NVIDIA RapidsMPF Documentation

Building high-performance GPU pipelines is hard. Each stage must move data efficiently between GPUs and processes, synchronize work, and manage limited device memory. NVIDIA RapidsMPF, including its nascent Streaming Engine and Out-of-Core (OOC) shuffle, provides a unified framework for asynchronous, multi-GPU pipelines using simple streaming primitives — Channels, Actors, and Messages — built on RAPIDS components: rmm and ucxx.

RapidsMPF’s design leverages Explicit Data Parallelism (SPMD-style coordination) combined with a local CSP-style streaming model, enabling the engine to overlap I/O, computation, and communication. This makes it possible to handle out-of-core processing efficiently (via Spilling) and integrate seamlessly with frontend query engines such as Polars. The result is clean, composable, and scalable GPU streaming — from single-node prototypes to large-scale, multi-GPU deployments. See the Glossary for definitions of key concepts.

Contents:

• Getting Started
  • Build from Source
  • MPI
  • UCX
  • rrun — Distributed Launcher
• Background
  • Shuffle Architecture
  • Streaming execution
  • Channels
  • Actors
• Python
  • Quickstart
  • API Reference
  • Integrations
• C++
  • Coverage
  • Shuffle Service
  • rrun — Distributed Launcher
• Configuration Options
  • Available Options
• Statistics
  • Example Output
• Glossary