior and context across sessions.

Over time, member activity on the platform has become increasingly dynamic and content consumption has become more sequence-driven, which has created an opportunity to improve how recommendations adapt to evolving user intent in real time. To address these factors, LinkedIn introduced a Generative Recommender, a transformer-based sequential ranking system designed to better model user trajectories and contextual signals (ref). The GR was deployed under strict production constraints and delivered significant improvements, including +2.10% session time spent in online Feed A/B tests.

Sequential and Generative Recommender models differ from traditional DLRM models in following ways: