flutter_edge_ai_onnx 0.5.2

ONNX Runtime engine for flutter_edge_ai: GenAI inference on 5 native platforms + web generation (Transformers.js) & embeddings (onnxruntime-web).

0.5.2

• Require flutter_edge_ai 2.x.

0.5.1

• Renamed from flutter_gemma_onnx.
• Fix web text generation failing with NoSuchMethodError on every call.
• Web setup pins onnxruntime-web 1.30.0 and Transformers.js 4.3.0.

0.5.0

• Native inference activeBackend reports null instead of echoing the requested backend.

0.4.0

• No longer depends on flutter_gemma_embeddings; asks core for a tokenizer, so register embeddingTokenizers:.

0.3.3

• Prefer pooler_output over last_hidden_state; re-index corpora from any graph exposing both.
• Refuse a SigLIP2 tokenizer.json instead of embedding it with Gemma's convention.
• Require flutter_gemma_embeddings 2.1.0, which fixes flutter build web for the web embedding arm.

0.3.2

• OnnxHuggingFaceResolver installs an ORT-GenAI model directory from a Hugging Face repo — lists the repo, picks a CPU execution-provider folder, downloads the whole bundle (#454).

0.3.1

• Fix a rebuilt native library not reaching the build.
• Fail the build on an unusable runtime — a failed download, a checksum mismatch, or an Android minSdk below 24 — instead of bundling nothing.

0.3.0

• Add web text generation via Transformers.js (@huggingface/transformers) — OnnxEngine web arm (WebGPU/WASM, HF-repo-id models).
• Fix iOS embeddings: probe ORT GetApi down from v27 (the ORT inside onnxruntime-genai predates 1.27).
• Fix Windows embeddings: pass the model path to OrtCreateSession as UTF-16 (ORTCHAR_T is wchar_t on Windows).
• Device-verify embeddings on macOS + Linux + Windows + Android + iOS (unified onnx_embedding_device_test).

0.2.0

• Add web support for OnnxEmbeddingBackend via onnxruntime-web (WebGPU/WASM, WordPiece models).
• OnnxEngine (text generation) declines on web for now — Transformers.js support is a fast-follow.

0.1.0

• Initial release: OnnxEngine — ORT-GenAI text generation over dart:ffi, macOS/Linux/Windows/Android/iOS arm64.
• Add OnnxEmbeddingBackend: plain ONNX Runtime embedding forward pass over dart:ffi (no ORT GenAI).
• Fix ORT/GenAI co-location via ORT_LIB_PATH (dladdr) — no Podfile/packaging step needed.
• Harden GenAiFfiClient lifecycle: cancel/stop/close races can no longer wedge the mutex or hang a caller.
• Platform-gate both engines to macOS/Linux/Windows/Android arm64 + iOS arm64; other hosts fail loud, not with a confusing native error.
• Add OrtClient/OrtFfiClient: injectable ORT 1.27.0 C API seam, fake-testable with zero dlopen.
• Verified against real all-MiniLM-L6-v2 (WordPiece) and EmbeddingGemma-300M-ONNX (SentencePiece) models on macOS.