rlx-sparse

Sparse linear algebra for RLX — CSR LU factor, sparse mat-vec, Conjugate Gradient. Downstream package: registers against rlx's custom-op scaffold (rlx_ir::OpExtension + per-backend op_registry) without touching rlx-ir, rlx-opt, or rlx-runtime core.

This is the JAX-shaped pattern for domain-specific ops: build the extension surface in core, ship the actual op in a separate crate.

What's here

SparseTensor — CSR ({row_ptrs, col_indices, values}) + density / nnz / shape metadata.
Op::Custom("rlx_sparse.lu") — symbolic + numeric CSR LU.
Op::Custom("rlx_sparse.spmv") — sparse mat-vec.
Op::Custom("rlx_sparse.cg") — Conjugate Gradient solver.
register() — call once per process to publish the OpExtension (shape inference + autodiff) plus per-backend kernels (CPU always, Metal / MLX behind features).

Install

[dependencies]
rlx-sparse = "0.1"

For Apple GPU acceleration:

rlx-sparse = { version = "0.1", features = ["metal", "mlx"] }

Quickstart

rlx_sparse::register();   // once per process

let mut g = rlx_ir::Graph::new("sparse");
let csr = rlx_sparse::SparseTensor::from_dense(&dense, n, n);
let x = rlx_sparse::cg(&mut g, csr, b);
g.set_outputs(vec![x]);

Status

CG + SpMV production-ready on CPU. Metal / MLX kernels behind their respective features.

License