Skip to contents

Installation

Source: vignettes/installation.Rmd
installation.Rmd

After the usual R package installation, torch requires installing other 2 libraries: LibTorch and LibLantern. They are automatically installed by detecting information about you OS if you are using torch in interactive mode. If you are running torch in non-interactive environments you need to set the TORCH_INSTALL env var to 1, so it’s automatically installed or manually call torch::install_torch().

Starting from torch v0.9.1.9000 it’s possible to install torch from pre-built package binaries available in a custom CRAN-like repository. This binaries include all shared objects necessary to run torch, thus it’s not required to install any additional software. See the Install from pre-built binaries session for more information.

We have provide pre-compiled binaries for all major platforms and you can find specific installation instructions below.

Windows

CPU

If you don’t have a GPU or want to install the CPU version of torch, you can install with:

Some Windows distributions don’t have the Visual Studio runtime pre-installed and you will observe an error like:

Error in cpp_lantern_init(normalizePath(install_path())): C:\Users\User\Documents\R\R-4.0.2\library\torch\deps\lantern.dll - The specified module could not be found.

See here for instructions on how to install it.

GPU

Since version 0.1.1 torch supports GPU installation on Windows. In order to use GPU’s with torch you need to:

  • Have a CUDA compatible NVIDIA GPU. You can find if you have a CUDA compatible GPU here.

  • Have properly installed the NVIDIA CUDA toolkit version 11.7. For CUDA v11.7, follow the installation instructions here. Note: The version of the CUDA toolkit must match exactly what’s mentioed above.

  • Have installed cuDNN - a version compatible with CUDA v11.7. Follow the installation instructions available here.

Once you have installed all pre-requisites you can install torch with:

If you have followed default installation locations we will detect that you have CUDA software installed and automatically download the GPU enabled Lantern binaries. You can also specify the CUDA env var with something like Sys.setenv(CUDA="11.7") if you want to force an specific version of the CUDA toolkit.

MacOS

CPU

We only support CPU builds of torch on MacOS. On MacOS you can install torch with:

Linux

CPU

To install the cpu version of torch you can run:

GPU

To install the GPU version of torch on linux you must verify that:

  • You have a NVIDIA CUDA compatible GPU. You can find if you have a CUDA compatible GPU here.

  • You have correctly installed the NVIDIA CUDA Toolkit versions 11.6 or 11.7, follow the instructions here.

  • You have installed cuDNN (a version compatible with your CUDA version). Follow the installation instructions available here.

Once you have installed all pre-requisites you can install torch with:

If you have followed default installation locations we will detect that you have CUDA software installed and automatically download the GPU enabled Lantern binaries. You can also specify the CUDA env var with something like Sys.setenv(CUDA="11.7") if you want to force an specific version of the CUDA toolkit.

Installing from pre-built binaries

As of torch v0.9.1.9000 it’s now possible to install torch from pre-built package binaries from a CRAN like repository hosted on Google Cloud Storage. We currently provide pre-built binaries for CPU (for macOS, Linux and Windows) and GPU (Windows and Linux).

Packages provided by the CRAN-like repository bundles all necessary for its execution, including CUDA and CUDNN in the case of the GPU builds. This means that by installing it you agree with the included software licenses. See PyTorch’s LICENSE and CUDA libraries EULA.

When installing from the pre-built binaries, you don’t need to manually install CUDA or cuDNN. If you have CUDA installed, it doesn’t need to match the installation ‘kind’ chosen below.

To install from the pre-built binaries, you can use the following:

options(timeout = 600) # increasing timeout is recommended since we will be downloading a 2GB file.
# For Windows and Linux: "cpu", "cu117" are the only currently supported
# For MacOS the supported are: "cpu-intel" or "cpu-m1"
kind <- "cu117"
version <- available.packages()["torch","Version"]
options(repos = c(
  torch = sprintf("https://storage.googleapis.com/torch-lantern-builds/packages/%s/%s/", kind, version),
  CRAN = "https://cloud.r-project.org" # or any other from which you want to install the other R dependencies.
))
install.packages("torch")

Troubleshooting

Large file download timeout

If you encounter timeout during library download, or if after a while, downloads end-up with a warning such as:

Warning messages:
1: In utils::download.file(library_url, temp_file) :
  downloaded length 44901568 != reported length 141774525
2: In utils::download.file(library_url, temp_file) :
  URL '...': Timeout of 60 seconds was reached
3: Failed to install Torch, manually run install_torch(). download from 'https://download.pytorch.org/libtorch/cpu/libtorch-macos-1.7.1.zip' failed

This means you encounter a download timeout. then, you should increase the timeout value in install_torch() like

install_torch(timeout = 600)

File based download

In cases where you cannot reach download servers from the machine you intend to install torch on, last resort is to install Torch and Lantern library from files. This is done in 3 steps :

1- get the download URLs of the files.

2- save those files into the machine filesystem. We will use /tmp/ here as an example .

3- install torch from files

# then after making both files available into /tmp/
Sys.setenv(TORCH_URL="/tmp/libtorch-v1.13.1.zip")
Sys.setenv(LANTERN_URL="/tmp/lantern-0.9.1.9001+cpu+arm64-Darwin.zip")
torch::install_torch()