PMFGLM.jl

Primordial magnetic fields on a single node

Modified

October 24, 2026

Single-node primordial magnetic-field simulations with full GLM magnetohydrodynamics, particle gravity, photon-baryon response, and primordial chemistry. Julia and KernelAbstractions share the numerical kernels across parallel CPU, Apple Metal, and NVIDIA CUDA. The package bundles its source dependencies; no Vespa checkout, private registry, or sibling science repository is required.

0.1 Start here

git clone https://github.com/yipihey/pmf-glm.git
cd pmf-glm
julia --project=. -t auto scripts/bootstrap.jl --backend=cpu
julia --project=. -t auto scripts/run_pmf.jl examples/smoke.toml \
  --output=results/first --backend=cpu

The installation guide targets a first small CPU run in about five minutes after Julia is installed; dependency downloads and first compilation can take longer. For Apple Silicon or NVIDIA, follow the separate GPU setup before changing the backend.

Your question	Read
Can I run this on my machine?	Installation, backends
What equations and units are evolved?	Physics, notation
How do I preserve a PMF realization between resolutions?	Initial conditions
What does a run write, and how do I restart it?	Outputs
What has actually passed here?	Recorded test results
How can I make a defensible science comparison?	Validation protocol
Where should I change a kernel or add a test?	Development, architecture

0.2 Scope

The research runner retains nonlinear finite-volume MHD throughout radiation coupling. The photon model evolves a monopole and velocity with an algebraic stress closure: it is not a full polarized Boltzmann hierarchy. Gaussian Lyman-alpha mixing and response-table chemistry are model approximations, not three-dimensional line radiative transfer. See the science boundary.

Evidence belongs to a revision

Inherited tests, historical Metal runs, and Schiff comparison tools describe lineage, not certification of this extraction. CUDA support in the source does not establish CUDA hardware validation. Only the evidence page reports measurements supplied with this documentation build.

0.3 Package surface

The public facade exposes configuration and run orchestration, plus selected MHD operations. The bundled MHDKernels, PoissonKernels, ChemistryKernels, and EmissionKernels remain accessible through PMFGLM. ^{module PMFGLM}

using PMFGLM
config = load_config("examples/cosmology-smoke.toml")
resolved = run_environment(config; output_dir="results/inspect")
run_simulation(config; output_dir="results/cosmology-first")

Use a new directory per experiment. The runner records the resolved options, source fingerprint, dependency-manifest hash, execution log, and completion status. A completed short run is an installation check, not a convergence result.

--- title: "PMFGLM.jl" --- Single-node primordial magnetic-field simulations with full GLM magnetohydrodynamics, particle gravity, photon-baryon response, and primordial chemistry. Julia and KernelAbstractions share the numerical kernels across parallel CPU, Apple Metal, and NVIDIA CUDA. The package bundles its source dependencies; no Vespa checkout, private registry, or sibling science repository is required. ## Start here {#sec-start} ```sh git clone https://github.com/yipihey/pmf-glm.git cd pmf-glm julia --project=. -t auto scripts/bootstrap.jl --backend=cpu julia --project=. -t auto scripts/run_pmf.jl examples/smoke.toml \ --output=results/first --backend=cpu ``` The [installation guide](getting-started.qmd) targets a first small CPU run in about five minutes after Julia is installed; dependency downloads and first compilation can take longer. For Apple Silicon or NVIDIA, follow the separate [GPU setup](backends.qmd) before changing the backend. | Your question | Read | |---|---| | Can I run this on my machine? | [Installation](getting-started.qmd), [backends](backends.qmd) | | What equations and units are evolved? | [Physics](physics.qmd), [notation](notation.qmd) | | How do I preserve a PMF realization between resolutions? | [Initial conditions](initial-conditions.qmd) | | What does a run write, and how do I restart it? | [Outputs](outputs.qmd) | | What has actually passed here? | [Recorded test results](test-results.qmd) | | How can I make a defensible science comparison? | [Validation protocol](validation.qmd) | | Where should I change a kernel or add a test? | [Development](development.qmd), [architecture](architecture.qmd) | ## Scope {#sec-scope} The research runner retains nonlinear finite-volume MHD throughout radiation coupling. The photon model evolves a monopole and velocity with an algebraic stress closure: it is **not a full polarized Boltzmann hierarchy**. Gaussian Lyman-alpha mixing and response-table chemistry are model approximations, not three-dimensional line radiative transfer. See the [science boundary](validation.qmd#sec-limitations). ::: {.callout-warning} ## Evidence belongs to a revision Inherited tests, historical Metal runs, and Schiff comparison tools describe lineage, not certification of this extraction. CUDA support in the source does not establish CUDA hardware validation. Only the [evidence page](test-results.qmd) reports measurements supplied with this documentation build. ::: ## Package surface {#sec-package-surface} The public facade exposes configuration and run orchestration, plus selected MHD operations. The bundled `MHDKernels`, `PoissonKernels`, `ChemistryKernels`, and `EmissionKernels` remain accessible through `PMFGLM`. {{< src "src/PMFGLM.jl" "module PMFGLM" >}} ```julia using PMFGLM config = load_config("examples/cosmology-smoke.toml") resolved = run_environment(config; output_dir="results/inspect") run_simulation(config; output_dir="results/cosmology-first") ``` Use a new directory per experiment. The runner records the resolved options, source fingerprint, dependency-manifest hash, execution log, and completion status. A completed short run is an installation check, not a convergence result.