nvidia-nvshmem-cu12

What it is and what it does

NVSHMEM is NVIDIA's implementation of the OpenSHMEM parallel programming standard, adapted for GPU clusters. It creates a unified memory address space spanning multiple GPUs, allowing kernels and CPU code to read and write remote GPU memory with fine-grained control. The package is built on top of CUDA 12 and is intended for researchers and engineers building distributed machine learning and scientific computing workloads that need efficient inter-GPU communication.

The package is in Beta status and actively maintained. It targets modern Python versions (3.5 through 3.11) and runs on Linux (both aarch64 and x86_64 architectures). Installation requires the nvidia-cuda-cccl-cu12 runtime dependency, and the platform-specific wheels mean you must match your system architecture exactly.

Use it for:

• Implement collective operations (allreduce, broadcast) across GPUs in a cluster without explicit message passing.
• Build distributed training loops where multiple GPUs access shared parameter buffers via global memory semantics.
• Develop scientific simulations that require fine-grained synchronization and data movement between GPU memories.
• Optimize communication in multi-GPU inference pipelines by leveraging GPU-initiated remote memory access.

Worth the install?

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

NVSHMEM provides a global address space for GPU cluster communication, enabling fine-grained GPU and CPU-initiated operations across multiple GPU memories using OpenSHMEM-based primitives.

Yes, if you are building multi-GPU CUDA applications on Linux and need efficient inter-GPU communication. The package is actively maintained, has no known vulnerabilities, and targets a real use case in distributed GPU computing. However, verify the unclear license terms and confirm Windows support is not actually available despite classifier claims. Medium install friction is acceptable for a specialized GPU library.

Install

nvidia-nvshmem-cu12 on PyPI

pip install nvidia-nvshmem-cu12

uv add nvidia-nvshmem-cu12

poetry add nvidia-nvshmem-cu12

Installing nvidia-nvshmem-cu12

Before you install

Medium install friction due to platform-specific wheels (aarch64 and x86_64 Linux only). Active maintenance with a recent release, though no public repository or commit history is available to verify ongoing development velocity.

License in practice

License terms are unclear—no SPDX identifier or raw license text is published. Verify licensing compatibility with your project before production use.

Quickstart

pip install nvidia-nvshmem-cu12
import nvshmem
# Initialize NVSHMEM for multi-GPU communication
nvshmem.init()

Requires NVIDIA CUDA 12.x runtime, NVIDIA GPUs, and Linux (aarch64 or x86_64); not available on Windows or macOS despite classifier claims.

Verify before relying

• Actual Windows support status—classifiers list Windows but wheels are Linux-only.
• Whether public documentation exists beyond the NVIDIA CUDA Zone homepage.
• Specific CUDA 12 minor version requirements and GPU compute capability minimums.
• Whether the package works standalone or requires additional NVIDIA libraries beyond nvidia-cuda-cccl-cu12.

Package facts

Evidence: nvidia_nvshmem_cu12-3.7.2-py3-none-manylinux2014_aarch64.manylinux_2_17_aarch64.whl; nvidia_nvshmem_cu12-3.7.2-py3-none-manylinux2014_x86_64.manylinux_2_17_x86_64.whl

Keywords: cuda, nvidia, runtime, machine learning, deep learning

Tags

More Software Development packages

Further reading