HSTU Generative Recommenders on Triton#

Hierarchical Sequential Transduction Units (HSTU) power Generative Recommenders (GRs): recommendation workloads reformulated as generative modeling over high-cardinality, non-stationary event streams. HSTU supports both retrieval and ranking style tasks.

Triton Inference Server can serve HSTU models through the PyTorch backend using ahead-of-time (AOT) Inductor packages (platform: "torch_aoti"). Training, export, KV-cache runtime, and end-to-end examples live in NVIDIA’s recsys-examples repository rather than in this tutorials tree.

Where to go next#

Resource

Description

HSTU overview

Architecture, training, and inference entry points

HSTU inference

Inference features, KV-cache, AOTInductor export, and KuaiRand examples

PyTorch AOTI on Triton

torch_aoti model repository layout and configuration

[!NOTE] Use the recsys-examples guides for building, exporting, and validating HSTU models. This page only points Triton users at that workflow and the Torch AOTI serving path.