sparsax exposes SuiteSparse’s CHOLMOD
(Cholesky, for symmetric positive definite matrices) and KLU and UMFPACK (LU,
for general matrices) to JAX as XLA custom calls. A sparse factorization
therefore runs at native speed inside jax.jit, lax.scan, and
lax.fori_loop, with no Python callback and no round trip between device and
host on each iteration. An optional PyTensor frontend gives PyMC the same
CHOLMOD core, so its NUTS sampler can differentiate through the sparse solve and
log-determinant.