NVIDIA NVSHMEM (CUDA 13)

3.6.5 · active · verified Thu Apr 09

nvidia-nvshmem-cu13 is a binary distribution package providing the NVIDIA NVSHMEM library, compiled specifically for CUDA Toolkit 13.x. NVSHMEM is a communication library that creates a global address space, enabling efficient and scalable data exchange across NVIDIA GPU clusters. This package itself does not contain Python-level bindings; instead, it serves as a foundational runtime dependency for other Python libraries (like `nvshmem-python`, Dask-CUDA, or CuPy) that utilize NVSHMEM's underlying C/C++ capabilities. The current version is 3.6.5.

Warnings

gotcha This `nvidia-nvshmem-cu13` package provides native NVSHMEM binaries for CUDA 13. It does NOT contain direct Python modules or APIs for import (e.g., `import nvidia-nvshmem-cu13` will fail). Its primary role is to act as a system-level runtime dependency for other Python libraries (e.g., `nvshmem-python`, Dask-CUDA, CuPy) that internally utilize NVSHMEM.
Fix: Do not attempt to directly import this package in Python. Install higher-level Python libraries (like `nvshmem-python`) to access NVSHMEM functionality via Python, ensuring this package is installed to provide the underlying native library.
gotcha CUDA Version Specificity: This package (`-cu13`) is built specifically for CUDA Toolkit 13.x. Ensure your system's CUDA installation and any other dependent libraries (e.g., PyTorch, TensorFlow, CuPy, Dask-CUDA) are compatible with CUDA 13.x. Mismatched CUDA versions can lead to runtime errors or crashes.
Fix: Verify your CUDA environment (`nvcc --version` or `nvidia-smi`) and ensure all related packages are compiled for the same CUDA major version (e.g., CUDA 13 for `nvidia-nvshmem-cu13`). If necessary, use a different `nvidia-nvshmem-cuXX` package corresponding to your CUDA version.
gotcha Requires MPI and Multi-GPU Environment: NVSHMEM is designed for high-performance communication across multiple GPUs, typically in a cluster environment. Meaningful usage often requires an MPI runtime (e.g., OpenMPI, MPICH) and multiple NVIDIA GPUs accessible to your application. Running NVSHMEM applications on a single GPU without MPI might still work but won't leverage its full potential and adds unnecessary complexity.
Fix: Set up a multi-GPU environment with a compatible MPI implementation (e.g., `mpiexec -n 2 python your_script.py`). Refer to NVSHMEM documentation for proper setup and execution of distributed applications.

Install

pip install nvidia-nvshmem-cu13 Install for CUDA 13

Imports

No direct Python imports
```
N/A
```
This package provides native NVSHMEM binaries for CUDA 13; it does not contain direct Python modules for import (`import nvidia-nvshmem-cu13` will fail). Python applications requiring NVSHMEM functionality typically use higher-level libraries (e.g., `nvshmem-python`, Dask-CUDA, CuPy) that link against these native libraries.

Quickstart

This quickstart clarifies that `nvidia-nvshmem-cu13` is a binary distribution and does not offer direct Python APIs. Its primary function is to satisfy the underlying native library dependency for other Python libraries that enable NVSHMEM functionality.

# This package (nvidia-nvshmem-cu13) provides native NVSHMEM binaries for CUDA 13.
# It does not expose direct Python modules for import or direct usage.
# Its role is to serve as a runtime dependency for other Python libraries
# (e.g., nvshmem-python, Dask-CUDA, CuPy) that internally leverage NVSHMEM
# for high-performance GPU-aware communication.

print("`nvidia-nvshmem-cu13` is installed as a system-level dependency.")
print("Python applications will use other libraries (e.g., `nvshmem-python`)")
print("that link against the NVSHMEM binaries provided by this package.")
print("\nTo truly 'use' NVSHMEM, you would typically write code using a library like `nvshmem-python`: ")
print("e.g., `pip install nvshmem-python` and then `import nvshmem; nvshmem.init()`")

# This quickstart does not include actual NVSHMEM operations because
# this specific package does not provide a direct Python API.

view raw JSON →