Backend-agnostic model train + infer for Orama Lake, behind one interface (design decision D19).
The same training/inference contract runs on a native backend per platform — MLX on macOS, libtorch or llama.cpp on Linux/Windows, CUDA in the cloud. Not MLX-specific: MLX is the first target, but the interface is there from day one so each environment uses its native engine. The specific models live with the components that use them and run on this runtime.
Small, single-responsibility module (D21). The decisions it is built on (D7, D16,
D19, D21, D26) are in docs/design.md.
Released: v0.1.1 (all releases).
cmake -S . -B build -G Ninja && cmake --build build # interface + tests, no backend
ctest --test-dir build --output-on-failureBackends are opt-in CMake toggles (-DORAMA_BACKEND_MLX=ON, …_LIBTORCH, …_LLAMACPP, …_CUDA) so the scaffold + CI build with no ML toolchain.
Real Gemma‑2 (pure C++/MLX, no Python): build/run/validate/benchmark the real model — see docs/gemma-runbook.md (weights download, gemma-cli, the sudo gemma-bench perf harness, the garbage-if-wrong checklist, troubleshooting).
C++23 (→ C++26 per D7), CMake ≥ 3.28 + Ninja, FetchContent, Catch2. Library-first
(D16); links as orama::model-runtime.
C++23 (→ C++26 per D7), CMake ≥ 3.28 + Ninja, FetchContent, Catch2. Library-first (D16); links as orama::model-runtime.
MIT, see LICENSE. Third-party components and their licenses are in THIRD_PARTY_LICENSES. Contributions are welcome under the Contributor License Agreement; see CONTRIBUTING.md.