Faster Than Light Optimizing Generative Recommender Training Efficiency At Linkedin

LinkedIn’s recommender-training account shows how much computing can be spent on empty space. Users have activity histories of different lengths, but the original training setup padded them to a maximum length, creating work that carried no additional history. The team changed its data loader and model to process packed sequences instead. For Feed training, it reports more than a 30% reduction in GPU hours from that change. The result illustrates a concrete efficiency opportunity: reduce the calculations devoted to padding while preserving the activity data the recommendation model learns from.