Installation

Currently, {anvil} is not available on CRAN, so you either have to install it via r-universe or from GitHub.

System Dependencies

You need a C++20 compiler, libprotobuf, and the protobuf-compiler.

Installing the latest release from GitHub

You can install the latest GitHub release via:

pak::pak("r-xla/anvil@*release")

You can install the latest release from r-universe via:

options(repos = c(
  rxla = "https://r-xla.r-universe.dev",
  CRAN = "https://cloud.r-project.org/"
))
install.packages("anvil")

To confirm that your CPU installation is working, run:

library(anvil)
nv_scalar(1, device = "cpu")

Installation the development version from GitHub

The development version can be installed via:

pak::pak("r-xla/anvil")

GPU Support

Running {anvil} with GPU support only works on Linux or via WSL2 on Windows. In principle, MPS support on macOS is available but many features are missing and the PJRT plugin is heavily outdated. To use {anvil} with GPU support, you need to have CUDA installed. The PJRT CUDA plugin was built with the following versions:

CUDA 12.8.1
CUDNN 9.8.0
NVSHMEM 3.2.5

Besides, there is this dockerfile which is used in the GitHub Actions CI, and the versions used therein should also work.

To confirm that your GPU installation is working, run:

library(anvil)
nv_scalar(1, device = "cuda")

If you are working on a server, you can also use the prebuilt Docker images, see the next section.

Docker

Prebuilt Docker images are available in r-xla/docker. This includes a CUDA and CPU build for amd64/x86-64 architecture:

Available Images

Image	Description
`anvil-cpu`	CPU support, based on `rocker/r-ver`
`anvil-cuda`	GPU support with CUDA 12.8
`anvil-cuda-base`	Base image for `anvil-cuda` without `anvil` installed

Note that running the GPU container requires the NVIDIA Container Toolkit to be installed on the host.