A backend needs more than the packages anvl declares as dependencies: the
"pjrt" backend runs on PJRT plugins that are downloaded rather than shipped
with the package, and the "quickr" backend needs quickr, which is only
suggested. This installs whichever of the two the given backend is missing.
Usage
install_anvl(backend = active_backend(), ...)Arguments
- backend
(
character(1))
Backend to install for. Defaults toactive_backend().- ...
Passed to the underlying installer:
pjrt::install_pjrt()for"pjrt",utils::install.packages()for"quickr".
Details
The PJRT plugins are downloaded on demand, but not silently: the first time a
plugin is needed, an interactive session asks for confirmation, while a
non-interactive session does not download at all. Call this to make the
download an explicit step instead, for instance in a Dockerfile layer of its
own or at the start of a script that later runs unattended. The PJRT_INSTALL
environment variable overrides the prompt: "1" always downloads without
asking, "0" never downloads.
For "pjrt", the CPU plugin is always installed, and the CUDA plugin too
when an NVIDIA GPU is detected on Linux (or cuda = TRUE is passed). The
CUDA plugin additionally needs the CUDA libraries, which come in the
pjrt.cuda R package from the r-xla r-universe; install_anvl() installs
it along with the CUDA plugin. See pjrt::install_pjrt() for details.