---
id: nccl4py
version: "0.4.1"
license: Apache-2.0
license_treatment: permissive
maintenance: active
---

# nccl4py — NCCL4Py: Python bindings for NCCL

License: permissive · Maintenance: active · Downloads: 323.5K/mo

## 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 above — 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

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

- License: Apache-2.0 (permissive)
- Python support: supports_current
- Install friction: medium
- Maintenance: active
- Downloads: 323.5K/month (top 15,000 on PyPI)
- Known vulnerabilities: none known

## Tags

gpu collective communication, nccl python bindings, multi-gpu distributed computing, nvidia nccl wrapper, gpu cluster communication, distributed training communication, cuda collective operations, gpu-computing, distributed-systems, nvidia-cuda

[View on SkillFed](https://skillfed.io/packages/nccl4py) · [View on PyPI](https://pypi.org/project/nccl4py/)