AMJax¶
AMJax converts algebraic multigrid hierarchies built by
PyAMG into JAX-compatible solvers and
preconditioners. PyAMG remains responsible for hierarchy construction; AMJax
focuses on repeated solves that can run under jax.jit, jax.vmap, GPU
execution, and automatic differentiation.
Installation¶
Where to start¶
- Use the quickstart for the PyAMG-to-AMJax solve pattern.
- Use benchmarks for the current recommendation and the latest generated benchmark summary.
- Use the API reference for AMJax's small public surface.
PyAMG Boundary¶
Use PyAMG documentation for AMG theory, hierarchy builders, and setup options:
- Classical solvers,
including
ruge_stuben_solver. - Aggregation solvers, including smoothed aggregation, root-node, pairwise, and AIR variants.
- PyAMG reference documentation for smoother, coarse-solver, and hierarchy construction details.
After MultilevelSolver.from_pyamg(...), AMJax converts the hierarchy into JAX
arrays and sparse BCOO matrices. Hierarchy setup still happens in Python, and
cycle, smoother, and coarse-solver choices should be treated as static JAX
configuration.