NVIDIA cuSPARSELt

0.9.0 · active · verified Thu Apr 09

NVIDIA cuSPARSELt is a Python binding to the NVIDIA cuSPARSELt library, providing highly optimized routines for sparse matrix-matrix multiplication (SpMM) on NVIDIA GPUs. This specific package, `nvidia-cusparselt-cu13`, targets CUDA Toolkit 13.x. It's a low-level library typically used for high-performance computing and deep learning workloads involving sparse data structures. Its release cadence is generally tied to major CUDA Toolkit updates.

Warnings

gotcha This package (`nvidia-cusparselt-cu13`) is specific to CUDA Toolkit 13.x. Using it with an incompatible CUDA Toolkit version on your system (e.g., CUDA 12.x) will likely result in runtime errors. Ensure your system's CUDA version matches the package suffix.
Fix: Install the `nvidia-cusparselt` package matching your system's CUDA version (e.g., `nvidia-cusparselt-cu12` for CUDA 12.x), or upgrade/downgrade your CUDA Toolkit.
gotcha Requires a compatible NVIDIA GPU and its corresponding drivers to be installed and properly configured on the system. Without a GPU, cuSPARSELt functions will fail.
Fix: Verify that an NVIDIA GPU is present, and the latest compatible NVIDIA drivers and CUDA Toolkit are installed and accessible to your Python environment.
breaking API stability is generally maintained by the `cuda-python` wrapper, but underlying changes in the C++ cuSPARSELt library between major CUDA Toolkit versions can lead to API breaking changes or behavioral differences, especially for advanced use cases.
Fix: Consult the official NVIDIA cuSPARSELt and `cuda-python` documentation for migration guides when upgrading to a new major CUDA Toolkit version. Test your code thoroughly after upgrading.
gotcha This library provides low-level bindings. For many deep learning applications, higher-level frameworks like PyTorch or TensorFlow integrate and utilize cuSPARSELt internally. Direct interaction is typically for advanced users optimizing specific sparse computations.
Fix: Evaluate if direct `cusparselt` usage is necessary. For common sparse operations, existing framework functions might suffice and abstract away the complexities of direct GPU programming.

Install

pip install nvidia-cusparselt-cu13 Install for CUDA 13.x

Imports

cusparselt
```
from cuda import cusparselt
```

Quickstart

This quickstart verifies the installation by attempting to retrieve the cuSPARSELt library version. It requires a compatible NVIDIA GPU and drivers to run successfully, as cuSPARSELt operates directly on the GPU.

from cuda import cusparselt
import sys

try:
    # Get cuSPARSELt properties (e.g., version)
    major, minor = cusparselt.cusparseltGetProperty()
    print(f"Successfully initialized cuSPARSELt. Version: {major}.{minor}")
except Exception as e:
    print(f"Error initializing cuSPARSELt: {e}", file=sys.stderr)
    print("Ensure an NVIDIA GPU is available and compatible CUDA drivers are installed.", file=sys.stderr)

view raw JSON →