openvino_genai 0.1.0

On-device LLM inference for Flutter via Intel OpenVINO GenAI: streaming chat completion on CPU/GPU with hardened load and fallback behaviour. Windows x64; unofficial.

0.1.0

Initial release.

• Streaming chat completion (generateChatStream / generateText) through the OpenVINO GenAI C API, with the model's own chat template applied natively.
• Device-ladder loading (devicePreference, default GPU → CPU; supports NPU → GPU → CPU) with a per-accelerator smoke test (opt-out via OpenVinoGenAiService.acceleratorSmokeTestEnabled) and process-lifetime demotion of failed devices.
• Isolate-offloaded model load/free, bounded load waits with a detached self-healing load, stall-based generation timeout, busy-refusal retry.
• Typed exception hierarchy (OpenVinoGenAiException).
• OpenVinoGenAiService.doctor() diagnostics (DLL path, bridge version, runtime pin, pipeline state).
• Windows x64 only, pinned to OpenVINO GenAI runtime 2026.0.0.0 via openvino_genai_windows.

Known limitation (documented, by design for 0.x): the native pipeline is a process-wide singleton — one model resident at a time, one generation at a time. A handle-based API is planned for 1.0.