NVIDIA-Merlin/Merlin

NVIDIA Merlin is an open source library providing end-to-end GPU-accelerated recommender systems, from feature engineering and preprocessing to training deep learning models and running inference in production.

What it solves

NVIDIA Merlin addresses the challenges of building and scaling high-performance recommender systems. It specifically targets bottlenecks in feature engineering, training large-scale deep learning models, and deploying these models into production, especially when dealing with datasets that reach hundreds of terabytes in size.

How it works

Merlin is a suite of GPU-accelerated libraries that optimize every stage of the recommendation pipeline:

  • NVTabular: Handles ETL and feature engineering for massive tabular datasets using GPU acceleration.
  • HugeCTR: A training framework designed to scale large embedding tables across multiple GPUs and nodes.
  • Merlin Models: Provides a library of standard recommendation model architectures and optimized data loaders for TensorFlow, PyTorch, and FastAI.
  • Transformers4Rec: Specializes in sequential and session-based recommendations using modular building blocks compatible with PyTorch.
  • Merlin Systems: Integrates models with feature stores and nearest neighbor search to create end-to-end pipelines served via Triton Inference Server.
  • Merlin Core: Provides the underlying dataset abstractions and common schemas used across the ecosystem.

Who it’s for

It is designed for data scientists, machine learning engineers, and researchers who need to build industrial-scale recommender systems that exceed the memory limits of single GPUs or CPUs.

Highlights

  • Massive Scalability: Optimized to support datasets up to hundreds of terabytes.
  • GPU Acceleration: Leverages RAPIDS cuDF and Dask for fast data manipulation and distributed processing.
  • Memory Management: Capable of distributing large embedding tables that exceed available GPU and CPU memory.
  • End-to-End Pipeline: Covers the entire lifecycle from raw data preprocessing to production serving with Triton Inference Server.