NVIDIA cuSOLVER CUDA 11 Runtime Libraries

11.4.1.48 · active · verified Sat Apr 11

nvidia-cusolver-cu11 provides the native runtime libraries for NVIDIA's cuSOLVER, a high-performance GPU-accelerated library for dense and sparse direct linear solvers and eigenvalue problems. It is a fundamental component of the CUDA Toolkit, enabling accelerated numerical computations on NVIDIA GPUs. This package is intended for runtime use rather than direct development of GPU kernels, typically serving as a foundational dependency for higher-level Python libraries like PyTorch or CuPy in CUDA 11 environments. It is maintained by the Nvidia CUDA Installer Team and generally follows a slow release cadence for new versions.

Warnings

breaking Mixing CUDA versions (e.g., -cu11 and -cu12 packages) in the same Python environment can lead to 'undefined symbol' errors or other runtime failures. Ensure all `nvidia-*` packages, as well as dependent libraries like PyTorch or TensorFlow, are built for a consistent CUDA version.
Fix: Use a virtual environment and ensure all CUDA-dependent packages specify the same CUDA major version (e.g., all `*-cu11` or all `*-cu12`). For PyTorch, explicitly install the version for your target CUDA (e.g., `pip install torch==X.Y.Z+cu118`).
gotcha This package provides native runtime libraries, not a direct Python API. You cannot `import nvidia_cusolver_cu11` and call functions directly from it. Its functionality is exposed through higher-level libraries (e.g., NVIDIA Warp, nvmath-python, PyTorch, CuPy) that utilize the underlying cuSOLVER C/C++ libraries.
Fix: Integrate with frameworks or libraries that provide Python bindings to cuSOLVER functionalities, such as `torch.linalg` for PyTorch, `cupy.linalg` for CuPy, or the NVIDIA Warp library.
gotcha The `pip` installation only provides the runtime libraries. It does not install the NVIDIA GPU driver or the full CUDA Toolkit (which includes development headers and compilers). Your system must have a compatible NVIDIA GPU driver installed that supports the CUDA 11 version. Incompatible drivers will prevent the libraries from functioning.
Fix: Manually install the appropriate NVIDIA GPU driver for your hardware and operating system. Verify driver compatibility with your CUDA version using `nvidia-smi`.
gotcha This package is not supported on macOS or ARM-based systems (like Apple Silicon Macs), as CUDA itself is an NVIDIA GPU technology. Attempts to install or use it on such systems will result in installation errors or runtime failures.
Fix: Run this library only on systems with compatible NVIDIA GPUs and Linux or Windows operating systems.
gotcha The license for `nvidia-cusolver-cu11` is 'NVIDIA Proprietary Software'. This imposes restrictions on its use, modification, and redistribution that differ from open-source licenses.
Fix: Review the NVIDIA Proprietary Software License to understand its terms and conditions before using the library in commercial or restricted environments.

Install

pip install nvidia-cusolver-cu11 Install latest version

Quickstart

This library does not expose a direct Python API for import. Its primary role is to provide native shared libraries that are used by other GPU-accelerated Python packages (e.g., PyTorch, CuPy) or C++/CUDA applications with Python bindings. The quickstart demonstrates how to verify that the underlying CUDA environment, which includes cuSOLVER, is correctly set up and accessible by a common dependent library like PyTorch.

import torch

if torch.cuda.is_available():
    print(f"CUDA is available. Device name: {torch.cuda.get_device_name(0)}")
    # Further usage would involve libraries that depend on nvidia-cusolver-cu11,
    # such as PyTorch for linear algebra operations on GPU tensors.
    # Example (PyTorch uses cuSOLVER indirectly for operations like torch.linalg.solve, SVD, etc.):
    a = torch.randn(3, 3, device='cuda')
    b = torch.randn(3, 1, device='cuda')
    x = torch.linalg.solve(a, b)
    print("Solved linear system (A@x = b) on GPU:")
    print(x)
else:
    print("CUDA is not available. Please check your NVIDIA driver and CUDA installation.")

view raw JSON →