cudass

High-performance CUDA sparse linear solver supporting multiple matrix types, cuDSS, cuSOLVER Dense, and cuSolverSp backends.

Docs: https://moca-technion.github.io/cudass/

Requirements

• PyTorch >= 2.0 with CUDA (from pytorch.org).
• nvidia-cudss-cu12 or nvidia-cudss-cu13 (>=0.6), chosen from torch.version.cuda (override: CUDASS_CUDA_MAJOR=12 or 13). Install or let the build install it.
• NVIDIA GPU and a driver compatible with your PyTorch CUDA (12.x or 13.x).

Installation

From PyPI (recommended)

pip install cudass

Builds from the sdist at install time. Requires nvcc (CUDA toolkit) on your system; the build pulls in Cython and torch, and picks nvidia-cudss-cu12 or cu13 from your PyTorch’s CUDA.

From source

pip install torch Cython   # nvidia-cudss-cu12 or cu13 is chosen from torch.version.cuda
# nvcc (CUDA toolkit): set CUDA_HOME or e.g. module load cuda/12.4 or cuda/13.0
pip install -e .
# Override: CUDASS_CUDA_MAJOR=13 pip install -e .  for CUDA 13

Quick Start

from cudass import CUDASparseSolver, MatrixType

solver = CUDASparseSolver(matrix_type=MatrixType.SPD, use_cache=True)
solver.update_matrix((index, value, m, n))
x = solver.solve(b)

License

Apache 2.0. See LICENSE.

This project uses third-party packages (e.g. PyTorch, NVIDIA cuDSS, Cython) and optional libraries (e.g. nvmath). See NOTICE for attributions and a pointer to their licenses. Your use of those components is subject to their terms.