PACKAGE · import sweep

sweep-solver

The differentiable wave-equation engine. Nine equation families — acoustic, elastic, VTI, TTI, VRZ, SEM — on torch / JAX / native-CUDA backends, with FWI, LSRTM and RTM building blocks.

Open the solver docs →

PACKAGE · import sweep_agent

sweep-agent

The natural-language control layer. Say what you want and a local LLM (Ollama / vLLM) turns it into a validated sweep run — forward modelling, benchmark models, plots.

Open the agent docs →

COMPANIONS · GROWING

…and the family

sweep-tasks (production FWI/LSRTM runner), sweep-loss, sweep-nn, sweep-tomo — each an independent package, added to sweepx as it ships.

Browse the org →

WHAT'S NEW

Recently shipped.

The latest user-visible additions across the stack. The badge says where each one lives: a released version you get from pip install, or dev for what is merged but not yet on PyPI — these docs are built from the development branch.

v0.3.32026-10

Attenuation, acoustic and elastic

Near-constant Q in 2-D with gradients for every model. New in 0.3.3: ViscoElastic, the generalized standard linear solid SPECFEM2D uses, with Qp and Qs on both backends. It joins the visco-acoustic form of Zhu & Harris (2014).

v0.3.22026-10

Install for your GPU and driver

Pick your GPU and driver, copy the matching torch + sweep command, and confirm it with a 15-line run. 0.3.2 also runs eager on any g++, and a wrong pml_type is now a clear error instead of a crash.

v0.3.02026-09

Prebuilt CUDA core

The wheel carries the compiled kernels, so impl='c' runs straight after pip install — no nvcc, no first-use build, one wheel for any PyTorch version.

v0.3.02026-08

Elastic TTI in 3-D

A tilted symmetry axis in three dimensions, where the azimuth finally matters — plus the Oh-2020 displacement form on (vh, η), held stable by a multiaxial PML. CUDA gradients for every parameter.

v0.3.02026-09

Domain decomposition, faster

A model too big for one GPU splits across several — one tile per rank, a halo per step, and a gradient bit-identical to the single-GPU one. 0.3.0 made it faster: up to 3.99× on 2-D elastic and 1.56× on 3-D elastic.

agent 0.0.32026-08

Natural-language control

Natural-language control on a plain pip install: forward modelling (acoustic and elastic) plus benchmark models, no sweep-tasks needed.

ONE INSTALL · MANY PACKAGES

One install.
Then import sweep.

sweepx is the umbrella — it carries no code of its own, it just pulls in the engine and its published companions. Install sweepx, but import sweep (same pattern as scikit-learn → sklearn).

# one install — pulls the engine + the agent:
#     pip install sweepx

import sweep                                       # ← sweep-solver, the engine
from sweep.propagator.torch import PropTorch       # torch / JAX / native-CUDA
from sweep.equations import Acoustic, ElasticTTI   # nine equation families

# …or drive it all in plain language — sweep-agent:
#     sweep-agent chat
#     >>> load Marmousi and run a forward — show the shot gather

Install & get started →

THE FAMILY

Every piece, installable on its own.

The same layout as the PyLops family — sweepx bundles the published packages; you can also pip install any single one.

PUBLISHED

sweep-solver

import sweep

Wave-equation engine — equations, propagators, FWI/LSRTM/RTM.

PUBLISHED

sweep-agent

import sweep_agent

Natural-language control via a local LLM.

COMPANION

sweep-tasks

coming

Production FWI/LSRTM runner — specs, YAML, multi-GPU, IO.

COMPANION

sweep-loss

coming

Misfit / loss functions for inversion.

COMPANION

sweep-nn

coming

Neural reparameterizations — INR / hash / SIREN encoders.

COMPANION

sweep-tomo

coming

First-arrival traveltime tomography — eikonal + SIRT / FATT.