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

Package	import	What it does
sweep-solver	sweep	Wave-equation engine: equations (acoustic / elastic / VTI / TTI / VRZ / SEM), propagators (torch / JAX / CUDA impl='c'), operators, boundary-saving, and FWI / LSRTM / RTM building blocks.
sweep-tasks	sweep_tasks	Production FWI / LSRTM runner — typed YAML task specs, a CLI, multi-GPU runs, field-data pipelines, with preprocessing (sweep_tasks.preproc) and plotting (sweep_tasks.viz) built in. Docs
sweep-loss	sweep_loss	Misfit / loss functions for FWI — two dozen, from L2 to optimal transport. Docs
sweep-io	sweep_io	Seismic I/O — SEG-Y, acquisition geometry, data and model plans. Docs
sweep-nn	sweep_nn	Neural reparameterizations and priors — INR / hash / SIREN encoders, TV and diffusion priors. Docs
sweep-agent	sweep_agent	Natural-language control layer — chat + files → sweep, through a local LLM (Ollama / vLLM). Runs forward modelling (acoustic and elastic), loads benchmark models, plots — all on the base install. pip install "sweep-agent[ui]" adds a web UI.

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.

Package	What it does
sweep-tomo	First-arrival traveltime tomography (eikonal + SIRT / FATT).
(planned)	sweep-opt.

sweep-viz and sweep-preproc are no longer separate packages: they are part of sweep-tasks, as sweep_tasks.viz and sweep_tasks.preproc.

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 a CUDA core for CUDA 12 (V100 and newer) and CUDA 13 (T4 and newer, driver >= 580) and loads the one matching your torch. Nothing compiles after pip install, with any torch version (Python >= 3.10 for sweepx); the eager / JAX backends are pure Python. Details are in the installation guide.

License

MIT — see LICENSE.