Accelerating Sequential Recommender Models in the PyTorch Ecosystem
Talkat KubeCon + CloudNativeCon + OpenInfra Summit + PyTorch Conference China 2026 · 2026-09-07..09-09·
summary
This talk covers rewriting the SASRec Transformer-decoder sequential recommender from TensorFlow to PyTorch, showing how PyTorch simplifies modeling and revealing gaps between academic prototypes and production systems. It presents HSTU, a specialized attention operator for sequential recommendation that differs structurally from standard Transformer attention, and discusses GPU-focused optimizations (kernel fusion, handling variable-length sequences) and PyTorch integration via FBGEMM. It ends with comparisons between sequential recommendation models and LLMs, and observations about modeling and systems optimization trends for recommendation workloads.
AI-written summary · about
related talks
08mTorchRec: A PyTorch domain library for large-scale recommender systems
33mNeural Networks for Recommender Systems
29mFurther enhancements of recommender systems using deep learning
25mScaling Inference of O(10K)-length Sequence Recommendation Models Using...
32mLet's talk about Recommendation Systems!
10mLightning Talk: Scaling Recommendation Systems To 2K GPUs and Beyond
32mResurrecting a Recommendations Platform
01h 21mUnlocking Personalization: A Deep Dive into Modern Recommendation Algorithms