tests/testthat/helper-torch.R

In rtemis: Schema-Driven Machine Learning and Visualization

# helper-torch.R
# ::rtemis::
# 2026- EDG rtemis.org

# Every test that fits a torch-backed algorithm -- MLP, TabNet -- gates on this.
# Two things have to be present and `requireNamespace()` answers only for the
# first: the R package installs without libtorch, which it downloads on first
# use. A check that stops at the package passes on a machine where the fit then
# fails inside `torch::cuda_is_available()`, which is what device resolution
# reaches before any module is built.

# %% torch_available ----
torch_available <- function() {
  requireNamespace("torch", quietly = TRUE) && torch::torch_is_installed()
}


# %% skip_if_no_torch ----
skip_if_no_torch <- function() {
  testthat::skip_if_not(torch_available(), "libtorch is not installed")
}


# %% libtorch threads ----
# libtorch runs on every core it sees, and CRAN allows two. A fit asks for its
# resolved algorithm workers, but the native backend (macOS) accepts one thread
# setting per process, so the first request wins. Pinning the test process here,
# before any torch fit, keeps every torch test within two threads wherever the
# fits' own requests land.
if (torch_available()) {
  set_torch_threads(
    min(2L, as.integer(torch::torch_get_num_threads())),
    verbosity = 0L
  )
}