Install and configure torch with GPU support

Installs the torch R package (if missing) and downloads the platform-appropriate libtorch + lantern binaries with GPU support. After installation, verifies that GPU acceleration works via check_torch().

Usage

setup_torch(reinstall = FALSE, type = NULL, verbose = TRUE)

Arguments

reinstall

Logical. Force re-download of libtorch/lantern binaries even if they already exist. Use this to fix broken CUDA installations (e.g., CPU-only lantern accidentally installed). Default: FALSE.

type

Character or NULL. Installation type, passed to torch::install_torch(). Common values: NULL (auto-detect), "cuda" (force CUDA on Windows/Linux), "cpu" (force CPU-only). Default: NULL (auto-detect based on available GPU).

verbose

Logical. Default: TRUE.

Value

Invisibly, the result of check_torch() after installation, or a partial list with needs_restart = TRUE if R must be restarted.

Details

The installation proceeds in stages:

1. If the torch R package is not installed, it is installed from CRAN.

2. torch::install_torch() downloads libtorch and lantern binaries. On Windows/Linux with an NVIDIA GPU and a CUDA toolkit, this automatically selects CUDA-enabled binaries.

3. A verification step runs check_torch() to confirm GPU works.

Important: If the torch package was already loaded in the current R session before calling setup_torch(), you must restart R for the new binaries to take effect. The function detects this and warns accordingly.

CUDA auto-detection

torch::install_torch() detects the CUDA version from:

• The CUDA environment variable (if set, forces a version)

• CUDA_PATH (Windows) or CUDA_HOME (Linux)

• nvcc --version on PATH

See also

check_torch() to diagnose the current setup, use_gpu() to enable GPU acceleration after setup.

Examples

if (FALSE) { # \dontrun{
setup_torch()                    # auto-detect GPU
setup_torch(type = "cuda")       # force CUDA
setup_torch(reinstall = TRUE)    # fix broken install
} # }