Package website: release | dev
{pjrt} is an R interface to PJRT, the runtime API of OpenXLA. It compiles StableHLO (or HLO) programs with the XLA compiler into executables for a specific device and runs them there. StableHLO programs do not depend on the framework that produced them, so a program exported from, e.g., JAX can be run by {pjrt}. To create StableHLO programs from R, see the stablehlo package.
{pjrt} is a low-level runtime. For array computing in R with JIT compilation and automatic differentiation, use anvl, which builds on {pjrt}.
install.packages("pjrt", repos = c("https://r-xla.r-universe.dev", getOption("repos")))
pjrt::install_pjrt()
Each backend (CPU or CUDA) is a PJRT plugin: a shared library that is
not bundled with the package but downloaded separately. install_pjrt()
downloads the CPU plugin and, if an NVIDIA GPU is detected, the CUDA
plugin together with the {pjrt.cuda} package, which provides the CUDA
runtime libraries. Without install_pjrt(), a plugin is downloaded the
first time its platform is used. See ?pjrt for the environment
variables that control this.
Below, we compile and run a StableHLO program that adds two f32
tensors of shape (2x2).
library(pjrt)
src <- r"(
func.func @main(
%x: tensor<2x2xf32>,
%y: tensor<2x2xf32>
) -> tensor<2x2xf32> {
%0 = "stablehlo.add"(%x, %y) : (tensor<2x2xf32>, tensor<2x2xf32>) -> tensor<2x2xf32>
"func.return"(%0): (tensor<2x2xf32>) -> ()
}
)"
program <- pjrt_program(src, format = "mlir")
program
#> PJRTProgram(format=mlir, code_size=221)
#>
#> func.func @main(
#> %x: tensor<2x2xf32>,
#> %y: tensor<2x2xf32>
#> ) -> tensor<2x2xf32> {
#> ...
executable <- pjrt_compile(program, device = "cpu")
x <- pjrt_buffer(c(1, 2, 3, 4), shape = c(2, 2), dtype = "f32")
x
#> PJRTBuffer
#> 1 3
#> 2 4
#> [ CPUf32{2x2} ]
y <- pjrt_buffer(c(5, 6, 7, 8), shape = c(2, 2), dtype = "f32")
y
#> PJRTBuffer
#> 5 7
#> 6 8
#> [ CPUf32{2x2} ]
pjrt_execute(executable, x, y)
#> PJRTBuffer
#> 6 10
#> 8 12
#> [ CPUf32{2x2} ]
NA handling.| OS | Architecture | CPU | CUDA |
|---|---|---|---|
| Linux | x86_64 | ✓ | ✓ |
| Linux | arm64 | ✓ | ✓ |
| Windows | x86_64 | ✓ | WSL2 |
| Windows | arm64 | ✗ | ✗ |
| macOS | x86_64 | ✓ | ✗ |
| macOS | arm64 | ✓ | ✗ |
✓ supported · WSL2: via the Windows Subsystem for Linux · ✗ not supported
./tools/copy-proto.R for which ones.