flutter_gemma_litertlm

What's Changed

• fix(litertlm): don't clamp maxTokens on NPU, and say which models the NPU runs by @DenisovAV in #508
• docs(translategemma): the GPU failure is silent padding, not an aborted load by @DenisovAV in #509
• test(translategemma): the smoke test's CPU pin had the old reason too by @DenisovAV in #510
• fix(core): .litertlm on iOS was wrapped in turn markers twice by @DenisovAV in #511
• fix(docs): platform permissions — macOS entitlements, Android GPU entries, speech microphone by @DenisovAV in #513
• fix(android): release builds on AGP 9 — MediaPipe R8 rules, agent webview opt-out, speech example by @DenisovAV in #514
• docs(codelabs): Package Skills codelab by @DenisovAV in #512
• docs: version pins for the 1.8.3 / 1.6.4 / 1.0.6 / 0.6.1 release by @DenisovAV in #515

Full Changelog: v1.8.2...v1.8.3

What's Changed

• docs(website): full content audit + Built-in AI page + Engines nav group by @DenisovAV in #455
• feat(rag_qdrant)!: migrate to the official qdrant_edge UniFFI SDK by @DenisovAV in #458
• fix(rag_sqlite, hooks): a rebuilt native library never reached the build, and a failed initialize() reported ready by @DenisovAV in #459
• docs(changelog): name the web corpus bug and the Android minSdk floor by @DenisovAV in #462
• docs(rag_sqlite): the README promised a fetch the package never did by @DenisovAV in #463
• fix(rag_sqlite): the web store could not take an ordinary number — 1.3.1 by @DenisovAV in #464
• docs(changelog): name the mustNot filter fix in 1.3.1 by @DenisovAV in #465
• feat(core): Hugging Face manifest-resolver seam (#454) by @DenisovAV in #461
• feat(litertlm): LitertlmManifestResolver — manifest-driven Hugging Face install (#454) by @john-rocky in #469
• feat: Hugging Face resolvers for all engines + core 1.7.0 release by @DenisovAV in #470
• feat(hf): one-call fromHuggingFace + engine-carried HF resolvers by @DenisovAV in #471
• fix(builtin-ai): iOS/macOS build failed on Xcode below 26.4 (#460) by @DenisovAV in #472
• feat(onnx): install ORT-GenAI models from Hugging Face as a directory (#454) by @DenisovAV in #473
• docs(agent): run_js tool takes no scriptName (0.2.5) by @DenisovAV in #474
• flutter_gemma_litertlm 1.6.1: resolver auto-registers from the engine (docs) by @DenisovAV in #475
• docs(genkit): description covers all models, not just Gemma by @DenisovAV in #476
• feat(genkit_hybrid): 0.2.0 — CapabilityStrategy, CostStrategy, cascadeModel + accept-throw fix by @DenisovAV in #477
• test(litertlm): offline manifest sweep on the live catalog + opt-in Hugging Face check (#454) by @john-rocky in #479
• docs: the per-session sampler claim was false — correct it, add a reverse probe by @DenisovAV in #432
• fix(example): a failed setUpAll made the web suite report "All tests passed" by @DenisovAV in #466
• feat(website): publish the hybrid-AI codelab at /codelabs on fluttergemma.dev by @DenisovAV in #481
• chore(genkit): support genkit 0.16 — genkit_flutter_gemma 0.6.0, genkit_hybrid 0.2.1 by @DenisovAV in #482
• fix(codelab): lookupAction takes ActionType in genkit 0.16 by @DenisovAV in #483
• docs(codelab): frame cascade accept as a demo proxy, name the real confidence signal by @DenisovAV in #484
• docs(codelab): say what the app is before teaching how to build it by @DenisovAV in #485
• feat(embeddings): SigLIP2 text-embedding profile (pooler_output + no-BOS/EOS/pad-64) by @DenisovAV in #467
• fix(macos): stage companions by repointing the load command that is there (#457) by @DenisovAV in #478
• docs(codelab): realign with the refactored app; teach the on-device context budget by @DenisovAV in #487
• fix(core): a web session right after install could not find the model — 1.7.2 by @DenisovAV in #468
• feat(website): a codelab catalogue at /codelabs by @DenisovAV in #488
• Two codelabs in the monorepo, on all six platforms, with a nightly gate by @DenisovAV in #489
• The hybrid-AI codelab, migrated into the monorepo as step directories by @DenisovAV in #491
• fix(embeddings): allow dart_sentencepiece_tokenizer 1.4.1 — embeddings 2.1.1 by @DenisovAV in #490
• docs(embeddings): document the bare-content tokenizer contract in the README by @DenisovAV in #495
• The multimodal codelab: two models, two modalities, on all six platforms by @DenisovAV in #493
• fix(core): createChat never passed tools; web to @litert-lm/core 0.17.0 by @DenisovAV in #498
• feat(example): LFM2.5-230M and the fromHuggingFace install path — core 1.7.4 by @DenisovAV in #499
• feat(stt): Whisper output language is per-transcription — core 1.8.0, speech 0.5.0 by @DenisovAV in #501
• site: make the author credit a link, matching litetune's footer by @DenisovAV in #502
• chore: add a root LICENSE, and relicense genkit_flutter_gemma to MIT by @DenisovAV in #503
• Add flush(), so a RAG index survives the process (#492) by @DenisovAV in #507
• feat: ship agent skills for consumers of the package by @DenisovAV in #505

New Contributors

• @john-rocky made their first contribution in #469

Full Changelog: v1.6.5...v1.8.2

• A chat stopped mid-reply answered every later message with nothing (#325).

• Google Play no longer rejects apps over 16 KB page sizes (#529).
• Native runtime LiteRT-LM v0.17.1: integer tool-call arguments stay integers.
• Web embeddings work: the four JS files ship here now, rebuilt on @litertjs/core 2.5.3.
• No longer depends on flutter_gemma_embeddings; asks core for a tokenizer, so register embeddingTokenizers:.
• Refuse a multi-input .tflite at load with a named error instead of a bare native status.

• Native runtime native-v0.17.0-a: tool calls no longer crash the app.
• FunctionGemma on .litertlm takes the runtime's tool path — needs core 1.8.4.
• Windows needs no Visual C++ runtime beyond what a Flutter app ships (#456).

• Native runtime: LiteRT-LM v0.17.0 (native-v0.17.0).
• Windows: embeddings and speech failed with CreateTensorBufferFromHostMemory status 3.

What's Changed

• fix(core): stop claiming background_downloader's updates stream from the host by @DenisovAV in #450
• chore: one CHANGELOG line per issue, and write that into the rules by @DenisovAV in #451
• fix(android,ios): drop KGP from the Android plugins (#440); core iOS floor 16 → 15 (#441) by @DenisovAV in #452

Full Changelog: v1.6.2...v1.6.4

• Web: move @litert-lm/core 0.14.0 -> 0.17.0.
• Web: maxOutputTokens now caps generation instead of being ignored.
• Web: function calling works; constrained decoding off (upstream #2434).

What's Changed

• docs(website): vision/audio per-encoder backend (1.5.9 / litertlm 1.4.2) by @DenisovAV in #430
• chore(release): guard-release-docs hook + skill HARD RULES / Definition-of-Done by @DenisovAV in #431
• feat(genkit): 0.15 caps — JSON output, native toolChoice, context-window middleware (genkit_flutter_gemma 0.5.0) by @DenisovAV in #434
• docs(release-skill): publish always from an up-to-date main by @DenisovAV in #435
• fix: make the checks able to fail — CI on all 11 packages, and the guards that outlived their reason by @DenisovAV in #438
• ci(win): assert the shipped OpenVino runtime matches the dispatch's SDK by @DenisovAV in #437
• fix(litertlm): sizeInTokens asks the tokenizer instead of guessing by @DenisovAV in #439
• fix(native): the library did not arrive, and nothing said so — dead lookup paths + a hook that reported success by @DenisovAV in #436
• feat(onnx): ONNX Runtime engine (ORT-GenAI inference + plain-ORT embeddings) + runtime-agnostic embeddings 2.0.0 by @DenisovAV in #443
• fix(rag): filters that returned the wrong rows without saying so by @DenisovAV in #442
• fix(litertlm): Android generation broke when embeddings loaded the library first by @DenisovAV in #448
• feat(onnx): web arm — generation (Transformers.js) + embeddings, iOS emb fix, device tests by @DenisovAV in #449

Full Changelog: v1.5.9...v1.6.2

• Docs: the README and library docs show the engine-carried resolver and the one-call fromHuggingFace(repo) install.

• Add LitertlmManifestResolver: manifest-driven Hugging Face install for repos shipping litertlm_manifest.json, auto-registered from LiteRtLmEngine (#454).

• Fix a rebuilt native library not reaching the build.

• Android: embeddings no longer poison the loader, fixing zero-chunk streams and SIGABRT (#447).

• Share the host-side native library lookup with core instead of a private copy.

• Add LiteRT embedding forward-pass + LiteRtEmbeddingBackend (moved from flutter_gemma_embeddings).
• internal: wire the new tokenizerFactory descriptor field; byte-identical vectors, no API change.

• fix: vision + audio encoders default to CPU (fixes GPU vision hard-fail on Metal/WebGPU); overridable.

• Fix Windows NPU: the OpenVino compiler DLLs shipped but were never bundled into the app.
• Drop the debug TBB variants that were being bundled alongside the release set.

• Migrate to LiteRT-LM v0.16.0 (native-v0.16.0) — fixes the Android OpenCL per-turn leak (#348, #402).
• Detect the stream-callback ABI at runtime — v0.15.0 changed it with no compat path.
• Fix Windows discrete GPU: a Bazel define we passed had been removed upstream.
• Rebuild both NPU dispatch stacks from the pin instead of shipping stale prebuilts.
• Pin the macOS deployment target of libStreamProxy.dylib to 11.0 instead of the build host's.
• Native platforms only — web stays on @litert-lm/core 0.14.0 (the 0.16.0 npm publish ships no dist/).

• Clearer engine-create error for GPU-only .litertlm models run on CPU (#390).

• Also expose the LiteRt interpreter (LiteRtBindings) for embeddings/speech; own web litert.js.

• Migrate FFI to LiteRT-LM v0.14.0 — native per-session sampler (opaque session-config); native-v0.14.0.
• Fix #214 GPU output garbage via the v0.14.0 runtime — verified on Xclipse.
• Bump web @litert-lm/core 0.12.1 → 0.14.0 (text path; API-compatible).
• Known regression: Windows discrete GPU broken upstream (LiteRT-LM #2957) — use CPU/NPU on Windows.

• Smooth UI during Android GPU prefill — flush the OpenCL queue every 2 ops (#364).

• Guard native cancel against a freed conversation — fixes a use-after-free SIGSEGV on close-mid-stream (#379).

• Create the native conversation off the main isolate to avoid ANRs on multimodal models (#365).
• Serialize native conversation create on the engine mutex to prevent a heap-corrupting race (#372).
• Cancel native decode before tearing a conversation down to avoid a multi-second ANR (#364, #373).

• Clamp maxTokens up to 1024 (min context for .litertlm) to fix the DYNAMIC_UPDATE_SLICE crash (#318).
• Honor maxOutputTokens (session + chat) via native set_max_output_tokens; skipped on NPU.

• Fix PreferredBackend.npu on Android (Qualcomm) + Windows (Intel): native-v0.13.1-a restores the NPU dispatch libs omitted from 1.0.0 (#155).
• Point homepage to fluttergemma.dev. No code change.

• Stable 1.0.0; spec imports redirected off the dart:io mobile lib for a wasm-clean web graph.

• Initial release: LiteRT-LM (.litertlm) on-device inference engine for flutter_gemma via dart:ffi.
• Provides LiteRtLmEngine (InferenceEngineProvider). Owns the shared LiteRT-LM native library.
• Android, iOS, macOS, Linux, Windows + web (@litert-lm/core, early preview).