nccl4py 0.3.1


pip install nccl4py

  Latest version

Released: Jun 11, 2026
Navigation
Meta
Author: NVIDIA Corporation
Requires Python: >=3.10
Classifiers

Intended Audience
  • Developers
  • Science/Research
  • End Users/Desktop

Natural Language
  • English

Operating System
  • POSIX :: Linux

Topic
  • Education
  • Scientific/Engineering
  • Software Development :: Libraries

Programming Language
  • Cython
  • Python :: 3 :: Only
  • Python :: 3.10
  • Python :: 3.11
  • Python :: 3.12
  • Python :: 3.13
  • Python :: 3.14
  • Python :: Implementation :: CPython

Environment
  • GPU :: NVIDIA CUDA
  • GPU :: NVIDIA CUDA :: 12
  • GPU :: NVIDIA CUDA :: 13
NVIDIA

nccl4py bridges Python’s simplicity with the performance of NVIDIA Collective Communications Library (NCCL), and provides a Pythonic interface to NCCL library’s functionality. It enables Python applications to leverage NCCL’s GPU-accelerated multi-GPU and multi-node communication capabilities for distributed computing workloads.

nccl4py follows the NCCL SLA. The details of the NCCL SLA is available here.

nccl4py is under active development. Feedback and suggestions are welcome!

Installation

For CUDA 12.x:

pip install nccl4py[cu12]

For CUDA 13.x:

pip install nccl4py[cu13]
0.3.1 Jun 11, 2026
0.2.0 Apr 24, 2026
0.1.1 Jan 09, 2026

Wheel compatibility matrix

Platform CPython 3.10 CPython 3.11 CPython 3.12 CPython 3.13 CPython 3.14 CPython (additional flags: t) 3.14
manylinux_2_24_aarch64
manylinux_2_24_x86_64
manylinux_2_28_aarch64
manylinux_2_28_x86_64

Files in release

Extras:
Dependencies:
packaging
numpy
cuda.core (~=1.0)
cuda-pathfinder (<2.0.0,>=1.5.4)
typing_extensions