JaxPM

JAX-powered Cosmological Particle-Mesh N-body Solver

Note

The new JaxPM v0.1.xx supports multi-GPU model distribution while remaining compatible with previous releases. These significant changes are still under development and testing, so please report any issues you encounter. For the older but more stable version, install:
pip install jaxpm==0.0.2

Install

Basic installation can be done using pip:

pip install jaxpm

For more advanced installation for optimized distribution on gpu clusters, please install jaxDecomp first. See instructions here.

Goals

Provide a modern infrastructure to support differentiable PM N-body simulations using JAX:

Keep implementation simple and readable, in pure NumPy API
Any order forward and backward automatic differentiation
Support automated batching using vmap
Compatibility with external optimizer libraries like optax
Now fully distributable on multi-GPU and multi-node systems using jaxDecomp working withJAX v0.4.35

Open development and use

Current expectations are:

This project is and will remain open source, and usable without any restrictions for any purposes
Will be a simple publication on The Journal of Open Source Software
Everyone is welcome to contribute, and can join the JOSS publication (until it is submitted to the journal).
Anyone (including main contributors) can use this code as a framework to build and publish their own applications, with no expectation that they need to extend authorship to all jaxpm developers.

Getting Started

To dive into JaxPM’s capabilities, please explore the notebook section for detailed tutorials and examples on various setups, from single-device simulations to multi-host configurations. You can find the notebooks' README here for a structured guide through each tutorial.