cmdlaplacer

A thin R wrapper around the Laplace compiler and cmdstanr, so you can go from a .laplace source file to a running cmdstanr model without touching the terminal.

Why

Laplace compiles .laplace files to plain, inspectable Stan (.stan) source — that output is always the source of truth and never depends on Laplace at runtime. This package doesn't change that. It just removes the manual step of running laplace build yourself and pointing cmdstanr::cmdstan_model() at the result.

• Default path: give it a .laplace file, get back a ready-to-use CmdStanModel object.
• Escape hatch: if you'd rather handle the Stan file yourself, ask for the path and do your own thing from there.

Nothing about compilation, caching, or diagnostics happens in R — that's all delegated to the laplace binary. This package is intentionally small.

Installation

cmdlaplacer requires two things to be installed separately first:

1. The laplace CLI, available on your system PATH.
2. cmdstanr, plus a working CmdStan installation. See the cmdstanr installation guide.

Then install the package itself:

# install.packages("remotes")
remotes::install_github("mlatinov/cmdlaplacer")

Installing the laplace CLI

laplace is built from source with Cargo. If you already have git and a Rust toolchain (cargo) on your PATH, you can install it directly from R:

library(cmdlaplacer)
laplace_install()

This clones the Laplace repository, builds it with cargo install --path . --root ~/.local, and adds ~/.local/bin to PATH for the current session. It's the same thing you'd get from running those git/cargo commands by hand — see the manual installation instructions if you'd rather do that yourself, or need to install git/Rust first.

You don't need to call laplace_install() yourself, though: if laplace_model() can't find laplace on your PATH in an interactive session, it will ask whether to install it for you before failing.

Usage

library(cmdlaplacer)

# Compile a .laplace file and load it directly as a cmdstanr model
mod <- laplace_model("models/linreg.laplace")
fit <- mod$sample(data = list(N = 10, y = rnorm(10)))

If you'd rather work with the compiled Stan file directly — to inspect it, version it, or drive cmdstanr yourself — pass stan_only = TRUE:

stan_path <- laplace_model("models/linreg.laplace", stan_only = TRUE)

# now do whatever you'd normally do with a .stan file
mod <- cmdstanr::cmdstan_model(stan_path)

laplace_model() runs laplace build inside the .laplace file's own directory, so it picks up that project's laplace.lock no matter what R's working directory is.

Adding a library from R

laplace_install_git() is laplace add --git from R. Like the CLI, it creates laplace.toml and laplace.lock if the project doesn't have them yet:

laplace_install_git(
  "transformations",
  "https://github.com/mlatinov/laplace-transform",
  tag = "0.1.0",
  subdir = "laplace",
  project_dir = "models"
)

Arguments

Argument	Description
path	Path to the .laplace source file to build.
stan_only	If TRUE, return the path to the compiled .stan file instead of a cmdstanr model object. Default FALSE.
out_dir	Where to write the compiled .stan file. Defaults to build/ next to the .laplace file — the same place laplace build writes to, so the R and terminal routes produce the same file.
split_functions	If TRUE, build with --split-functions (one <pkg>.stanfunctions file per imported package) and pass out_dir to cmdstanr as an include path. Default FALSE.
...	Passed straight through to cmdstanr::cmdstan_model() (e.g. compile = FALSE).

What this package does not do

• It does not reimplement or wrap cmdstanr's fitting API ($sample(), $optimize(), etc.) — use cmdstanr directly on the returned model object.
• It does not cache builds in R — laplace build is responsible for deciding whether a rebuild is needed.
• It does not hide the compiled Stan source — the .stan file is always written to disk and can be inspected, diffed, or committed independently of this package.

Error handling

If the laplace CLI isn't found on your PATH:

• In an interactive session, laplace_model() asks whether to install it via laplace_install() before failing.
• In a non-interactive session (scripts, Rscript, CI), it never prompts — it just raises an informative error pointing at laplace_install() or the manual installation instructions.

If laplace build fails, or if cmdstanr isn't installed and you haven't set stan_only = TRUE, you also get a clear, specific error rather than a raw system-call failure or package-loading error.

License

MIT