cpp_options = list(stan_threads = TRUE) does not work when included in src/install.libs.R.

[toddpocuca]

My install.libs.R file has been modified like so:

# ...
callr::r(
  func = function(bin_stan) {
    instantiate::stan_package_compile(
      models = instantiate::stan_package_model_files(path = bin_stan),
      cpp_options = list(stan_threads = TRUE) ## <-- extra line
    )
  },
  args = list(bin_stan = bin_stan),
  show = TRUE,
  stderr = "2>&1"
)

However whenever I load and run the model I get this warning:

1: 'threads' is set but the model was not compiled with 'cpp_options = list(stan_threads = TRUE)' so 'threads' will have no effect!

And my current workaround is to just re-compile the model:

    model <- instantiate::stan_package_model(
        name = "disize",
        package = "disize",
        compile = TRUE,
        force = TRUE,
        cpp_options = list(stan_threads = TRUE)
    )

A reproducible example is available at https://github.com/toddmccready/disize/tree/errors by running the tests in tests/testthat/test-disize.R

[wlandau]

I installed https://github.com/toddmccready/disize/tree/errors and noticed STAN_THREADS=true in the output:

stan_version_major = 2
stan_version_minor = 36
stan_version_patch = 0
STAN_THREADS=true
STAN_MPI=false
STAN_OPENCL=false
STAN_NO_RANGE_CHECKS=false
STAN_CPP_OPTIMS=false
** R
** byte-compile and prepare package for lazy loading
** help
*** installing help indices
** building package indices
** testing if installed package can be loaded from temporary location
** testing if installed package can be loaded from final location
** testing if installed package keeps a record of temporary installation path
* DONE (disize)

Inside disize::disize() just before model$optimize(), model$cpp_options() returns:

$STAN_THREADS
[1] TRUE

$STAN_VERSION
[1] "2.36.0"

So it is actually compiling with threads enabled. If you run:

library(cmdstanr)
library(disize)
library(tidyverse)
debugonce(cmdstanr:::assert_valid_threads)

and then run https://github.com/toddmccready/disize/blob/e140984f50f33d4aaae4e5f06ea27bb584c6a1ec/tests/testthat/test-disize.R#L8-L78, you will see threads is 5 and cpp_options equals the list above. The problem is that cmdstanr is checking cpp_options[["stan_threads"]] instead of cpp_options[["STAN_THREADS"]]. That's why it thinks threads are disabled. Please submit a bug report to cmdstanr to capitalize "stan_threads" in https://github.com/stan-dev/cmdstanr/blob/5f2a8c850890e54e6fccc8108e5a31cb08010ab8/R/cpp_opts.R#L149-L171:

assert_valid_threads <- function(threads, cpp_options, multiple_chains = FALSE) {
  threads_arg <- if (multiple_chains) "threads_per_chain" else "threads"
  checkmate::assert_integerish(threads, .var.name = threads_arg,
                               null.ok = TRUE, lower = 1, len = 1)
  if (is.null(cpp_options[["stan_threads"]]) || !isTRUE(cpp_options[["stan_threads"]])) { # should be cpp_options[["STAN_THREADS"]]
    if (!is.null(threads)) {
      warning(
        "'", threads_arg, "' is set but the model was not compiled with ",
        "'cpp_options = list(stan_threads = TRUE)' ",
        "so '", threads_arg, "' will have no effect!",
        call. = FALSE
      )
      threads <- NULL
    }
  } else if (isTRUE(cpp_options[["stan_threads"]]) && is.null(threads)) { # should be cpp_options[["STAN_THREADS"]]
    stop(
      "The model was compiled with 'cpp_options = list(stan_threads = TRUE)' ",
      "but '", threads_arg, "' was not set!",
      call. = FALSE
    )
  }
  invisible(threads)
}

[stouffer]

I recently came across this same issue and have been frustrated to see how slowly the fix upstream via cmdstanr has taken.

Just in case @toddpocuca or anyone else wants to include an unsatisfactory "fix" to still allow a model to make use of threading, the following hacky line of code inserted after calling instantiate::stan_package_model() worked for me:

my_model <- function() {
  model <- instantiate::stan_package_model(
    name = 'foo',
    package = 'bar'
  )
  model$.__enclos_env__$private$cpp_options_$stan_threads <- TRUE
  return(model)
}

[jgabry]

Very sorry for the delay in cmdstanr! I have an open PR for it now: stan-dev/cmdstanr#1212. I had to focus for a while on some other projects with looming deadlines, but making a big push now towards a major release of cmdstanr. In hindsight, this was a simple enough fix that I should have made it a priority to handle it before shifting work to those other projects.

[jgabry]

This should finally be addressed by stan-dev/cmdstanr#1212. I apologize for the delay!