mnn_engine 0.1.0

Run MNN LLMs on Android and expose an on-device OpenAI-compatible API server from Flutter.

0.1.0 (unreleased)

• Upgrade the pinned MNN source and bundled Android libraries to 3.6.1.

• Add CPU/OpenCL/Vulkan backend selection, capability checks and resident backend reporting.

• Build Vulkan buffer by default; retain opt-in Hexagon host/DSP builds and deployment.

• Report unavailable backends without silently retrying on CPU; separate model caches by backend and version.

• Share native event streams between multiple Dart listeners.

• Extend GitHub Actions with optional Hexagon SDK builds, complete native bundles, failure logs and independent example APK verification.

• Add a pinned Snapdragon Docker build using Hexagon SDK 6.6, a Community SDK domain adapter, toolchain provenance and DSP dependency/symbol validation.

• Bundle v73/v75/v79/v81 DSP runtimes behind one ARM64 stub, query the device ISA with FastRPC, and extract only the matching runtime. Report the detected architecture to consumers.

• Default GitHub Actions and local builds to CPU/OpenCL/Vulkan; Hexagon requires explicit opt-in, including for its host backend. Tag builds use the default CPU/GPU bundle.

• Upgrade JNI adapter ABI to 8, keep bounded native error context through a direct MNN log bridge, and remove temporary logger hooks, Logcat fallback and verbose diagnostic dumps.

• Preserve MNN 3.6.1 MTOK added-token decoding, generation error checks and the experimental Hexagon model-format guard; maintain native/tokenizer regression tests independently of device investigation tools.

• Release failed generation sessions and clear resident-model state so loading the same model rebuilds the runtime; preserve reuse after normal completion, length limits and cancellation.

• Reject new or pending generation requests when server shutdown begins, without queueing them or reopening admission when an old request finishes.

• Check cancellation between prefill chunks (128 tokens by default for standard causal models), preserve explicit model chunk settings and stops received before native entry, and align multimodal positions and deep-stack inputs with chunked embeddings.

• Align multi-turn tool-call messages and completion responses with llama-server.

• Support both JSON and tagged-parameter MNN tool-call output formats.

0.0.2

• Use the imported model directory name as the MNN runtime and API model ID.
• Add model-directory renaming with case-insensitive duplicate-name checks.
• Add import-result APIs with optional automatic conflict-free model naming.

0.0.1

• Bundle verified Android arm64 native libraries for pub.dev consumers.
• Stop compiling MNN from consumer Gradle builds.
• Use the Ktor CIO server engine to avoid Netty packaging conflicts in host applications.
• Add reproducible GitHub Actions native builds and public artifact metadata.
• Add bilingual package documentation and pub.dev publishing guidance.
• Make server binding authoritative, reuse safe local ports, and report real bind conflicts without a race-prone preflight in the main startup path.
• Add Android arm64 MNN 3.6.0 native build and JNI runtime.
• Add complete model directory import and validation.
• Add foreground Ktor OpenAI-compatible server and SSE chat completions.
• Add runtime state and independent log event streams.