JAXMg: A multi-GPU linear solver in JAX

JAXMg

JAXMg brings distributed matrix solvers to JAX, allowing calculations to run across multiple GPUs and nodes. This enables solvers to scale to matrices near the combined memory capacity of the available GPUs, far beyond native JAX routines, while retaining a familiar JAX interface.

JAXMg currently provides a jittable API for the following routines:

• potrs: Solves the system of linear equations $Ax=B$, where $A$ is an $N \times N$ symmetric or Hermitian positive-definite matrix, using a Cholesky decomposition. It can also return the log determinant of $A$.
• lu_solve: Solves the system of linear equations $Ax=B$, where $A$ is an $N \times N$ general nonsingular matrix, using a pivoted LU decomposition.
• syevd: Computes the eigenvalues $\lambda_i$ and optional eigenvectors $v_i$ of an $N \times N$ symmetric or Hermitian matrix $A$, satisfying $Av_i=\lambda_i v_i$.
• gesvd: Computes the singular-value decomposition of an $M \times N$ matrix $A$ ($A = U \Sigma V^{\dagger}$), returning the singular values and optional left and right singular vectors.

How JAXMg works

JAXMg connects JAX to NVIDIA's distributed cuSOLVERMp routines through a native C++/CUDA backend.

Supply a JAX matrix sharded over a one- or two-axis device mesh, and JAXMg handles the local memory-layout conversion, redistribution into cuSOLVERMp's 2D block-cyclic layout, distributed numerical computation, and restoration of the result to its original JAX layout. The matrix data remains GPU-resident throughout, with in-place transformations and bounded scratch storage minimizing memory overhead.

The operations are implemented using:

• potrs: cusolverMpPotrf and cusolverMpPotrs
• lu_solve: cusolverMpGetrf and cusolverMpGetrs
• syevd: cusolverMpSyevd
• gesvd: cusolverMpGesvd

For more details, see the API reference and the accompanying paper.

Deprecation notice: JAXMg 0.x, based on NVIDIA's deprecated cuSOLVERMg backend, has been superseded by JAXMg 1.0. Version 1.0 requires one Python process per GPU and removes potri; the previous API remains available with pip install "jaxmg<1". See the 0.0.9 release and 0.0.9 documentation.

Installation

Install the package with the extra matching your CUDA setup:

Prebuilt Linux wheels are provided for x86_64 and aarch64 and build for CUDA architecures sm_70, sm_80, sm_90, sm_120 and compute_90 . The supported NVIDIA GPU families are:

Note: pip install jaxmg installs CPU-only JAX. Select one of the CUDA extras above when installing JAXMg for GPU use. See Installation for details.

Example

JAXMg runs with one Python process per GPU. After launching one process for each GPU, initialize distributed JAX before constructing the device mesh. See Distributed execution for launch details.

For JAX arrays A and b, a Cholesky solve and log-determinant calculation requires only the distributed mesh, array placement, and solver call:

import jax
from jax.sharding import NamedSharding, PartitionSpec as P
from jaxmg import potrs


jax.distributed.initialize()

mesh = jax.make_mesh((jax.process_count(), 1), ("pr", "pc"))
jax.set_mesh(mesh)

A = jax.device_put(A, NamedSharding(mesh, P("pr", "pc")))
b = jax.device_put(b, NamedSharding(mesh, P("pr", None)))

x, logdet = potrs(A, b, T_A=256, return_logdet=True)

import jax
jax.config.update("jax_enable_x64", True)

import jax.numpy as jnp
from jax.sharding import PartitionSpec as P, NamedSharding
from jaxmg import potrs


jax.distributed.initialize()

T_A = 3
dtype = jnp.float64

num_procs = jax.process_count()
N = T_A * num_procs

A = jnp.diag(jnp.arange(N, dtype=dtype) + 1)
b = jnp.ones((N, 1), dtype=dtype)

mesh = jax.make_mesh((num_procs, 1), ("pr", "pc"))
jax.set_mesh(mesh)
a_sharding = NamedSharding(mesh, P("pr", "pc"))
b_sharding = NamedSharding(mesh, P("pr", None))
A = jax.device_put(A, a_sharding)
b = jax.device_put(b, b_sharding)

out, logdet = potrs(A, b, T_A=T_A, return_logdet=True)
out.block_until_ready()
logdet.block_until_ready()

expected_out = 1.0 / (jnp.arange(N, dtype=dtype) + 1)
expected_logdet = jnp.sum(jnp.log(jnp.arange(N, dtype=dtype) + 1))
is_correct = jnp.allclose(out.flatten(), expected_out) & jnp.allclose(
    logdet, expected_logdet
)
is_correct.block_until_ready()

if jax.process_index() == 0:
    print(out)
    print(logdet)
    print(is_correct)

[[1.        ]
 [0.5       ]
 [0.33333333]
 [0.25      ]
 [0.2       ]
 [0.16666667]
 [0.14285714]
 [0.125     ]
 [0.11111111]
 [0.1       ]
 [0.09090909]
 [0.08333333]]
19.987214495661885
True

Call potrs or lu_solve directly for standard solves. For use inside a larger jax.jit-compiled function, see the advanced Cholesky and LU examples.

Projects that use JAXMg

• JAXMg Benchmarks: Benchmarks for various Multi-GPUs setups.
• JAXMg + Netket: Implementation of the MinSR Netket driver that uses JAXMg for inverting the S-matrix. Tested on Multi-node settings.
• JAXMg for blurred sampling: Implementation of t-VMC that makes use JAXMg for inverting the QGT.

Citations

@misc{2601.14466,
Author = {Jacob Tutt and Roeland Wiersema},
Title = {JAXMg: A multi-GPU linear solver in JAX},
Year = {2026},
Eprint = {arXiv:2601.14466},
}

Acknowledgements

We acknowledge support from the Flatiron Institute. The Flatiron Institute is a division of the Simons Foundation.