Evaluate sundials4py (official LLNL bindings) to replace the Cython SUNDIALS wrapper

[davidcortesortuno]

Summary

Evaluate replacing our hand-written Cython SUNDIALS wrapper (fidimag/common/sundials/cvode.pyx, ~750 lines) — and possibly the bin/install-sundials.sh + local/ SUNDIALS build — with the official LLNL Python bindings, sundials4py.

Potential upside: it is first-party (authored by the SUNDIALS team), tracks SUNDIALS releases for us (already at 7.8.0 vs our 7.6.0 pin), BSD-3-Clause, and could delete a large chunk of maintenance surface.

Note on performance: our current cv_rhs already calls a Python RHS callback per step (unpacking the N_Vector into numpy views), so sundials4py's Python-callback model is unlikely to regress the hot path — the usual dealbreaker does not apply here.

Blockers to verify before adopting

Our integration leans on several advanced SUNDIALS features. Each must be confirmed in sundials4py:

• OpenMP N_Vector — we wrap state with N_VMake_OpenMP for parallel vector ops. Confirm nvector_openmp is exposed, not just the serial vector.
• Zero-copy callback arrays — our cv_rhs writes ydot in place via a zero-copy view. Confirm the callback receives writable, zero-copy numpy views (no per-step O(N) copy in/out).
• Matrix-free SPGMR + Jacobian-times-vector — we pass a jtimes_fun to a GMRES solver. Confirm support for the equivalent of CVodeSetJacTimes with a Python jtimes callback, plus reinit, tolerances (CVodeSStolerances), get_stats, and max-steps.
• Do the wheels bundle SUNDIALS? — decides whether this actually removes the build burden (bundled → we can drop install-sundials.sh + local/; not bundled → we still maintain the native install).
• Python floor — sundials4py requires Python ≥ 3.12; our pyproject.toml currently says >=3.9. Adopting it drops 3.9–3.11 (minor, since the uv branch targets 3.14).
• API completeness — verify everything else the wrapper uses is covered (CVODE reinit, rootfinding if used, stats, custom tolerances).

Suggested approach

• Prototype behind the existing Integrator abstraction (the CVODE wrapper is already isolated), keeping cvode.pyx as the fallback.
• Run the LLG + NEB test suites and a timing comparison against the current wrapper.
• Only remove the Cython wrapper / build scripts once all boxes above are checked.

References

• PyPI: https://pypi.org/project/sundials4py/ (v7.8.0, 2026-06-26, BSD-3-Clause)
• SUNDIALS docs: https://sundials.readthedocs.io/
• Current wrapper: fidimag/common/sundials/cvode.pyx

[davidcortesortuno]

Update — zero-copy callbacks are not a performance lever here.

Prototyped eliminating the RHS/jtimes output memcpy in the current Cython wrapper (cvode.pyx) by handing the callback a zero-copy numpy view onto the N_Vector. Correct, but no measurable speedup: the per-eval copies are ~0.6–3% of runtime (dominated by the effective-field computation — demag FFT / exchange stencils), the input copy is a required state-sync that can't be removed, and the view creation offsets the saved copy.

So the "zero-copy callback arrays" point above should be treated as a maintainability consideration only, not a speed one. Any sundials4py adoption should be decided on maintainability grounds.