ucxx-cu12

What it is and what it does

ucxx-cu12 is a Python interface to UCX, a low-latency communication library optimized for GPU clusters. It wraps the C++ UCX library to enable efficient point-to-point messaging between processes, with native support for CUDA device memory, managed memory, and asynchronous transfers. The package depends on libucxx-cu12 (the compiled C++ backend), cuda-core, numpy, nvidia-ml-py, and rmm-cu12 (RAPIDS memory manager for GPU buffers).

The library is designed for high-performance distributed computing scenarios where GPU-to-GPU communication is critical. It supports multiple memory types (host, CUDA device, managed, and async) and can be used in both synchronous (blocking) and asynchronous (event-driven) modes. The package is actively maintained and targets current Python versions (3.11+), making it suitable for modern GPU-accelerated applications that need low-latency inter-process or inter-node communication.

Use it for:

• Implement GPU-to-GPU communication in distributed deep learning frameworks without copying data to host memory.
• Build low-latency messaging systems between CUDA-enabled processes on the same or different nodes.
• Benchmark and optimize communication performance between GPUs using the included send/receive benchmark tools.
• Integrate UCX transport into custom distributed computing applications requiring fine-grained memory control.
• Test CUDA memory handling (device, managed, async) in multi-GPU environments with standardized communication patterns.

Worth the install?

AI-flagged interpretation of the facts on this page — verify before relying

ucxx-cu12 provides Python bindings for the Unified Communication X (UCX) library, enabling efficient point-to-point communication with native support for CUDA memory types and GPU-to-GPU transfers.

Yes, if you are building GPU-accelerated distributed systems or need low-latency inter-GPU communication. The package is actively maintained, permissively licensed, and has no known vulnerabilities. Medium install friction is acceptable for its specialized use case. Not recommended for CPU-only or single-GPU workloads where simpler communication libraries suffice.

Install

ucxx-cu12 on PyPI

pip install ucxx-cu12

uv add ucxx-cu12

poetry add ucxx-cu12

Installing ucxx-cu12

Before you install

Medium install friction due to compiled dependencies (libucxx-cu12, cuda-core, rmm-cu12) and platform-specific wheels. Active maintenance with recent releases; last commit 2026-08-14 and version 0.51.1 released 2026-08-13 suggest current development.

License in practice

BSD-3-Clause permissive license allows commercial and private use with minimal restrictions, requiring only license and copyright notice retention.

Quickstart

pip install ucxx-cu12

import ucxx
import numpy as np

# Create a context for communication
ctx = ucxx.create_context()
# Use ctx for point-to-point communication with UCX

Requires CUDA 12.x runtime and compatible NVIDIA GPU; libucxx-cu12 and cuda-core must be available in the environment; Python >=3.11.

Verify before relying

• Whether the package provides synchronous and asynchronous APIs as described in benchmarks section, or if those are separate modules.
• Specific Python API surface and whether it exposes UCX connection establishment, send/receive, and memory management directly.
• Whether numpy and rmm-cu12 are optional or required at runtime for all use cases.

Package facts

Evidence: ucxx_cu12-0.51.1-cp311-abi3-manylinux_2_26_aarch64.manylinux_2_28_aarch64.whl; ucxx_cu12-0.51.1-cp311-abi3-manylinux_2_27_x86_64.manylinux_2_28_x86_64.whl

Tags

More Python Modules packages