oird — OIR native inference daemon

Native C++ daemon that owns model residency and inference for the Open Intelligence Runtime. Runs as a system service under its own SELinux domain (u:r:oird:s0), registers oir_worker with servicemanager, and serves requests from OIRService over AIDL.

What it does

• Loads LLM / VLM / ONNX / whisper models on demand.
• Shares loaded models across every app that asks for the same capability — one copy in memory, N callers.
• Pools inference contexts per model (ContextPool for llama-backed, WhisperPool for whisper) with priority-aware wait queues.
• Accounts KV-cache memory in the resident budget so eviction decisions are accurate.
• Dispatches across backends (llama.cpp, whisper.cpp, ONNX Runtime, libmtmd) based on capability.

Tree location

Installs as /system_ext/bin/oird via prebuilt_etc. Lives at system/oird/ in the AOSP tree.

Building

oird is built as part of a JibarOS tree:

cd ~/aaosp
source build/envsetup.sh
lunch aosp_cf_x86_64_phone-trunk_staging-userdebug
m -j8 oird

Dependencies

• AIDL interfaces from oir-framework-addons
• platform_external_llamacpp
• platform_external_whispercpp
• platform_external_onnxruntime

See also

github.com/Jibar-OS/JibarOS for architecture + capability model.