Manage Julia

Inference runs in Julia, and MCMC.js can install and manage it for you. This guide covers provisioning the toolchain, checking it, managing Julia versions, and the persistent worker daemon.

Provision and check

mcmc setup installs the Julia toolchain (juliaup and Julia) and precompiles the managed project.

mcmc setup            # install and precompile
mcmc setup --dry-run  # show what would be installed, change nothing

mcmc doctor reports the toolchain MCMC.js needs and whether you are ready to fit. mcmc engines lists the inference engines the CLI knows about and their capabilities.

mcmc doctor
mcmc engines

julia    ready  setup, versions, fit, predict

Managing Julia versions

mcmc julia version wraps juliaup channels.

Command	What it does
mcmc julia version list	list installed Julia versions
mcmc julia version status	installed versions plus the juliaup and Julia toolchain
mcmc julia version add <version>	install a version or channel, e.g. 1.10 or release (--default to also set it)
mcmc julia version remove <version>	uninstall a version or channel
mcmc julia version default <version>	set the default version or channel
mcmc julia version update [version]	update one channel, or all installed channels
mcmc julia version gc	reclaim disk from uninstalled Julia versions

mcmc julia version add 1.11 --default
mcmc julia version status

juliaup  1.20.7  /home/you/.juliaup/bin/juliaup
julia    1.12.6  /home/you/.juliaup/bin/julia

* release    1.12.6     .../julia-1.12.6/bin/julia
  1.11       1.11.9     .../julia-1.11.9/bin/julia

All of these accept --json, and the mutating commands accept --verbose for the raw juliaup output.

Running on a chosen version

A fit runs on the Julia channel pinned in the spec’s [backend].version (a specific version, not a moving channel, so a run reproduces the resolved package set). Override it per command with --julia-version:

mcmc run model.jl --julia-version 1.11

Version matrices

mcmc fit can run one spec across several Julia versions, or several versions of one managed package, each in its own isolated managed environment:

mcmc fit model.toml --versions 1.11,1.12 -o out/
mcmc fit model.toml --package-versions Turing=0.44,0.45 -o out/

See Run inference for the matrix flags.

Persistent workers (daemon)

Each fit normally starts a fresh Julia process, paying Julia’s startup and load cost every time. A persistent worker keeps a Julia process warm and runs fits through it.

Start one by adding --daemon to a fit (or set MCMC_DAEMON=1):

mcmc run model.jl --daemon

Manage the workers with the daemon group:

mcmc daemon status   # list known workers and whether they answer
mcmc daemon stop     # stop all workers and remove their sockets

no workers; start one with `mcmc run <model> --daemon`

Both subcommands accept --json.