nccl4py

What it is and what it does

nccl4py is a Python wrapper around NVIDIA's Collective Communications Library (NCCL), designed to bring GPU-accelerated communication to Python applications running on multi-GPU and multi-node clusters. It abstracts NCCL's C API into a Pythonic interface, allowing distributed computing frameworks and custom applications to coordinate GPU computations across multiple nodes without dropping to C code.

The package targets researchers and engineers building distributed machine learning systems, high-performance computing applications, and other workloads that need efficient all-reduce, broadcast, and other collective operations across GPUs. It depends on numpy for array handling, packaging for version management, and CUDA runtime libraries (cuda-core and cuda-pathfinder) to interface with GPU hardware. Installation requires CUDA 12.x or 13.x and is limited to Linux on x86_64 or aarch64 architectures.

Use it for:

• Coordinate gradient synchronization across multiple GPUs during distributed deep learning training.
• Implement custom all-reduce and broadcast operations in multi-node HPC applications.
• Build distributed data-parallel inference pipelines that need efficient GPU-to-GPU communication.
• Integrate NCCL communication into Python-based simulation or numerical computing frameworks.
• Prototype distributed algorithms that require low-latency collective operations on GPU clusters.

Worth the install?

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

nccl4py provides Python bindings to NVIDIA's Collective Communications Library (NCCL), enabling GPU-accelerated multi-GPU and multi-node communication for distributed computing workloads on Linux clusters.

Yes, if you are building distributed GPU applications on Linux clusters and need low-latency collective communication. The active maintenance, permissive Apache-2.0 license, and strong repository signals indicate a well-supported project. Medium install friction (CUDA runtime dependency, platform-specific wheels) is typical and expected for GPU libraries. No known security vulnerabilities. Not suitable for Windows or macOS, or for single-GPU workloads.

Install

nccl4py on PyPI

pip install nccl4py

uv add nccl4py

poetry add nccl4py

Installing nccl4py

Before you install

Medium install friction due to CUDA runtime dependencies (cuda-core, cuda-pathfinder) and platform-specific wheels (x86_64 and aarch64 Linux only). Active maintenance with a recent release (3 days old) and strong repository signals (4996 stars, last commit 2026-08-14).

License in practice

Apache-2.0 permissive license allows commercial and private use with minimal restrictions, making it suitable for production deployments in research and industry settings.

Quickstart

pip install nccl4py[cu12]

import nccl4py
import numpy as np

# Initialize NCCL communicator for multi-GPU operations
comm = nccl4py.NcclComm()
data = np.array([1.0, 2.0, 3.0])
comm.AllReduce(data)

Requires NVIDIA CUDA 12.x or 13.x installed and NVIDIA GPU hardware; Linux (x86_64 or aarch64) only; Python 3.10 or later.

Verify before relying

• Whether nccl4py requires explicit NCCL library installation or bundles it with the wheel.
• Performance characteristics compared to direct NCCL C API or other Python distributed frameworks.
• Supported collective operations beyond AllReduce and their API signatures.

Package facts

Evidence: nccl4py-0.4.1-cp310-cp310-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl; nccl4py-0.4.1-cp310-cp310-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl; nccl4py-0.4.1-cp311-cp311-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl; nccl4py-0.4.1-cp311-cp311-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl; nccl4py-0.4.1-cp312-cp312-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl; nccl4py-0.4.1-cp312-cp312-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl; nccl4py-0.4.1-cp313-cp313-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl; nccl4py-0.4.1-cp313-cp313-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl; nccl4py-0.4.1-cp314-cp314-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl; nccl4py-0.4.1-cp314-cp314-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl; nccl4py-0.4.1-cp314-cp314t-manylinux_2_24_aarch64.manylinux_2_28_aarch64.whl; nccl4py-0.4.1-cp314-cp314t-manylinux_2_24_x86_64.manylinux_2_28_x86_64.whl

Tags

More Libraries packages

Further reading