Installation

From PyPI (recommended)

sweep works with any PyTorch version, but the PyTorch build has to match your NVIDIA driver. A plain pip install sweepx pulls PyPI's default torch, which on Linux is built for CUDA 13 and needs driver >= 580. With an older driver torch cannot start CUDA (UserWarning: ... The NVIDIA driver on your system is too old) and impl='c' reports no visible GPU. Install the matching torch first, then sweep:

Package

sweepx (solver + agent) sweep-solver (solver only)

Installer

pip uv

Your GPU

V100 T4 · A100 · H100 · RTX 20–40 B200 · RTX 50 (Blackwell) No NVIDIA GPU

Compute Platform

Auto CUDA 12.1 CUDA 12.4 CUDA 12.6 CUDA 12.8 CUDA 12.9 CUDA 13.0 CUDA 13.2 CPU

Run this Command

• Which CUDA? nvidia-smi shows the newest CUDA Version your driver supports; pick that or lower. Combinations your GPU cannot run are greyed out.
• Older CUDA 12 builds come with an older torch (cu121: 2.5, cu124: 2.6, cu128: 2.11, cu129: 2.13). sweep runs on all of them. CUDA 11 supports impl='eager' only.
• Wrong torch already installed? Reinstall just torch with the command above plus --force-reinstall; sweep stays as is.

Note

sweepx (Python >= 3.10) installs the solver plus the sweep-agent companion; pip install sweep-solver installs the solver alone (Python >= 3.9). Either way you import sweep: the bare name sweep is taken on PyPI.

From source

Install torch first, as above, then:

git clone https://github.com/DeepWave-KAUST/sweep
cd sweep
pip install .
python -m sweep.build   # optional: build the CUDA core now, not on the first impl='c' call

A clone carries no prebuilt core: the CUDA core is compiled once for your card (2–5 min, needs an nvcc of your torch's CUDA major) and cached.

Verification

Run a small forward model:

import numpy as np, torch, sweep
from sweep.equations import Acoustic
from sweep.propagator.torch import PropTorch
from sweep.signal import ricker

dev = torch.device("cuda" if torch.cuda.is_available() else "cpu")
solver = PropTorch(Acoustic(device=dev), shape=(64, 64), dh=10.0, dt=1e-3, dev=dev)
wavelet = ricker(np.arange(500) * 1e-3 - 0.1, f=15.0).astype(np.float32)
gather = solver(wavelet, np.array([[32, 4]]), np.array([[[ix, 4] for ix in range(64)]]),
                models=[torch.full((64, 64), 2000.0, device=dev)])

print("sweep", sweep.__version__, "| torch", torch.__version__)
print("device:", dev, "| backend:", solver.impl)
print("gather:", tuple(gather.shape), "| finite:", bool(torch.isfinite(gather).all()))

On a GPU the output should look like:

sweep 0.3.1 | torch 2.13.0+cu129
device: cuda | backend: c
gather: (1, 500, 64, 1) | finite: True

backend: c means the compiled CUDA backend is working. If you see device: cpu on a GPU machine, torch cannot use your GPU: reinstall it with the selector above. If you see device: cuda | backend: eager, run python -c "import sweep; sweep.precompile()": its error says why impl='c' is unavailable. Without a GPU, backend: eager is expected. With g++ older than 10, torch.compile cannot build the CPU step: sweep warns and runs it uncompiled (slower, same result).

How the core is picked, when nvcc is needed, building without a GPU and developer builds: Building the CUDA core.