Recsys reimagined as generative modeling at billion-user scale

What it does

The project hosts the reference implementation for Meta’s Generative Recommenders (GR), a framework that recasts the standard DLRM paradigm as generative modeling over sequences of user actions. It bundles the HSTU architecture and M-FALCON algorithms with optimized CUDA and Triton kernels, plus reproducible benchmarks for MovieLens, Amazon Reviews, and a synthetic 3-billion-rating dataset. A full DLRM-v3 training and inference stack is also included for production-scale evaluation.

The interesting bit

The central bet is that recommendation is fundamentally a generation problem, not just a prediction one. By treating user actions as tokens in a sequential transducer, the authors claim they can push models toward trillion-parameter scale and observe scaling laws in live, billion-user systems for the first time. The efficiency story is equally aggressive: custom FlashAttention-3-based kernels and fractal data expansion purportedly yield 10× to 1000× speedups over standard pipelines.

Key highlights

• HSTU-large improves over SASRec by up to 23.1% HR@10 on MovieLens-20M and 56.7% on Amazon Books using identical configurations.
• Includes DLRM-v3 benchmarks compatible with MLPerf loadgen for multi-GPU training and inference.
• Ships with custom ops/cpp/hstu_attention kernels built on FlashAttention V3, tuned for H100 efficiency.
• Supports fractal expansion to blow up MovieLens-20M into a 3-billion-rating synthetic dataset.
• Verified reproducible results as of April 2024 via provided Gin config files.

Caveats

• The trillion-parameter and billion-user production claims come from the ICML'24 paper; the open-source release focuses on smaller public and synthetic benchmarks.
• Public experiments require a GPU with 24GB+ HBM and a narrow dependency stack (CUDA 12.4, Python 3.10, FBGEMM, TorchRec).
• Baseline comparisons are asymmetrical: the README notes BERT4Rec and GRU4Rec use full negatives while HSTU results rely on 128/512 sampled negatives, which slightly favors the baselines.

Verdict

A strong candidate if you are building large-scale sequential recommenders and want a hardware-efficient, research-backed successor to SASRec-style architectures. Look elsewhere if you need a framework-agnostic, plug-and-play library; this is a PyTorch-and-CUDA research artifact tightly coupled to Meta’s RecSys stack.