NVIDIA cuSOLVER (CUDA 12)

11.7.5.82 · active · verified Sun Mar 29

The `nvidia-cusolver-cu12` package provides the native runtime libraries for NVIDIA's cuSOLVER, a collection of GPU-accelerated dense and sparse direct linear solvers and Eigen solvers. It is part of the NVIDIA CUDA Toolkit and is designed to be consumed by higher-level Python libraries such as CuPy, PyTorch, and TensorFlow for dynamically loading CUDA libraries. The current version is 11.7.5.82, and its release cadence typically follows major and minor updates of the CUDA Toolkit.

Warnings

gotcha The `nvidia-cusolver-cu12` package primarily provides native shared libraries and is not intended for direct Python import and use of symbols. Its functionality is exposed indirectly through other high-level Python libraries like CuPy, PyTorch, or TensorFlow.
Fix: Use a Python library that integrates with cuSOLVER (e.g., CuPy, PyTorch) and ensure `nvidia-cusolver-cu12` is installed in the environment.
breaking The `cu12` suffix in the package name signifies compatibility with CUDA Toolkit 12.x. Installing this package with an incompatible CUDA Toolkit, GPU driver, or older GPU architecture can lead to runtime errors or performance issues.
Fix: Ensure your NVIDIA GPU, driver, CUDA Toolkit installation, and all `nvidia-*` Python packages (e.g., `nvidia-cublas-cu12`, `nvidia-cuda-runtime-cu12`) are consistent with the `cu12` version suffix. Major CUDA version changes (e.g., `cu11` to `cu12`) often require a complete reinstallation of all CUDA-related components and potentially recompilation of CUDA-dependent code.
gotcha Pip's dependency resolver can struggle with the complex interdependencies among `nvidia-*` packages, leading to slow resolution times or conflicts.
Fix: When installing multiple `nvidia-*` packages or libraries that depend on them, explicitly pin exact versions in your `requirements.txt` or install order. For example, `pip install 'nvidia-cusolver-cu12==11.7.5.82'`.
deprecated Certain cuSOLVER APIs, particularly within the `cuSOLVERMg` (multi-GPU) and some `cuSOLVERSp` (sparse) modules, have been deprecated in recent CUDA Toolkit versions.
Fix: Refer to the NVIDIA CUDA Toolkit documentation for your specific CUDA version to identify deprecated APIs and their recommended replacements. Ensure your application code is updated to use current cuSOLVER APIs.

Install

pip install nvidia-cusolver-cu12 Install latest stable version

Imports

N/A
```
This package provides native shared libraries (e.g., .so, .dll) for cuSOLVER.
```
The `nvidia-cusolver-cu12` package is not designed for direct Python import of symbols. Instead, it serves as a runtime dependency, providing the underlying CUDA binaries that other Python libraries (like CuPy, PyTorch, or TensorFlow) dynamically load and link against to offer CUDA-accelerated functionalities.

Quickstart

This quickstart demonstrates how a higher-level library, CuPy, leverages the underlying cuSOLVER libraries provided by `nvidia-cusolver-cu12`. When installed, `cupy.linalg.solve` and other linear algebra functions can utilize cuSOLVER for GPU-accelerated computations. Ensure CuPy is also installed (`pip install cupy-cuda12x`).

import cupy as cp

# This example uses CuPy, which dynamically loads cuSOLVER routines
# via packages like nvidia-cusolver-cu12 if available.

# Create a random positive definite matrix A and a vector b on the GPU
A = cp.random.rand(5, 5, dtype=cp.float64)
A = A @ A.T + cp.identity(5) # Make it symmetric positive definite
b = cp.random.rand(5, dtype=cp.float64)

# Solve the linear system Ax = b using CuPy's linalg.solve
# CuPy internally dispatches to cuSOLVER routines for this operation.
x = cp.linalg.solve(A, b)

print("Matrix A:\n", A)
print("Vector b:\n", b)
print("Solution x:\n", x)

# Verify the solution (A @ x - b should be close to zero)
print("Verification (A @ x - b):\n", A @ x - b)
print("Norm of residual (should be close to zero):", cp.linalg.norm(A @ x - b))

view raw JSON →