sweepx 0.4.1

sweepx 0.4.1 follows sweep-solver 0.3.6 and requires it (sweep-solver>=0.3.6,<0.4).

pip install -U sweepx

Upgrading from 0.4.0

• The LSRTM equations follow sweep-solver 0.3.6's rename to Born (DeepWave-KAUST/sweep#135, no aliases): a task YAML that names equation: AcousticLSRTM must name AcousticBorn. The LSRTM template, example, schemas and plotting use AcousticBorn.
• Topography follows sweep-solver 0.3.6: under physics.topography a line geometry's depth now counts down from the surface of each column (z = surface[x] + depth), and the agent places its explicit points the same way. The topography snapshot view (sweepx.tasks.viz.wavefield, make_wavefield_gif) masks the air by default instead of resampling.
• ElasticAPM is gone from the equation lists and from introspection, as it is from the solver.
• sweepx 0.4.0 also accepts sweep-solver 0.3.6, but its LSRTM task does not run there (the equation was renamed) and its topography line depths are read differently, so upgrade the two together.

Fixed

• The agent demos work with a real LLM: a 32k context and server errors shown; absolute output paths; an unknown tool argument is an error; shot gathers on the receivers' x axis; the elastic wavefield tools use a vertical force, so S waves appear; run_fwi and build_fwi_spec use every shot and bound vp; tools pick the GPU and the CUDA kernels by default (device / backend_impl 'auto'); an answer that claims a figure no tool wrote is flagged.
• Tasks: optimizer.kind: lbfgs is sweepx.opt's L-BFGS (initial_step, refused on shot batches), and a non-finite loss fails the run.

Docs

• What's new at https://sweepx.deepwave.group/: five sweep-solver 0.3.6 cards (spectral elements, Biot, elastic topography, qP on the CUDA core, elastic Born) and the Viking frequency-selection walkthrough. The task runner and the agent do not drive sweep.sem yet.

The merge is #13. The files below are the ones on PyPI.

sweepx 0.4.0

sweepx 0.4.0 makes the toolkit one package: io, loss, nn, opt, tasks and agent now ship together as sweepx.<module>. The engine stays its own package, sweep-solver (import sweep), and pip install sweepx brings it along.

pip install -U sweepx

Upgrading from 0.3.x

• Imports move: sweep_tasks → sweepx.tasks, sweep_loss → sweepx.loss, sweep_io → sweepx.io, sweep_nn → sweepx.nn, sweep_agent → sweepx.agent. There is no alias for the old names.
• The sweep-tasks and sweep-agent commands are unchanged; sweepx tasks … and sweepx agent … reach the same CLIs.
• The separate packages (sweep-tasks, sweep-io, sweep-loss, sweep-nn, sweep-agent) get no further releases. Once your scripts import sweepx.<module>, you can uninstall them.

New

• sweepx.opt: FWI optimizers behind the torch.optim interface (L-BFGS, L-BFGS-B, nonlinear CG, truncated Newton and Gauss-Newton, trust region, Adam, Lion, SAM, Shampoo, SOAP), with pseudo-Hessian and depth-weighting preconditioners and bound and DIP-FWI parameterizations.
• Optional extras: segy, hdf5, animate, agent-ui, mcp, vllm, all.
• Docs at https://sweepx.deepwave.group/, with the six modules under one sweepx tab.

The merge is #7. The files below are the ones on PyPI.