sweepx

The sweep seismic wave-equation ecosystem — one install, then import sweep.

sweepx is the umbrella install for sweep — a GPU seismic wave-equation engine (forward modelling, RTM, FWI/LSRTM) — and its companion packages. It carries no code of its own: pip install sweepx pulls in the engine and the published companions so you can import sweep and go.

pip install sweepx

The ecosystem

Every piece is a separate, independently-installable package (the same layout as the PyLops family). sweepx bundles the published ones; you can also pip install any single package on its own.

Published — pulled in by sweepx

Companions — not on PyPI yet

Shipping as they mature; when a companion is published it is added to sweepx's dependencies, so you keep the same pip install sweepx.

Quick start

Drive the engine directly:

import sweep
from sweep.propagator.torch import PropTorch     # PyTorch propagator
from sweep import equations, propagator          # native submodules

…or in plain language, via the agent (needs a local LLM):

sweep-agent chat
>>> load the Marmousi benchmark model and run a forward — show the shot gather

Import name

After pip install sweepx, import as sweep, not sweepx — same pattern as pip install scikit-learn → import sklearn. The import name sweep was already taken on PyPI, so the installable is named sweepx.

CUDA backend

sweep's GPU backend (impl='c') ships prebuilt: the sweep-solver wheel carries one CUDA core per CUDA major -- cu12 (V100 through H100/H200, Blackwell via PTX) and cu13 (T4/RTX 20 and newer, Blackwell native, driver >= 580) -- and loads the one your torch's CUDA major names. After pip install nothing compiles: no nvcc, no C++ compiler, no CUDA headers, and the same wheel works with any torch version and any Python 3.

import sweep
sweep.precompile()   # optional: check that a CUDA core is in place and load it (compiles nothing)

An nvcc of your torch's CUDA major is needed only when no shipped core fits -- a torch built for another CUDA major, a GPU older than the shipped archs (e.g. Pascal), or an install from source -- and then only the core is built, once, locally. The pure-Python eager / JAX backends need none of this.

import sweep
print(sweep.is_torch_binding_available())          # torch + CUDA GPU + a usable core?
print(sweep.backend.torch.binding.diagnostics())   # usable / reason / shipped_core

License

MIT — see LICENSE.