---
id: libucxx-cu12
version: "0.51.1"
license: BSD-3-Clause
license_treatment: permissive
maintenance: active
---

# libucxx-cu12 — Python Bindings for the Unified Communication X library (UCX)

License: permissive · Maintenance: active · Downloads: 173.9K/mo

## What it is and what it does

libucxx-cu12 is a Python interface to UCX, a low-latency communication library designed for high-performance computing. It wraps a C++ implementation that provides object-oriented abstractions over UCX's transport layer, with built-in support for CUDA device memory, managed memory, and asynchronous operations. The package is part of the RAPIDS ecosystem and targets distributed computing workloads where GPU-to-GPU or host-to-GPU communication needs to be fast and efficient.

The library supports both synchronous and asynchronous communication patterns, with benchmarking tools included for performance testing. It integrates with librmm-cu12 for GPU memory allocation and can work with standard host memory, CUDA device memory, or unified memory. Installation requires a CUDA 12 environment and the two compiled runtime dependencies; the package itself is distributed as pre-built wheels for Linux on x86_64 and aarch64 architectures.

Use it for:

- Build distributed GPU computing clusters where multiple nodes need fast inter-GPU communication with minimal latency.
- Implement custom communication layers for data-parallel machine learning frameworks that need fine-grained control over memory transfers.
- Benchmark and profile UCX transport performance across different memory types and message sizes in a CUDA environment.
- Integrate UCX-based messaging into data processing pipelines for multi-GPU analytics.
- Develop low-latency point-to-point communication between CUDA applications without going through higher-level abstractions.

## Worth the install?

AI-flagged interpretation of the facts above — verify before relying.

Provides Python bindings for UCX (Unified Communication X), enabling object-oriented C++ interfaces for high-performance inter-process communication with native CUDA memory support.

Yes, if you need GPU-aware inter-process communication in a CUDA 12 environment and are willing to manage the compiled dependency chain. The package is actively maintained with no known vulnerabilities and carries a permissive BSD-3-Clause license. Medium install friction is acceptable for specialized HPC and distributed computing use cases; not suitable for general Python projects or non-CUDA environments.

## Install

pip install libucxx-cu12
uv add libucxx-cu12
poetry add libucxx-cu12

## Installing libucxx-cu12

Before you install: Medium install friction due to compiled binary wheels for specific architectures (x86_64, aarch64) and two runtime dependencies on CUDA-specific libraries (librmm-cu12, libucx-cu12). Active maintenance with release from 2026-08-13.

License in practice: BSD-3-Clause permissive license allows use in most projects without significant restrictions, though attribution is required.

Quickstart:

pip install libucxx-cu12

# Requires librmm-cu12 and libucx-cu12 runtime dependencies to be available

Requires CUDA 12 runtime environment and compatible GPU hardware; wheels are architecture-specific (x86_64 or aarch64 Linux only).

Verify before relying:

- Whether Python version support is truly unspecified or if there are undocumented minimum/maximum Python version constraints.
- Whether librmm-cu12 and libucx-cu12 are automatically installed as dependencies or must be pre-installed separately.
- Concrete usage examples beyond the build and benchmark documentation provided in the description excerpt.

## Package facts

- License: BSD-3-Clause (permissive)
- Python support: unspecified
- Install friction: medium
- Maintenance: active
- Downloads: 173.9K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

UCX python bindings, high-performance communication library, CUDA-aware messaging, inter-process communication, GPU memory transfer, distributed computing transport, low-latency networking, gpu-communication, cuda-native, hpc

[View on SkillFed](https://skillfed.io/packages/libucxx-cu12) · [View on PyPI](https://pypi.org/project/libucxx-cu12/)