pjrt

PJRT runtime libraries, each shipped with the exact C ABI headers it was built against. Headers are vendored in-tree; the libraries are fetched on demand.

PJRT C ABI headers from OpenXLA at 91888df6 — API version 0.115.

The repository holds xla/ (headers), devices/ (fetched runtimes), fetch.sh, pjrt.cmake, VERSION (machine-readable provenance, kept in sync with the headers line by scripts/sync_headers.sh), and the maintainer scripts/.

Fetch

./fetch.sh all        # everything in the release
./fetch.sh cpu        # CPU plugin for this arch
./fetch.sh cuda12     # cuda/12 for this arch (also fetches cpu)
./fetch.sh rocm7      # rocm/7
./fetch.sh mps        # Apple GPU (arm64)

Libraries land under devices/:

devices/cpu/{x86_64,aarch64}/
devices/cuda/{11/x86_64, 12/{x86_64,aarch64}, 13/{x86_64,aarch64}}/
devices/rocm/{6,7,10}/x86_64/
devices/xpu/x86_64/   devices/tpu/x86_64/   devices/mps/arm64/

GitHub publishes a SHA256 digest for every release asset.

Use

include(pjrt.cmake)
target_link_libraries(app PRIVATE pjrt::pjrt)

#include "xla/pjrt/c/pjrt_c_api.h"
#include "xla/pjrt/pjrt_loader.h"   // pjrt::load, PJRT_HAS, PJRT_EXT, find_extension

pjrt::load opens a plugin (dlopen → GetPjrtApi). Gate core and extension fields on struct_size with PJRT_HAS / PJRT_EXT, and look up extensions with find_extension (which also handles the pre-0.45 layout). Compile as C++17.

Maintain

scripts/sync_headers.sh   # refresh xla/ and the headers line from openxla/xla
scripts/collect.sh        # extract runtimes from public wheels
scripts/release.sh        # publish one pjrt-YYYY-MM-DD release

CI syncs the headers, opening a PR on drift, and publishes a daily release.