Using {cmdstanr} with {rixpress}

Setting up the Environment

As with any {rixpress} pipeline, the first step is to define the execution environment using {rix}:

library(rix)

rix(
  date = "2025-04-29",
  r_pkgs = c("readr", "dplyr", "ggplot2"), # Add other R packages as needed
  system_pkgs = "cmdstan", # Crucial: include cmdstan as a system dependency
  git_pkgs = list(
    list(
      package_name = "cmdstanr",
      repo_url = "https://github.com/stan-dev/cmdstanr",
      commit = "79d37792d8e4ffcf3cf721b8d7ee4316a1234b0c" # Pin to a specific commit
    ),
    list(
      package_name = "rixpress",
      repo_url = "https://github.com/ropensci/rixpress",
      commit = "HEAD" # Or pin to a specific commit
    )
  ),
  ide = "none", # Or your preferred IDE
  project_path = ".",
  overwrite = TRUE
)

Key points in this environment definition:

• cmdstan is included in system_pkgs. This makes the cmdstan executables available to the pipeline.
• {cmdstanr} is installed from its GitHub repository, as it’s not available on CRAN. Pinning to a specific commit is recommended for maximum reproducibility.

With the environment set up, we can define the pipeline:

Setting up the pipeline

The Stan model code itself should reside in a .stan file. We use rxp_r_file() to bring its contents into the pipeline as a character string.

rxp_r_file(
  bayesian_linear_regression_model,
  "model.stan",
  readLines
)

Next, we define parameters and simulate some data for our model.

  rxp_r(
    parameters,
    list(
      N = 100,
      alpha = 2,
      beta = -0.5,
      sigma = 1.e-1
    )
  ),
  rxp_r(
    x,
    rnorm(parameters$N, 0, 1)
  ),
  rxp_r(
    y,
    rnorm(
      n = parameters$N,
      mean = parameters$alpha + parameters$beta * x,
      sd = parameters$sigma
    )
  ),
  rxp_r(
    # Prepare the data list for cmdstanr
    inputs,
    list(N = parameters$N, x = x, y = y)
  ),

Compiling and Sampling the Model

Interfacing with cmdstan from within {rixpress} requires a specific strategy due to the hermetic nature of Nix sandboxes. We’ll use a wrapper function to handle model compilation and sampling within a single rxp_r() step.

First, let’s define the wrapper function (e.g., in a functions.R file that we’ll include):

# In functions.R
cmdstan_model_wrapper <- function(
  stan_string = NULL, # The Stan model code as a character string
  inputs,             # Data list for the model
  seed,               # Seed for reproducibility
  ...                 # Additional arguments for cmdstan_model or sample
) {
  # Create a temporary .stan file within the sandbox
  stan_file <- tempfile(pattern = "model_", fileext = ".stan")
  writeLines(stan_string, con = stan_file)

  # Compile the Stan model
  # cmdstanr will find cmdstan via the CMDSTAN environment variable
  model <- cmdstanr::cmdstan_model(
    stan_file = stan_file,
    ...
  )

  # Sample from the posterior
  fitted_model <- model$sample(
    data = inputs,
    seed = seed,
    ...
  )

  return(fitted_model)
}

Now, we use this wrapper in our pipeline:

# ... (continuation of pipeline_steps list)
  rxp_r(
    model, # Target name for the fitted model object
    cmdstan_model_wrapper(
      stan_string = bayesian_linear_regression_model,
      inputs = inputs,
      seed = 22
    ),
    user_functions = "functions.R",
    encoder = "save_model",
    env_var = c("CMDSTAN" = "${defaultPkgs.cmdstan}/opt/cmdstan")
  )

Explanation of the Wrapper Approach:

1. stan_string = bayesian_linear_regression_model: We pass the model code (read by rxp_r_file) as a string to our wrapper.
2. writeLines(stan_string, con = stan_file): Inside the wrapper, the Stan code is written to a temporary .stan file. This file exists within the sandbox of the current rxp_r step. This is crucial because cmdstan_model needs a file path. Attempting to pass the original model.stan path directly via additional_files to cmdstan_model can lead to permission or path issues when cmdstan tries to compile it from a different working directory or context.
3. cmdstanr::cmdstan_model(): Compiles the model from the temporary stan_file.
4. model$sample(): Samples from the compiled model.
5. Single Step: Both compilation and sampling must happen within the same rxp_r step (and thus the same sandbox). This is because the model object returned by cmdstan_model() contains paths to the compiled executable. If these were separate steps, the paths from the compilation sandbox wouldn’t be valid in the sampling sandbox.
6. env_var = c("CMDSTAN" = "${defaultPkgs.cmdstan}/opt/cmdstan"): This sets the CMDSTAN environment variable within the sandbox for this specific step. {cmdstanr} uses this variable to locate the cmdstan installation. The ${defaultPkgs.cmdstan} is a Nix interpolation that resolves to the path of the cmdstan package in the Nix store. If the environment providing cmdstan were named differently, for example cmdstan-env.nix, then you would need to use ${cmdstan_envPkgs.cmdstan}.

save_model <- function(fitted_model, path, ...) {
  fitted_model$save_object(file = path, ...)
}

Summary

Using {cmdstanr} with {rixpress} involves these key considerations:

• Include cmdstan in system_pkgs and {cmdstanr} (from Git) in your {rix} environment definition.

• Read your .stan file into the pipeline using rxp_r_file().

• Implement a wrapper function that:

  • Takes the model code string and writes it to a temporary .stan file inside the wrapper.
  • Calls cmdstanr::cmdstan_model() on this temporary file.
  • Calls model$sample() to fit the model.
  • Returns the fitted model object.

• Perform model compilation and sampling within the same rxp_r() call using the wrapper.

• Set the CMDSTAN environment variable for the rxp_r() step that runs the wrapper, pointing to the Nix store path of cmdstan.

• Use {cmdstanr}’s $save_object() method via a custom encoder for robust saving of the fitted model.

This approach ensures that cmdstan can operate correctly within the isolated and reproducible environment provided by {rixpress} and Nix.