llamadart

Add LlamaEngine.scoreNextToken(...) for next-token log-probabilities on
native llama.cpp and WebGPU bridge assets v0.1.52+, matching llama-server
n_probs; check
supportsNextTokenScoring first
(#694).

Add example/laya_command_bar, a Flutter text field that reshapes into a
reminder, message, calculation or other command as you type, read by a
Laya decision model, by EmbeddingGemma and labelled examples, or by small
LLMs' next-token scores, including the decision model decider-2b.

Throw LlamaModelException when native llama.cpp cannot find or load a
multimodal projector, and LlamaUnsupportedException when the runtime lacks
the mtmd functions; LlamaEngine.supportsAudio also throws the latter.
Speech-to-text capabilities now say when no projector is loaded, using the
new LlamaEngine.hasMultimodalProjector
(#325).

Honour LlamaEngine.cancelGeneration() issued right after listening to a
create, generate or ChatSession.create stream, before it reaches the
backend, instead of running the whole generation
(#602).

Cancel an active text-to-speech synthesis on LlamaEngine.unloadModel() and
dispose() instead of waiting for it to finish
(#628).

Cancel an active Qwen3-ASR transcription on LlamaEngine.unloadModel() and
dispose() instead of completing it with the transcript cut at the unload
(#670).

Send LiteRT-LM tool calls and tool results in the runtime's own message
format, so Gemma 4 reads tool output and Qwen3 tool histories no longer
fail
(#681).

Start a native llama.cpp generation requested right after a cancel once the
cancelled run stops, instead of failing with generation is already in progress. An overlap with a running generation that was not cancelled now
throws LlamaStateException
(#655).

Render Qwen3 prompts as llama.cpp does: an earlier assistant tool-call
turn without reasoning no longer gets an empty <think> block
(#691).

Require dinja 1.1.0. Its Jinja string comparison makes three more chat
templates render as llama.cpp does: MiniMax-M1 adds no empty
system block for an empty or whitespace-only system message; NVIDIA
Nemotron Nano v2 drops the blank line before a tool call, the blank lines
before its tool instructions when tools come with an empty or
whitespace-only system message, and an empty final assistant turn without
a generation prompt; and Functionary v3.2 tool declarations drop stray
// Format=<|NONE|> lines and spell out nested object parameters
(#351).

Cancel a generation's backend run as soon as its stream subscription is
cancelled, instead of at its next token, which during prompt evaluation
meant after the whole prompt. A native llama.cpp generation requested right
after such a cancel now waits for it instead of throwing
LlamaStateException, and native llama.cpp sees a cancel between text
prompt micro-batches (ModelParams.microBatchSize, 512 tokens by default)
or, with speculative decoding, between batches (ModelParams.batchSize)
(#663,
#660).

Render every result of a tool message holding several
LlamaToolResultContent parts, as one tool message per result like
llama.cpp, instead of only the first; LlamaChatMessage.toJson lists them
all (#683).

Render Gemma 4 tool calls and tool results as llama.cpp does, so Gemma 4
GGUF models can read tool output
(#669).

Stop a Qwen3-TTS audio decode at its next chunk boundary when native
text-to-speech is cancelled, instead of finishing the native step in
progress first. This needs llamadart-native v0.4.1-1 or later; older
runtimes keep the previous behaviour
(llamadart-native#86,
#322).

Add an experimental DecisionEngine for Laya-style decision models (a
ModernBERT encoder GGUF plus a safetensors head) on native llama.cpp, with
typed ChoiceKey, ScoreKey and NoulKey questions
(#604).

Add example/basic_app/bin/llamadart_decision_example.dart, a console demo
that triages a support ticket with DecisionEngine
(#604).

Add example/laya_tetris, a Flutter app in which a Laya decision model
plays real-time Tetris through DecisionEngine
(#604).

Run example/laya_tetris on Web through the WebGPU bridge, with a live demo
at https://leehack-flutter-laya-tetris.static.hf.space.

Add a notebook in example/laya_tetris/training/ that fine-tunes a Laya
decision head for the Tetris example and exports it for DecisionEngine
(#604).

Run DecisionEngine on WebGPU through the bridge decision API
(apiVersion 1), which bridge assets v0.1.47+ include
(#604).

Stop native image and audio requests from seeding the repeat penalty with
leftover memory, which made output depend on the previous request
(#603).

After a failed native prompt decode, the next reusePromptPrefix request no
longer runs on the wrong KV cache or keeps failing
(#601).

Reject embed() and embedBatch() on rank-pooled reranker GGUFs with
LlamaUnsupportedException on native, instead of returning memory read
past llama.cpp's classifier-score buffer
(#583).

Throw LlamaInferenceException from native embed() and embedBatch()
when input to an encoder-only model or a model without a KV cache (such as
BERT-family and ModernBERT GGUFs) does not fit one microBatchSize pass,
instead of aborting the process or embedding only the last chunk
(#607).

Fail Apple builds before native symbol lookup when the resolved llama.cpp
companion does not match the core native runtime, with an actionable upgrade
diagnostic instead of allowing ABI-incompatible frameworks.

Updated the default llama.cpp native runtime pin to
leehack/llamadart-native@v0.4.0, regenerated matching Dart FFI bindings, refreshed
the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and
aligned current README/website native override docs. Updated multimodal calls
for the matching ABI. Saved native sessions from older runtimes must be
regenerated. Apple companion 0.0.18 supplies the matching native runtime.

Aligned the default WebGPU bridge assets to v0.1.43 for Web/native
llama.cpp v0.4.0@5266f24da75dc449bd56cbed7addb9c8e4a6a73e parity.
Web @litert-lm/core@0.15.0 and native LiteRT pins are unchanged. Immutable
manifest: 111eefc3588842cebfe665b363378edca34924764610263e1eda5280dfcfaa27.

Known upstream limitation: llama.cpp v0.4.0 can reject large grammar
repetitions, such as root ::= "a"{2000}. The post-v0.4.0 upstream
correction is tracked in llamadart-native#76 and is not part of this release.

Updated llamadart_llama_cpp_flutter to 0.0.17 with the
Apple SwiftPM v0.3.0 runtime pin.

Updated the default llama.cpp native runtime pin to
leehack/llamadart-native@v0.3.0, regenerated matching Dart FFI bindings, refreshed
the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and
aligned current README/website native override docs.

Aligned the default WebGPU bridge assets to v0.1.41 for corrected
TypeScript declarations and TTS recovery guidance, retaining Web/native
llama.cpp v0.3.0@c1d0e7a004015f23bc0233470b747b596f29b264 parity and Web
@litert-lm/core@0.15.0. Immutable manifest:
fe97604daabaad6aefa223a8637d5fd9dcac09dd4a61b2ef19cd6aabb39392b9.

Consolidated native release tag grammar across Dart, Python, Bash, workflows,
and documentation via a machine-readable fixture contract (#404).

Aligned the default WebGPU bridge assets to v0.1.39 (immutable manifest
b355d01040604f6ae2c5c5fe5bb42b858101a96f03f67e4b27b32fe41ce3b2bf),
restoring Web/native llama.cpp upstream v0.2.0@bb4caa7540188872173c44d161602d9271386413
parity with native anchor llamadart-native@v0.2.0-1 while preserving
approved Web @litert-lm/core@0.15.0 packaging.

Fixed native Qwen3-ASR transcription by applying the model chat template to
audio turns, while preserving the validated raw-prompt Web bridge contract.
Empty ASR output now fails explicitly instead of reporting an empty result.

Fixed Qwen 2.5/3 LiteRT-LM ToolChoice.required requests silently running
without their required-call grammar and finishing with no call. They now fail
with an actionable LlamaUnsupportedException before generation when the
backend cannot enforce the declared tool schema; Gemma 4 compatibility and
auto/none tool routing are unchanged.

Fixed Gemma 4 thinking-budget output so split channel controls and tool-call
envelopes stay out of visible assistant content while preserving ordinary
whitespace. The chat example now also validates custom tool declarations,
executes declared host handlers exactly once, appends tool results, and
performs a bounded continuation for both llama.cpp and LiteRT-LM backends.

Patched the website's vulnerable nanoid and uuid dependency paths. Until
Docusaurus replaces its unpatched image parser, automatic local Markdown
images are rejected; website contributors should use static pathname URLs.

Fixed llamadart_native_runtimes values none, off, and the string
false selecting every runtime family instead of none;
the build hook fails with its No native runtimes selected error again, as
it did before 0.8.0. A YAML boolean false clears the selection too. Unset,
empty, and all-unrecognised config still select every family.

The published package no longer ships the doc/ directory; that contributor
and maintainer documentation is maintained on GitHub, and the packaged files
that link to it now use absolute URLs.

Fixed unanchored docs/ and website/ publish-exclusions that matched those
directory names at any depth and dropped tool/docs/ plus the
llamadart_server example's OpenAPI spec and Swagger UI sources from the
package, leaving the published example unable to analyze. Both patterns are
now root-anchored.

Narrowed the dinja dependency constraint to >=1.0.0 <1.1.0 so chat
template capability detection cannot silently resolve against an unverified
Jinja parser minor. A 1.0.x patch can still reorganise the private sources
the analyzer imports; a new coupling test turns that into a named failure.

Fixed Command R7B, Hermes, and Hunyuan V3 tool grammars so distinct tool or
parameter names cannot collide after conversion to internal GBNF rule names.

Fixed DeepSeek V3.2 DSML tool calls using their upstream
<｜DSML｜function_calls> envelope while preserving DeepSeek V4's distinct
<｜DSML｜tool_calls> grammar and parser behavior.

Fixed partial GLM 4.5, Poolside Laguna, and Muse Glimmer tool envelopes
leaking into streamed assistant content, while preserving completed calls,
malformed final output, and ordinary text surrounding Muse recipient
channels.

Fixed schema-constrained tool calls for Kimi K3, MiniMax M1/M3, DeepSeek
V3.2/V4, and Muse Glimmer, including exact escaped names, required fields,
declared value types, matching MiniMax M3 element tags, zero-argument calls,
and strings containing delimiter characters. Required-tool mode now accepts
each format's reasoning/content prefix while still requiring a call.
MiniMax M3, DeepSeek DSML, Muse Glimmer, Poolside Laguna, and GLM 4.5 now
reconstruct argument values from the declared tool schema instead of
guessing from text. Added ToolParam.nullType for null-only JSON Schema
properties.

llama.cpp backend initialization failures now complete the worker startup
handshake with a typed LlamaBackendInitializationException and collected
native-loader diagnostics. A failed or incompatible worker is torn down
instead of being reported ready and leaving later requests waiting forever.

Added template-aware parsing for Kimi K3, MiniMax M1/M3, DeepSeek V3.2/V4,
Muse Glimmer, and Poolside Laguna, preventing their native tool calls from
silently falling back to plain content.

Fixed XML-style tool-call parsing to honor raw-versus-JSON argument values
and final-value delimiters. Apriel 1.5 and Xiaomi MiMo now parse nested JSON
values without splitting on inner commas, while malformed payloads remain
ordinary assistant content.

Made native video-input capability truthful without claiming end-to-end
support. LlamaVideoContent requests now fail with an actionable
LlamaUnsupportedException, LlamaEngine.supportsVideo reports public
consumability as false, and the llama.cpp worker uses the behavioral
mtmd_helper_support_video result instead of exported helper symbols. The
published v0.2.0-1 archive has not been qualified for end-to-end video;
full path/byte input remains blocked on cross-platform FFmpeg/ffprobe
packaging and Dart frame lifecycle wiring.

Native release synchronization and build-hook overrides now accept stable
vMAJOR.MINOR.PATCH artifacts and ordered vMAJOR.MINOR.PATCH-N wrapper
rebuilds plus nightly bNNNN-N rebuilds, while preserving historical
bNNNN and bNNNN-llamadart.N artifacts. Sync rejects invalid tags,
leading-zero nightly versions, rollback, wrapper/nightly latest results,
incompatible manifest contracts, missing bundles, and release/manifest
checksum or version skew without changing the default pin.

Fixed Web/native backend API parity. WebAutoBackend now forwards grammar
constraint support from its active runtime, so strict structured output fails
early with an actionable error on unsupported Web backends, and the Web-safe
LiteRtLmRuntimeClient stub now exposes the native client's thinking-tag
configuration method.

A failed llama.cpp model load now reports the startup diagnostics collected
during native library discovery, so a missing or unloadable runtime library
explains itself instead of surfacing as a bare load failure. Platforms that
record no diagnostics keep their previous message unchanged.

llama.cpp worker errors now keep their type. Every backend method routes an
ErrorResponse through the file's own error mapper instead of rebuilding a
bare Exception, tokenize and detokenize no longer discard the worker's
error entirely, and a core UnsupportedError raised for an unavailable native
capability is classified rather than flattened. State-file failures now throw
LlamaStateException.

Fixed multimodal media placeholders being normalized inconsistently. <img>,
<|img|>, <start_of_image> and indexed markers such as <|image_1|> are now
rewritten to the mtmd marker on every path, rather than depending on which
layer rendered the prompt. MiniMax-M2 and MiniCPM-5 also now detect a
forced-open thinking block using the same rule as every other handler.

aLoRA adapters are now rejected with LlamaUnsupportedException instead of
being applied like ordinary LoRA adapters. An aLoRA adapter must activate only
after its invocation tokens appear in the prompt, so applying it from the
start of generation silently changed output. Missing metadata-inspection
symbols in custom native runtimes also fail closed with the same typed error,
and rejected adapters are released when the cleanup ABI is available. LoRA
errors from the worker keep their typed exception instead of arriving as a
bare Exception, and a failed adapter load now throws
LlamaModelException.

Deprecated LiteRtLmRuntimeClient.conversationTokenCount() and
replaceConversationWithClone(). Both are unused and are scheduled for
removal in the next major release; open an issue if you depend on either.

NativeLlamaBackend.modelLoadFromUrl now throws LlamaUnsupportedException
instead of UnimplementedError, bringing it into the LlamaException
hierarchy. It is the same exception type LlamaEngine.loadModelFromUrl
already throws for this condition; each keeps its own message.

Updated the default llama.cpp native runtime to the immutable
leehack/llamadart-native@v0.2.0-1 release (llama.cpp v0.2.0), adding LFM2
DSpark support plus current upstream correctness and backend performance
fixes. Matching Dart FFI bindings, including the new multimodal
projector-device field, and the Apple SwiftPM artifact checksum were
refreshed. Linux libmtmd.so.0 now loads without the old
libmtmd.so.SOVERSION compatibility alias.

Removed the abandoned Dart-side MTP/n-gram speculative-decoding scaffolding
from the llama.cpp backend. Speculative decoding is unchanged: it continues to
run through the native wrapper, and SpeculativeDecodingStrategy keeps every
existing option.

Bumped llamadart_llama_cpp_flutter to 0.0.16 so the v0.2.0-1 Apple
SwiftPM pin actually publishes; 0.0.14 was already on pub.dev, so release
automation skipped it and Apple builds would have kept the b10514 runtime.

Corrected the WebGPU bridge docs, which claimed the pinned v0.1.37 bridge
assets match the default native llama.cpp runtime. They embed b10514 and now
trail the native v0.2.0-1 pin.

Chat-template capability detection now logs a debug message naming the
probe (string-content, typed-content, system-role, tools) when its
render throws, so a template that fails to render is distinguishable from
one that genuinely lacks the capability.

Added code_assets 2.x compatibility while retaining support for 1.x native
asset toolchains.

Updated the default llama.cpp native runtime pin to
leehack/llamadart-native@b10514, adding BailingMoE3,
GraniteSWA/GraniteMoeSWA, speculators-format DSpark checkpoints, and current
upstream multimodal/backend improvements. Matching Dart FFI bindings and the
llamadart_llama_cpp_flutter Apple SwiftPM artifact were refreshed.

Updated the default WebGPU bridge assets to v0.1.37, embedding llama.cpp
b10514 to restore native/Web parity. The bridge provisions an explicit 1 MiB
Wasm stack for wasm32 and memory64, preventing Qwen3-ASR memory64 context
construction from overflowing the default stack.

Improved Web microphone transcription by warming up browser capture before
showing the recording-ready state, trimming the warmup silence, and
rejecting too-short, silent, or unsupported PCM WAV captures before
inference.

Added logical batch-size (n_batch) and micro-batch-size (n_ubatch)
controls for llama.cpp/WebGPU models in the Flutter chat example.

Improved Android Auto backend selection by probing the packaged Vulkan device
before memory planning, avoiding unnecessary CPU fallback on capable models.

Updated the native LiteRT-LM runtime to v0.16.0-native.2. The Apple companion
packages the iOS Gemma constraint provider and Metal accelerator/sampler
plugins required by the published runtime.

Added an experimental SpeechToTextEngine.liteRtLm path with capability
discovery, worker-isolated CPU inference, bounded mono 16 kHz float PCM,
partial/final transcript events, finalization, and cancellation.

Added experimental live dictation to the native Flutter chat example for
chat models using selectable Moonshine Tiny and Parakeet TDT 0.6B sidecars.
Live dictation is CPU-only, English-only, capped at five minutes, and
unavailable on Linux and Web.

Improved Flutter chat example model downloads with bounded retries for
transient network failures, safe resume after truncated responses, and a
distinct integrity-verification state after transfer reaches 100%.

Redesigned the Flutter chat example onboarding and Lab surfaces, preserved a
completed model card's viewport position when downloads reorder the catalog,
and stopped streaming responses from pulling users away from chat history.

Added an experimental typed TextToSpeechEngine for native llama.cpp and
WebGPU with Qwen3-TTS models, including capability discovery, speaker
references, cancellable progress, complete 24 kHz PCM output, and WAV
encoding. The Flutter chat example adds synthesis, playback, replay, and WAV
save controls. Apple apps discover the TTS ABI in the embedded llama
framework; current LiteRT-LM artifacts fail explicitly as unsupported.

Added an experimental typed SpeechToTextEngine with an explicit Qwen3-ASR
adapter profile for whole-file llama.cpp transcription. The Flutter chat
example includes a checksum-pinned Qwen3-ASR 0.6B preset plus file and
microphone transcription on native and WebGPU. Web accepts WAV bytes only;
native LiteRT-LM live dictation remains a separate implementation.

Added Ask with voice to the native Flutter chat example for Gemma 4 E2B
through LiteRT-LM direct media and audio-capable GGUF + projector paths. It
sends a microphone recording through multimodal chat and remains separate
from typed speech-to-text.

Added experimental, opt-in native llama.cpp DSpark speculative decoding
through SpeculativeDecodingConfig.draftDspark(...) with a compatible
external draft model.

Updated the default llama.cpp native runtime pin to
leehack/llamadart-native@b10333, regenerated matching Dart FFI bindings,
refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and aligned
current README/website native override docs.

Updated WebGPU bridge assets to v0.1.27 (llama.cpp b10333), keeping the
native and Web GGUF runtimes on the same upstream revision.

Fixed corrupt Qwen3.5 output on Android Vulkan by preserving the KQV
offload required for correct hybrid model inference while retaining the
remaining conservative Android context settings.

Updated the default llama.cpp native runtime pin to
leehack/llamadart-native@b10276, picking up Qwen3-TTS model-loading
primitives, explicit bundled-MTP loading, automatic model-specific token
suppression, and recent model, multimodal, speculative-decoding, backend,
and runtime fixes. Regenerated matching Dart FFI bindings, migrated model
loading to llama.cpp's load-mode ABI and the penalty sampler to the new
vocabulary-sized ABI, and refreshed the llamadart_llama_cpp_flutter Apple
SwiftPM checksum. Speech generation is not yet exposed through the public
Dart API.

Updated the default LiteRT-LM runtimes to
leehack/litert-lm-native@v0.15.0-native.3 and
@litert-lm/core@0.15.0. The native artifact includes a corrected v0.15
streaming callback bridge and an Android Dawn rollback for Mali-G715 GPU
device loss; incompatible callback runtimes now fail safely before
generation, and concrete macOS app, framework, and cache libraries take
precedence over process-linked assets.

Disabled automatic WebGPU fetch-backed model loading by default. Streamed
loading remains the safe default; controlled range-capable deployments can
opt in explicitly.

Updated WebGPU bridge assets to v0.1.26 (llama.cpp b10276), refreshing
both WebAssembly runtimes while preserving the existing bridge API.

Updated the default llama.cpp native runtime to
leehack/llamadart-native@b10075 with matching bindings and Apple artifacts.

Added Tencent Hunyuan V3 chat-template, reasoning, and tool-call support.

Fixed Gemma 4 LiteRT-LM text generation in the Web chat app after model
loading completed successfully.

Restored GGUF loading in deployed Web chat apps by packaging the pinned
WebGPU runtime assets with Flutter Web builds.

Updated the default llama.cpp native runtime to leehack/llamadart-native@b9982, including regenerated bindings and safer multimodal UTF-8 prompt handling.

Improved llama.cpp batching defaults and ChatSession context management for more predictable generation under constrained contexts.

Added llama.cpp presencePenalty sampling and thinking-budget controls, including the server's thinking_budget_tokens extension. Unsupported WebGPU and LiteRT-LM paths now fail explicitly.

Reworked the runnable TUI coding agent with a focused Pi-style workflow, Unsloth Qwen3.6 defaults, shared model-source loading, and clearer reasoning and final-answer presentation.

Improved the OpenAI-compatible server with standard client-managed tool-call transcripts, named tool_choice, configurable thinking behavior, and shared model-source loading.

Added clipboard media attachments to the runnable chat app. Desktop and web users can paste screenshots or copied image/audio files with Cmd/Ctrl+V, mobile users can choose Paste attachment, and ordinary text paste remains unchanged.

Restored the runnable macOS chat app build phase that embeds and signs LiteRT-LM runtime libraries inside the sandboxed app bundle, and enabled LiteRT-LM Metal selection on iOS with the consolidated upstream runtime. Updated the default LiteRT-LM runtime to leehack/litert-lm-native@v0.14.0-native.2, which fixes Android GPU plugin symbol resolution and uses the checksum-pinned official Apple XCFrameworks.

Fixed native LiteRT-LM generation incorrectly treating the requested maximum response length as a forced benchmark decode count. Short responses no longer wait for every allowed token before streaming, and LiteRT-LM chat flushes its first token immediately.

Added an app-owned FIFO model-download queue to the runnable chat app. Only one model transfers at a time, queued cards show their position and can leave the queue independently, and a responsive shell progress pill keeps the active download visible after the settings panel closes.

Replaced the runnable chat app's broad built-in model catalog with a focused Unsloth-first set. Added cross-platform Gemma 4 E4B plus native-desktop Gemma 4 12B/26B-A4B/31B and Qwen3.6 35B-A3B presets. The model library now promotes downloaded models, supports name/capability search and Mobile & Web/Desktop filters, explains incompatible choices, and uses quieter cards with compact compatibility, capability, and recommended-setting summaries. Custom entries retain independent remove-from-library and downloaded-file actions.

Enabled Gemma 4 audio attachments in the runnable chat app for the current native GGUF projector and LiteRT-LM bundle, while keeping LiteRT-LM Web correctly text-only. Model capability settings now distinguish direct media input from external mmproj input and persist that distinction across app launches.

Redesigned the runnable chat app with a quieter responsive shell, compact runtime details, full-screen mobile settings, streamlined message/composer surfaces, accessible controls, protected conversation deletion, and copy/regenerate actions for assistant responses. Model downloads now remain active when settings closes or switches between drawer and pinned layouts.

Hardened the runnable chat app for narrow windows, 200% text scaling, and macOS assistive technology; removed stale assistant placeholders after empty or failed generations; kept context accounting stable when regenerating a response; and redacted signed model URL parameters from labels and generation errors.

Clarified the chat app's backend and GPU-layer controls, exposed the active loaded backend beside its preference, and restored GPU offload when switching from a zero-layer CPU configuration back to Auto or a GPU backend. Native Max now maps to full llama.cpp offload, while Auto uses model size, available device memory, safe system headroom, and requested context to choose full or partial offload and only reduces context when the model does not fit. Auto intent now persists separately from resolved layer/context values so device headroom is recalculated on every model load and after app restarts.

Updated the runnable chat app's web runtimes to pinned WebGPU bridge assets v0.1.18 (llama.cpp b9915) and @litert-lm/core@0.14.0, keeping hosted and local inference on reproducible runtime versions.

Improved the runnable chat app's model-cache UX by reusing cached GGUF and LiteRT-LM web bundles for benign catalog URLs such as ?download=true, preserving browser model caches during app cache cleanup, adding a text-only projector skip path, and polishing download/load progress states.

Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9935, regenerated matching Dart FFI bindings, refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and aligned current README/website native override docs.

Fixed load lifecycle guards so repeated model loads preserve the active model state, URL load unsupported-runtime diagnostics stay typed, and unload cancels active generation before freeing llama.cpp handles.

Tightened LiteRT-LM runtime validation and local smoke coverage by requiring complete macOS arm64 runtime caches, requiring an explicit model path for the LiteRT-LM chat feature smoke scenario, normalizing the chat app's LiteRT-LM auto context size, and adding the missing iOS-compatible SwiftPM Gemma provider target. Flutter macOS LiteRT-LM companion-package builds now fall back to hook-managed native assets, and the companion package no longer links the incomplete macOS LiteRT-LM SwiftPM artifact set.

Reworked the README into a shorter entry point, fixed stale docs/examples found during the documentation review, and aligned release, Android smoke, WebGPU mem64, native sync, and capability-support wording with the current workflows and runtime behavior. WebGPU runtime LoRA calls now throw an unsupported-operation error instead of reporting no-op success.

Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9873-llamadart.2, keeping the b9873 llama.cpp ABI/bindings while picking up wrapper fixes for native release provenance and backend-selected speculative sampler acceptance. Refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum and aligned current README/website native override docs.

Aligned llama.cpp n-gram speculative rejected-tail rollback with upstream server behavior by trimming rejected target tokens before falling back to checkpoint replay, avoiding early token-count/EOG drift in accepted-draft ngram-map-k runs. The speculative benchmark can now emit full generated text with --include-output for parity investigations.

Fixed llama.cpp n-gram speculative configuration mapping so draftTokenMax no longer implicitly overrides upstream ngramSizeM, and documented upstream comparison commands plus measured n-gram benchmark results.

Added a durable flag-based llama.cpp speculative benchmark runner and wired it into the local E2E scenario list, testing matrix, README, and website performance guide. The runner can now generate a llama.cpp-compatible static n-gram cache file for ngram-cache E2E validation, and renders benchmark prompts with configured or loaded GGUF chat templates before the generic fallback instead of silently falling back to a hard-coded Gemma prompt.

Documented compatible DFlash GGUF metadata, a known-good public target/draft model pair, and troubleshooting guidance for incompatible dflash-draft or missing dflash.target_layers artifacts.

Hardened generic llama.cpp external draft-model speculative decoding so draft-context processing does not request unused logits, avoiding a draft-simple native abort while preserving target verification logits.

Extended llama.cpp speculative performance diagnostics with draft-attempt, target-verification-token, and replay-token counters. Enabled speculative runs now preserve explicit zero counters instead of collapsing them to null, and local benchmark JSON includes the new fields.

Hardened LiteRT-LM generation validation so llama.cpp-only speculative decoding knobs fail loudly instead of silently degrading to LiteRT-LM's boolean speculative toggle.

Added llama.cpp upstream speculative decoding parity through SpeculativeDecodingConfig constructors for draft-simple, EAGLE3, MTP, DFlash, ngram-simple, ngram-map-k, ngram-map-k4v, ngram-mod, ngram-cache, and mixed n-gram plus one draft-model strategy, including generic native wrapper bindings, docs, and local benchmark matrix coverage.

Added LlamaStructuredOutput and LlamaEngine.createStructuredJson(...) helpers for strict JSON-object / JSON-schema generation with final-output validation and typed decoding.

Added LlamaEngine.loadMultimodalProjectorSource(...) so GGUF multimodal projector files can use the same ModelSource resolver, native download/cache manager, authentication, checksum, and progress options as loadModelSource(...), while preserving the existing loadMultimodalProjector(...) path/string API.

Improved the runnable chat app's Manage Models cache UX so model and mmproj asset cache states are shown separately, missing multimodal projectors can be re-cached without re-fetching already cached model assets, and runtime media capability mismatches surface as user-readable warnings. Custom signed or tokenized Hugging Face URLs now require confirmation before they are saved.

Updated the default LiteRT-LM native runtime pin to leehack/litert-lm-native@v0.14.0-native.1, refreshed matching native-assets checksums, aligned the llamadart_litert_lm_flutter Apple SwiftPM checksums, and documented the newly exposed native LiteRT-LM load/generation controls: per-request max output tokens, native sampler params, thread count, one default-scale initial text LoRA adapter, activation data type, prefill chunk size, parallel file-section loading, and Android LiteRT dispatch library directory.

Hardened LiteRT-LM runtime packaging and local runtime preparation for the 0.14 line, including Linux/Windows runtime dependency resolution, Android Dawn companion libraries, Apple SwiftPM checksums, and macOS fallback app runtime copying.

Hardened release automation by adding CODEOWNERS coverage for publication-sensitive files and making pub.dev/GitHub Release propagation waits configurable with longer defaults.

Added post-merge release automation so a merged release-prep PR can publish missing companion package versions, push the core release tag, wait for pub.dev, and confirm the GitHub Release without a separate manual tag step.

Added LlamaEngine.getModelFileType() for llama.cpp/GGUF models, exposing native model file type / quantization metadata from llama_model_ftype and llama_ftype_name when available.

Refreshed llama.cpp b9860 template parity for DeepSeek V4 and MiniCPM5, including MiniCPM5 XML tool-call detection, rendering, and parsing.

Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9860, regenerated matching Dart FFI bindings, refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and aligned current README/website native override docs.

Potentially breaking behavior change: native model cache defaults changed without breaking Dart source compatibility. DefaultModelDownloadManager() now prefers the platform shared cache on desktop/server instead of the process temp directory, and mobile DefaultModelDownloadManager.auto() without an explicit app-private directory now uses a best-effort temporary/cache fallback instead of throwing. Apps or tests that asserted the old temp path or mobile exception should pass an explicit cache directory or follow MIGRATION.md.

Added optional androidAppPrivateCacheDirectory and iosAppPrivateCacheDirectory arguments to DefaultModelDownloadManager.auto(...) so apps can provide platform-specific mobile cache roots without constructor-level Platform.isAndroid / Platform.isIOS branching.

Updated the default native DefaultModelDownloadManager() constructor to use the per-user shared model cache on desktop/server platforms and the mobile app-private cache fallback, so plain LlamaEngine(...) remote source loads use a platform-appropriate default while preserving a temporary fallback for hosts that cannot expose a desktop cache environment.

Broadened the hooks dependency constraint to support both the existing build-hooks package family and the latest stable release, restoring the pub.dev dependency freshness score without breaking downstream packages that still resolve hooks 1.x.

Made web-safe backend stubs the default conditional import/export targets, preserving native dart:io selection while avoiding false WASM compatibility deductions in pub.dev analysis.

Added a CI release-doc version consistency check so current README/website install snippets and companion package READMEs stay aligned with package pubspec.yaml versions, and documented that companion/core package publishing happens only after release-prep merge with explicit maintainer approval for each tag.

Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9803, regenerated matching Dart FFI bindings, refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and aligned current README/website native override docs.

Added DefaultModelDownloadManager.auto(...) plus explicit model cache root constructors for shared desktop caches, app-private mobile caches, user-selected model libraries, and App Group containers. Implicit shared cache resolution now fails loudly on mobile and web where the OS cannot provide a hidden cross-developer model folder.

Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9744, regenerated matching Dart FFI bindings, refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and aligned current README/website native override docs.

Expanded llama.cpp chat-template parity coverage for the latest upstream template fixtures, including Cohere2 MoE, LFM2.5 tool-call, and Granite 4.1 templates. LFM2.5 templates that use plain List of tools: [...] prompts with <|tool_call_start|> / <|tool_call_end|> now route through the LFM2 handler like upstream llama.cpp, and ToolChoice.required now uses grammar-constrained LFM2 tool-call generation.

Fixed streaming tool-call parsing so partial North/Cohere bare action arrays are not emitted as content before the complete tool call is parsed.

Expanded the local GGUF chat feature smoke to cover thinking, tool-call, and optional multimodal turns through the unified local E2E runner.

Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9694, regenerated matching Dart FFI bindings, refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum, and updated the default WebGPU bridge asset pin to leehack/llama-web-bridge-assets@v0.1.17 (llama.cpp b9699). The WebGPU backend now caps unset large-model browser batches so Gemma 4 mem64 loads do not fall back to context-sized compute buffers.

Added BackendGpuEnumeration.listGpuDevices({probeBackends}) (exposed via LlamaEngine.listGpuDevices) to enumerate GPU-class devices for offload selection — backend, per-backend mainGpu index, name, description, device id, type, and free/total memory per device. With an empty probeBackends only already-registered backends are inspected, so an unsupported GPU runtime cannot crash the process during enumeration; pass specific backends to opt into loading just those modules first. Web/WebGPU return an empty list.

Added Cohere2 MoE / North Code chat-template detection and parsing so <|START_TEXT|> responses and <|START_ACTION|> tool-call arrays are handled separately from older Command-R templates.

• Split SwiftPM-linked Apple runtime packaging out of the core package into llamadart_llama_cpp_flutter for GGUF/llama.cpp and llamadart_litert_lm_flutter for .litertlm/LiteRT-LM. These companion packages live under packages/ in this repository and publish as separate pub.dev packages.
• Removed Flutter plugin metadata from llamadart so pure Dart/native-assets consumers can keep using the core package without taking a Flutter SDK constraint.
• Started the companion packages at 0.0.1; native pin sync bumps only the affected companion package patch version. Companion package publishing uses package-specific tags after the first manual pub.dev publish, and skips companion versions that already exist on pub.dev.
• Changed unset or empty llamadart_native_runtimes to include all available runtime families. For Flutter iOS/macOS app builds, installed companion packages decide Apple SPM runtimes; for every other build, llamadart_native_runtimes remains the selector.
• Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9587, regenerated matching Dart FFI bindings, and refreshed the llamadart_llama_cpp_flutter Apple SwiftPM checksum.

• Added llama.cpp speculative decoding perf diagnostics for decode timing, draft/accepted token counts, draft verification timing, and acceptance rate so MTP benchmarks can separate backend decode cost from drafting overhead.
• Extended local macOS and chat app benchmark outputs with the new diagnostics and added focused llama.cpp MTP smoke/benchmark tools for baseline-vs-MTP comparisons.

• Added SpeculativeDecodingConfig.mtp(draftModelPath: ...) for llama.cpp external draft-model MTP sessions, with draft model caching and cleanup tied to the target model lifetime.
• Removed the Android Vulkan MTP allow-list dart define and the model-name based Android Vulkan acceleration shortcut; Vulkan MTP now runs only when callers explicitly request Vulkan plus MTP in runtime parameters.

• Added responseFormat routing to LlamaEngine.create(...) for grammar-capable backends, deprecated the legacy chatTemplate(...) jsonSchema shortcut, and made strict response-format requests fail early on LiteRT-LM instead of silently degrading to unconstrained generation.

• Routed eligible native .litertlm text chat through LiteRT-LM Conversation APIs so structured history, system messages, tool declarations, and template extra context reach the runtime without a Dart-rendered prompt. Unsupported cases still fall back to the existing Dart chat-template path.

• Added opt-in native .litertlm ModelParams for liteRtLmActivationDataType, liteRtLmPrefillChunkSize, liteRtLmParallelFileSectionLoading, and liteRtLmDispatchLibDir, forwarding the pinned LiteRT-LM v0.13.1-native.1 engine-settings C APIs while keeping defaults unchanged.
• Extended the LiteRT-LM engine smoke tool with matching environment variables so real-model runs can validate load time, prefill throughput, decode throughput, and selected runtime settings.
• Documented support decisions for each candidate native knob and kept LiteRT-LM web rejecting these native-only settings explicitly.

• Added Flutter iOS/macOS Swift Package Manager integration so Apple apps link the pinned leehack/llamadart-native and leehack/litert-lm-native XCFramework artifacts through darwin/llamadart/Package.swift.
• Disabled the legacy hook-managed Apple bundle path for Flutter iOS/macOS builds, avoiding wrapper/framework MinimumOSVersion mismatches in App Store uploads. Standalone Dart macOS keeps the native-assets dylib fallback.
• Raised the Flutter Apple runtime floors to iOS 16.4 and macOS 14.0 to match the published XCFramework artifacts.

• Android native builds still include both llama_cpp and litert_lm by default; iOS, macOS, Linux, and Windows now default to llama_cpp only.
• Added native release pin automation so the maintainer sync workflow updates Apple SPM checksums from published native release asset digests.
• Added SpeculativeDecodingConfig as a backend-neutral generation option for selecting speculative decoding strategies such as MTP while keeping the existing GenerationParams.speculativeDecoding flag as a compatibility switch.
• Added llama.cpp native MTP speculative decoding for compatible GGUF models through SpeculativeDecodingConfig.mtp(...), defaulting to a conservative one-token draft depth unless callers tune draftTokenMax.
• Updated the default llama.cpp native runtime pin to leehack/llamadart-native@b9547, including the MTP wrapper exports and llama-common runtime packaging.
• Added ModelParams.speculativeRollbackTokenMax so llama.cpp contexts can reserve recurrent-state rollback snapshots required by Qwen3.5 MTP-style models.
• Guarded llama.cpp MTP on Android Vulkan by default because the upstream draft-mtp backend-sampling path can abort with vk::DeviceLostError; CPU and other supported backends remain available, and a dart-define debug override is available for reproductions.

• Cached and retried tiny GGUF test-model downloads used by VM integration tests so main-branch CI is less exposed to Hugging Face 429 rate limits.
• Excluded local SwiftPM artifact caches from pub archives; Flutter Apple consumers still resolve the published remote XCFramework targets.

• Added first-class .litertlm routing through LlamaBackend() on native and web targets, with native bundle downloads from leehack/litert-lm-native and web loading through @litert-lm/core.
• Added ModelParams.liteRtLmBackend so callers can select LiteRT-LM CPU, GPU, or Android NPU execution where supported. auto chooses GPU on Android/macOS and CPU elsewhere on native targets.
• Added cached Hugging Face .litertlm loading through loadModelSource(...), preserving the selected LiteRT-LM backend after the cache manager resolves the local file.
• Added native LiteRT-LM tokenization, detokenization, log-level control, runtime metrics, and high-level ChatSession token counting support.
• Added hooks.user_defines.llamadart.llamadart_native_tag, llamadart_native_repository, and llamadart_native_path so apps can test a different compatible native runtime source without patching llamadart.
• Updated the Windows runtime fallback scanner to discover custom GitHub and local archive cache namespaces when .dart_tool/lib is unavailable.

• Added GenerationParams.speculativeDecoding for native LiteRT-LM. The default remains disabled; llama.cpp, WebGPU, and LiteRT-LM web reject the option until their speculative paths are implemented.
• Fixed Gemma 4 .litertlm thinking and tool calling by replacing the stub template with the canonical Gemma 4 chat template, parsing the runtime thought channel as reasoning, and suppressing reasoning deltas when callers set enableThinking: false.
• Added a filename-keyed .litertlm chat-template registry seeded with Gemma 4/3/3n and Qwen 2.5/3. Pass ModelParams.chatTemplate to override detection for other models.
• Fixed LiteRT-LM tool calling for grammar-using handlers by forwarding supportsGrammarConstraints from the active NativeAutoBackend delegate.
• Stopped structured tool-call streams from leaking raw Hermes/Qwen JSON or Gemma <|tool_call> markers as assistant content before the final tool_calls chunk.

• Added ModelParams.preferMemory64 and ModelParams.modelBytesHint so large WebGPU GGUF models such as Gemma 4 E2B can choose the 64-bit bridge core before hitting the wasm32 address-space limit.
• Fixed web .litertlm chat-app turns by swallowing unsupported token-count refreshes, avoiding unsupported minP/penalty parameters for LiteRT-LM web generation, and replacing the stuck "Loading model 0%" label with an indeterminate load message.
• Halved web .litertlm load time by skipping WebGPU CacheStorage prefetch for LiteRT-LM models, which are fetched directly by @litert-lm/core.
• Fixed web GGUF downloads reporting success before the bridge was ready by awaiting window.__llamadartBridgeReadyPromise, requiring the bridge prefetch API, and surfacing actionable errors for old bridge assets.
• Allowed benign Hugging Face ?download=true URLs to be prefetched into the browser cache while still skipping credentialed or signed URLs.

• Fixed iOS .litertlm loading by resolving embedded LiteRtLm and StreamProxy frameworks from the app bundle, matching the macOS runtime path behavior.
• Improved LiteRT-LM diagnostics before model load, including selected CPU/GPU/NPU backend reporting, platform availability errors, and complete dynamic-library candidate failures.
• Validated platform-specific LiteRT-LM companion libraries during native-asset setup so incomplete runtime bundles fail at build time.
• Hardened native and LiteRT-LM cancellation/disposal so in-flight generation no longer races token release, worker teardown, engine deletion, closed response ports, or stream writes after cancellation.
• Freed multimodal prompt buffers on tokenize/eval error paths, serialized multimodal projector load/unload, and closed a leaked native-backend handshake reply port.

• ChatSession now forwards empty-choices completion chunks instead of throwing, strips multiple <think> blocks, and trims history only on user-message turn boundaries.
• LlamaEngine.generate wraps unexpected backend errors in LlamaInferenceException so callers catching LlamaException see the documented error type.
• Tool-call parsing now uses stable fallback ids, tolerates code-fence language tokens without trailing delimiters, and keeps commas inside quoted argument values.
• JSON-schema-to-GBNF conversion now resolves $refs nested inside other $ref targets and fails loudly on unresolvable or external $refs.
• Array grammar generation validates minItems/maxItems, model downloads use connection and idle-read timeouts, and partial-download resume is restricted to files with stored validators.

• Added fair Gemma 4 LiteRT-LM versus llama.cpp/GGUF benchmark tooling for Android, macOS, and web, with speculative-decoding metrics, Pixel benchmark failure detection, and target-specific timeouts.
• Added tool/gguf_chat_features_smoke.dart and the chat-app-web-gemma4-webgpu-smoke E2E scenario for real-model parser and WebGPU mem64 validation.
• Updated README, website docs, and doc/litert_lm_templates.md for backend selection, platform/runtime support, package-size controls, benchmark results, model templates, and current LiteRT-LM capability limits.

• Updated native hook pinning and regenerated bindings through leehack/llamadart-native@b9371, picking up llama.cpp b9371.
• Picked up the Apple mobile Metal stability fix that disables Metal residency sets on iOS/tvOS/visionOS native bundles, avoiding affected device context-creation failures such as MTLLibraryErrorDomain Code=3.

• Fixed native getVramInfo() so it reports free/total VRAM from llama.cpp GPU-class backend devices when available, using props-based memory reporting first and the legacy memory probe as a fallback.
• Routed native VRAM probing through the ggml registry fallback path so Windows split bundles resolve backend-device symbols from the runtime that owns the device registry.

• Improved browser recovery for large remote WebGPU model/projector loads by retrying wasm32 model-staging aborts with the wasm64 core before surfacing memory-pressure failures.
• Improved the runnable chat app's web remote-model startup path so model assets are prefetched into browser cache when available, browser CacheStorage failures fall back to direct network loading, and credentialed/signed model URLs skip persistent browser cache storage.

• Improved the runnable chat app's mobile download behavior so lifecycle pauses no longer deliberately cancel active foreground downloads; the app now lets short screen-lock/background interruptions continue when the OS permits and still keeps explicit pause/dispose cancellation paths.
• Added in-app and docs guidance for mobile large-model downloads, including resumable partial files, foreground Dart lifecycle limits, and the need for opt-in native background download/model-store integrations for robust cross-app GGUF management.

• Fixed GLM-OCR and other multimodal chat-template workarounds so image and audio content parts are preserved when tool-call normalization runs, system prompts are merged before leading media parts, and invalid tool-call serialization fails loudly instead of silently falling back to the wrong template shape.

• Added tool/testing/run_local_e2e.dart as a discovery and orchestration entry point for heavyweight local-only Dart E2E, Flutter device, and Web/Playwright smoke scenarios.
• Hardened the upstream llama.cpp chat/template E2E runner against current llama.cpp target renames, dynamic backend library lookup, and full test-chat server/mtmd build requirements.
• Documented that real-model/device/WebGPU scenarios remain skipped from default CI and should be opted into explicitly with --list and --dry-run first.

• Updated the default WebGPU bridge asset pin to leehack/llama-web-bridge-assets@v0.1.16 (llama.cpp b9165), picking up the published JS bridge build, TypeScript declaration asset, and refreshed bridge docs.

• Added WebGPU readiness guidance covering browser capability checks, cross-origin isolation, bridge asset/version diagnostics, fallback behavior, model/configuration pressure, and the Flutter Web real-model smoke path.

• Added ModelDownloadController, a dependency-free helper that turns ModelDownloadManager cache/download work into app-facing lifecycle states for resolving, cache checks, downloads, verification, ready, failed, cancelled, and retry flows.
• Wired the runnable chat app example through a ModelDownloadManager adapter so its model-management UI demonstrates the controller while preserving the example's multi-asset and web-cache service behavior.

• Added ModelSource for local paths, HTTP(S) URLs, and Hugging Face hf://owner/repo/path/to/model.gguf references, including deterministic cache keys and redacted metadata/log identities for signed URLs.
• Added ModelLoadOptions, ModelCachePolicy, resolver targets, and download/cache metadata/progress value models for package-managed model download and cache management.
• Added native/file-backed DefaultModelDownloadManager support for streaming HTTP downloads, .part files, atomic promotion, persisted metadata, authenticated bearer/custom headers, cancellation, retry, Range resume, cache hit/refresh/cache-only/no-cache policies, SHA-256 verification, cache listing, removal, clearing, and age/size pruning.
• Improved Hugging Face source ergonomics: hf:// references now accept ?revision=... for branch/ref names containing slashes, and docs clarify current single-file behavior, private/gated bearer-token usage, separate mmproj asset handling, sharded-GGUF limitations, and redaction guarantees.
• Serialized concurrent stable-cache downloads for the same remote cache entry across manager instances so duplicate callers do not race on shared .part files or metadata, while distinct cache entries can still download in parallel and waiting-caller cancellation does not cancel the active download.
• Hardened versioned cache metadata recovery: completed files can rebuild missing, malformed, or unsupported-schema sidecars without network access, while byte-count and stored/caller SHA-256 mismatches are treated as cache misses and safely re-downloaded.
• Clarified ModelSource.path(...) option semantics: local paths now reject remote/download-only options (non-default cache policies, cache directories, authenticated headers, resume, and retry overrides) while continuing to support cancellation and optional local SHA-256 verification.
• Added LlamaEngine.loadModelSource(...) to route local sources through the existing native local loader, remote sources through the native download cache before local loading, and simple remote sources through URL-capable web backends when available.
• Migrated server/testing helpers away from ad-hoc model downloads so examples dogfood the package-managed cache manager.

• Added LlamaEngine.supportsStatePersistence, LlamaEngine.stateSaveFile(...), and LlamaEngine.stateLoadFile(...) so callers can persist and restore llama.cpp KV-cache state for fast raw-prompt resume/fork workflows.
• Added BackendStatePersistence, BackendStatePersistenceSupport, and StateLoadResult for custom backend implementers and diagnostics.
• Documented that state files are opaque llama.cpp artifacts tied to the same model and runtime/build, that native paths use the app filesystem while web paths use the bridge WASMFS virtual filesystem, and that ChatSession message history must be persisted separately.
• Added WebGPU bridge state persistence wiring for bridge assets v0.1.15+, including Dart JS interop, backend forwarding, and browser integration test coverage.

• Updated native hook pinning to leehack/llamadart-native@b9016, picking up the CUDA 12.8 Blackwell-capable native bundles.
• Updated default web bridge asset pinning to leehack/llama-web-bridge-assets@v0.1.14 (llama.cpp b9016) so native and web runtimes track the same upstream revision.
• Picked up the bridge-side Qwen UTF-8 streaming stabilization and multimodal fallback narrowing, while preserving control-token output for parser consumers.
• Picked up the bridge-side BERT embedding thread-pool sizing fix so automatic thread selection does not exceed the compiled WebAssembly pthread pool.

• Added ModelParams.useMmap (default true) and ModelParams.useMlock (default false), wired to llama_model_params.use_mmap / use_mlock. Lets callers turn off mmap for platforms where memory-mapped weights hurt throughput, or pin weights in RAM to avoid first-token paging spikes.
• Added ModelParams.flashAttention with the FlashAttention.{auto, enabled, disabled} enum, wired to llama_context_params.flash_attn_type. Explicit settings win over the existing automatic Android/Vulkan heuristics; auto preserves prior behavior.
• Added ModelParams.cacheTypeK and ModelParams.cacheTypeV with the KvCacheType.{f16, q8_0, q4_0} enum, wired to llama_context_params.type_k / type_v. Enables KV-cache quantization (Q8_0 ≈ halves KV memory; Q4_0 ≈ quarters it). When the user requests a non-F16 KV type with flashAttention: auto, the service auto-promotes flash attention to enabled — llama.cpp requires it for KV quantization.
• Added ModelParams.kvUnified (nullable) for explicit override of llama_context_params.kv_unified. null keeps the existing auto-enable-when-multi-sequence behavior.
• Added ModelParams.ropeFrequencyBase and ModelParams.ropeFrequencyScale (both nullable) for context-extension overrides on llama_context_params.rope_freq_base / rope_freq_scale. null keeps the model's trained values.
• Forwarded native-compatible ModelParams load tuning knobs through the WebGPU bridge path, including maxParallelSequences, flash attention, KV-cache type, KV-unified, RoPE, split-mode, and main-GPU options.
• Matched native batch defaults on the WebGPU path so unset batchSize / microBatchSize cascade to n_batch = n_ctx and n_ubatch = n_batch, avoiding first-embedding aborts for BERT-class/non-causal encoder models while preserving explicit caller values and Qwen3.5 web tuning.

• Added ModelParams.mainGpu and wired it to llama.cpp llama_model_params.main_gpu.
• Added ModelParams.splitMode and wired it to llama.cpp llama_model_params.split_mode, enabling explicit single-GPU selection with ModelSplitMode.none.

• Resolved ggml backend registry/device APIs from the loaded ggml runtime DLL when the generated default FFI asset cannot see those symbols, restoring explicit Vulkan device selection in Windows split bundles.

• Filtered backend-owned runtime dependencies during native asset bundling so CUDA runtime DLLs and OpenBLAS runtime libraries are emitted only when their owning backend module is selected.
• Kept unknown non-core runtime libraries bundled for compatibility with future native bundle layouts.

• Updated native hook pinning and regenerated bindings through leehack/llamadart-native@b8955.

• Parsed streamed <|channel>thought ... <channel|> blocks into thinking deltas instead of leaking Gemma 4 thought markers into content output.
• Added engine coverage for Gemma 4 thought-channel chunks split across native stream boundaries.

• Tracked the chat app lockfile so generated Flutter plugin metadata stays stable in CI and release validation.

• Updated native hook pinning and regenerated bindings through leehack/llamadart-native@b8638.

• Converted native multimodal prompt-evaluation overflow paths into Dart exceptions instead of allowing downstream sampling asserts.
• Downscaled staged chat-app image picks to a 384px max edge across Android, iOS, macOS, and Web to reduce multimodal context pressure.
• Added a local-only macOS Qwen3.5 multimodal repro harness plus CI-safe provider coverage for the new overflow guidance.

• Added built-in Gemma 4 template detection, rendering, and parsing support, including thinking and tool-call handling.
• Added runtime projector capability checks so multimodal flows and the chat app gate image/audio input against supportsVision / supportsAudio instead of model-family assumptions.
• Documented current Gemma 4 projector behavior in the docs site and chat app guidance.

• Documented that iOS builds require a minimum deployment target of 16.4 or newer across the README, docs site, and example docs.
• Updated example/chat_app iOS Podfile and Runner project settings to use deployment target 16.4.

• Honored ggml_backend_score during asset-based backend fallback so unsupported Android CPU variant libraries are skipped before initialization.
• Changed Android auto backend resolution to prefer CPU by default while keeping Vulkan available for explicit opt-in.
• Clarified that changing hooks.user_defines requires flutter clean && flutter pub get before rebuilding.

• Updated native hook pinning and regenerated bindings to leehack/llamadart-native@b8480.
• Refreshed generated low-level FFI bindings to match the synced upstream headers.

• Updated native hook pinning and regenerated bindings to leehack/llamadart-native@b8373.
• Hardened Linux bundle loading for packaged apps and accepted versioned libllamadart mappings so colocated native dependencies resolve more reliably at runtime.

• Fixed Hermes handler parsing when whitespace appears between <tool_call> and the JSON payload.

• Updated native hook pinning to leehack/llamadart-native@b8216.
• Updated default web bridge asset pinning to leehack/llama-web-bridge-assets@v0.1.10 (llama.cpp b8216).

• Switched bundled Qwen3.5 presets to Unsloth Q4_K_M GGUFs across the example catalog and tooling.
• Added Android-native perf diagnostics chips (p_eval, eval, sample, reuse) backed by llama.cpp context timings with manual timing fallback when built-in counters report zero.
• Restored a targeted Android Vulkan fast path for local Qwen3.5 0.8B / 2B / 4B models by re-enabling KQV/op-offload/flash-attention where stable.
• Updated Android chat app defaults to prefer CPU for Qwen3.5 0.8B and 2B, and reduced Android 0.8B context to 2048 for lower first-token latency.
• Hardened Android multimodal handling by downscaling staged images in the chat app and forcing Qwen3.5 0.8B projector work onto CPU on Android.
• Fixed WebGPU Qwen prompt/control-token handling and committed companion bridge-side streaming/multimodal fixes required by the local chat app runtime.

• Added LlamaEngine.embed(...) and LlamaEngine.embedBatch(...) for direct vector generation.
• Added optional backend capability interface BackendEmbeddings for custom backend implementers.
• Added optional backend batch capability BackendBatchEmbeddings and worker-side batch embedding request/response path to reduce isolate round-trip overhead in embedBatch(...).
• Added ModelParams.maxParallelSequences (n_seq_max) so contexts can reserve multiple sequence slots for true multi-sequence embedding batches.
• Wired native isolate/worker/service embedding flow to llama.cpp embedding outputs with optional L2 normalization.
• Added embedding-focused tests for engine behavior and worker message contracts.

• Added example/basic_app/bin/llamadart_embedding_example.dart.
• Added example/basic_app/bin/llamadart_sqlite_vector_example.dart for local embedding retrieval with SQLite vector search.
• Updated example docs and top-level README with embedding usage snippets.
• Added tool/testing/native_embedding_benchmark.dart to compare sequential embedding calls vs embedBatch(...) throughput (with optional --json-out).
• Added tool/testing/native_embedding_sweep.dart to run max-seq sweeps and dump CSV speedup reports for plotting.

• Added WebGPU bridge embedding APIs and wired web backend support for LlamaEngine.embed(...) / embedBatch(...).
• Updated default web bridge asset pinning to leehack/llama-web-bridge-assets@v0.1.8.
• Validated the v0.1.8 bridge bundle through local fetch-script checksum verification.

• Reduced bridge log noise and improved runtime profile diagnostics for web sessions.
• Stabilized multimodal backend switching using resolved runtime mode behavior and added an E2E regression gate.
• Tuned streaming/typewriter pacing and token callback overhead to improve incremental render smoothness.
• Added GPU-path multimodal image-size capping to reduce runtime pressure on large image inputs.

• Updated example/chat_app recommended Qwen presets to the Qwen3.5 lineup (0.8B, 2B, 4B, 9B) and removed older Qwen2.5/Qwen3 defaults from the in-app library.
• Added multimodal projector (mmproj) wiring for Qwen3.5 model cards and tuned safer multimodal defaults (contextSize: 8192, maxTokens: 1024).
• Fixed Flutter text paint crashes caused by malformed UTF-16 streaming boundaries by aligning incremental reveal to surrogate-pair boundaries and sanitizing text/tool payload rendering paths.
• Added sanitizer unit coverage and refreshed chat-app README architecture/troubleshooting sections for multimodal and UTF-16 guidance.

Multimodal projector offload alignment:

• Updated native multimodal projector initialization to follow effective model-load configuration.
• CPU-only model settings (preferredBackend: cpu or gpuLayers: 0) now also disable mmproj GPU offload.

Package metadata cleanup:

• Removed unused Flutter-only constraints/dependencies from the root pubspec.yaml (environment.flutter, flutter, path_provider, json_rpc_2, integration_test) to keep the core package pure Dart.
• Kept Flutter-specific dependencies scoped to Flutter example apps.

Backend selection safety and status accuracy:

• Added strict CPU-mode behavior in native backend preparation so preferredBackend: cpu no longer initializes optional GPU backends during startup/model load probing.
• Disabled context-time GPU offload knobs (offload_kqv, op_offload, flash-attention auto path) when effective GPU layers resolve to zero, preventing GPU allocation attempts during context creation in CPU mode.
• Added ModelParams.batchSize (n_batch) and ModelParams.microBatchSize (n_ubatch) so context batch sizing can be tuned independently from contextSize while preserving legacy defaults.
• Split backend reporting into two semantics: selectable backend options (getAvailableBackends) vs active runtime backend (getBackendName).
• Added optional BackendAvailability capability and LlamaEngine.getAvailableBackends() to support safe settings UIs without forcing GPU initialization.
• Added optional BackendRuntimeDiagnostics capability and LlamaEngine.getResolvedGpuLayers() to expose resolved native load-time layer count for runtime diagnostics.
• Updated example/chat_app to populate backend selector options from safe availability discovery while keeping active-backend status bound to effective runtime backend.
• Improved native auto/explicit backend status resolution to avoid false CPU labeling on Apple consolidated runtimes and false GPU labeling when explicit backend falls back.

Web model cache + large-model UX improvements (chat app):

• Updated web Download flow to prefetch model/mmproj bytes into browser Cache Storage with live progress and cancellation support.
• Added best-effort cache eviction for web model delete actions.
• Added large-model web load fallback to fetch-backed worker runtime path (bridge) to reduce contiguous ArrayBuffer pressure.
• Added dedicated web bridge worker entry wiring and worker fallback diagnostics to improve worker startup reliability.
• Reduced synthetic load-progress dominance so bridge/network progress appears earlier during web model load.
• Added warning-only UI guidance for very large web models that may exceed browser memory limits at load time.

Web model-load resilience:

• Updated WebGpuLlamaBackend to retry web model loads with reduced context sizes (and CPU fallback as last attempt) when bridge errors indicate browser memory pressure.
• Added bridge config plumbing for optional wasm64 core assets (llama_webgpu_core_mem64) with automatic fallback to wasm32 when unsupported.
• Added explicit runtime diagnostics and error normalization for worker-thread and cross-origin-isolation requirements in large web model load flows.
• Updated default bridge asset pinning in chat app/docs/fetch script to leehack/llama-web-bridge-assets@v0.1.5.
• Updated HF static chat-app deploy workflow to emit COI custom_headers in generated Space README frontmatter.

Android arm64 CPU variant policy and loader hardening:

• Updated native hook tag pin from b8138 to b8157 to consume Android arm64 CPU-variant runtime bundles.
• Added Android arm64 CPU policy keys in hook config: cpu_profile (full default, compact) and advanced cpu_variants override.
• Added hook tests and Android hook integration coverage to verify pubspec-driven CPU variant packaging behavior.
• Hardened Android runtime backend loading to resolve CPU variant modules even when backend module directory discovery is unavailable.
• Added Android runtime smoke helper (scripts/android_runtime_smoke.sh) and smoke-plan docs for device verification.
• Compatibility note: no public API breaking changes. android-arm64 now defaults to cpu_profile: full, which may increase package size compared with baseline-only CPU packaging.

• Synced bundled native runtime/assets and regenerated bindings from b8099 to b8138.
• Pulled in Android arm64 ISA compatibility hardening (including STLUR guard changes) to prevent launch-time crashes on older devices.

• Reduced settings-write overhead during frequent parameter adjustments.
• Improved model manager responsiveness during download progress updates.
• Smoothed chat streaming auto-follow and rendering to reduce unnecessary UI work.

• Updated web "Download" behavior to verify remote model/mmproj availability without pre-buffering large GGUF payloads in app memory.
• Clarified that web cache population occurs when a model is first loaded.

• Added safe fallback handling for invalid persisted log-level settings.
• Added regression tests for persisted settings fallback behavior.

• Added example/tui_coding_agent, a nocterm-based terminal coding agent with tool-calling loop, workspace-scoped file/command tools, and runtime model switching.
• Default model source is GLM 4.7 Flash (unsloth/GLM-4.7-Flash-GGUF:UD-Q4_K_XL) with support for custom local paths/URLs/Hugging Face shorthand.
• Added stable text-protocol tool mode as the default (native template grammar tool-calling remains available via --native-tool-calling for experimentation).

• Reduced request overhead by caching model metadata and skipping unnecessary prompt token counting in create(...).
• Improved native stream throughput with worker-side token chunk batching and configurable thresholds (streamBatchTokenThreshold, streamBatchByteThreshold).
• Added prompt-prefix reuse for native text generation (reusePromptPrefix, enabled by default) with conservative full-replay fallback to preserve deterministic parity.
• Optimized ChatSession context trimming using bounded turn-offset search to avoid repeated linear recount loops on long histories.

• Added tool/testing/native_inference_benchmark.dart for TTFT, throughput, and latency measurement with tunable generation settings.
• Added tool/testing/native_prompt_reuse_parity.dart and curated prompt sets for deterministic prompt-reuse parity validation.
• Added CI prompt-reuse parity checks to catch native reuse regressions.

Publishing compatibility fix:

• Moved hook backend-config support code out of hook/src/ into lib/src/hook/ because pub.dev currently only allows hook/build.dart under hook files.
• Updated hook/test imports accordingly to keep native-assets backend selection behavior unchanged.

llama.cpp parity expansion (Dart-native template/parser pipeline):

• Reworked template detection/render/parse routing to align with llama.cpp semantics across supported chat formats, including format-specific tool-call parsing and fallback behavior.
• Added PEG parity components in Dart (peg_parser_builder, peg_chat_parser) and integrated parser-carrying render/parse flow for PEG-native/constructed formats.
• Removed brittle fallback coercions that could mutate valid tool names/argument keys, preserving model-emitted tool payloads for dispatch parity.
• Hardened template capability detection with Jinja AST + execution probing, while preventing typed-content false positives caused by raw content stringification.
• [BREAKING] Removed legacy custom template-handler APIs: ChatTemplateMatcher, ChatTemplateRoutingContext, ChatTemplateEngine.registerHandler(...), ChatTemplateEngine.unregisterHandler(...), ChatTemplateEngine.clearCustomHandlers(...), ChatTemplateEngine.registerTemplateOverride(...), ChatTemplateEngine.unregisterTemplateOverride(...), ChatTemplateEngine.clearTemplateOverrides(...), and per-call customHandlerId / parse handlerId routing.
• Removed silent render/parse fallback paths so handler/parser failures are surfaced instead of downgraded to content-only output.
• Added llama.cpp-equivalent per-call template globals/time injection via chatTemplateKwargs and templateNow.

Parity test coverage and tooling:

• Added vendored llama.cpp template parity integration coverage for detection + render + parse paths.
• Added upstream llama.cpp chat/template suite runners and local E2E harness (run_llama_cpp_chat_tests.sh, run_template_parity_suites.sh).
• Added mirrored unit tests for new internal template components (peg_parser_builder, template_internal_metadata) to satisfy structure guards.

Test cleanup and maintainability:

• Reduced noisy diagnostics in template integration tests and centralized format sample parse payload fixtures for easier parity maintenance.

Native integration cleanup (llamadart-native migration):

• Added tool/testing/prepare_llama_cpp_source.sh to fetch/refresh ggml-org/llama.cpp into .dart_tool/llama_cpp (or LLAMA_CPP_SOURCE_DIR) pinned to a resolved ref (LLAMA_CPP_REF, default latest release tag).
• Updated tool/testing/run_llama_cpp_chat_tests.sh to use prepared .dart_tool source instead of third_party/llama_cpp, so local upstream chat-suite runs no longer depend on vendored source.
• Updated template parity tests to resolve fixtures from LLAMA_CPP_TEMPLATES_DIR or .dart_tool/llama_cpp/models/templates instead of third_party/llama_cpp.
• Clarified README backend matrix notes: KleidiAI/ZenDNN are CPU-path optimizations, not selectable runtime backend modules.
• Runtime backend probing for split-module bundles now runs during backend initialization (not only after first model load), so device/backend availability is visible earlier in app flows.
• Native-assets hook output now refreshes emitted native files per build to prevent stale backend module carryover when backend config changes.

Linux runtime/link validation and backend loader hardening:

• Hardened split-module backend loading to avoid probing backends that are not bundled for the active platform/arch, reducing noisy optional-backend load failures.
• Added failed-backend memoization so missing optional modules are not retried on every model load.
• Tightened Linux cache source selection to the current ABI bundle (linux-arm64 vs linux-x64) when preparing runtime dependencies.
• Added Linux backend/runtime setup guidance in README, including distro-specific package baselines (Ubuntu/Debian, Fedora/RHEL/CentOS, Arch).
• Added reproducible Docker link-check flows for baseline (cpu/vulkan/blas) and optional cuda/hip module dependency resolution.
• Added scripts/check_native_link_deps.sh helper plus dedicated validation images: docker/validation/Dockerfile.cuda-linkcheck and docker/validation/Dockerfile.hip-linkcheck.

Chat example backend UX cleanup:

• Removed user-facing Auto backend option from settings; only concrete runtime-detected backends are shown.
• Added migration behavior that resolves legacy saved Auto preference to the best detected backend at runtime.

llama.cpp parity hardening:

• ChatTemplateEngine now preserves handler-provided tokens even when grammar is attached via params, avoiding token-loss regressions in tool/thinking formats.
• Native stop-sequence handling now skips preserved tokens so parser-critical markers are not terminated early.
• Generic tool-instruction system injection now follows llama.cpp semantics more closely (replace first system content when supported, otherwise prepend to first message content).
• LFM2 output parsing now extracts reasoning more consistently across tool and non-tool output shapes.

Chat example loop/lifecycle hardening:

• Improved tool-loop guards (first-turn force-only behavior, duplicate/equivalent call suppression, per-tool budget, and loop-stop messaging).
• Added response fallback that can ground final answers from recent tool results when the model emits stale real-time disclaimers.
• Added assistant debug badges (fmt:*, think:*, content:json, fallback:tool-result) and strengthened detach/exit disposal paths.

Parity/integration test robustness:

• tool_calling_integration_test now accepts both structured tool_calls deltas and XML-style <tool_call> payloads.
• llama.cpp template-detection integration expectations were updated for current Ministral-family routing outcomes.

Documentation updates:

• Clarified chat app behavior when models return JSON-shaped assistant content (for example {"response":"..."}) and documented content:json diagnostics.
• Documented example server sampling defaults (penalty=1.0, top_p=0.95, min_p=0.05) and added a CLI README batch parity-matrix usage example.

Chat app backend/status fixes:

• Backend switching now preserves configured gpuLayers while still allowing load-time CPU enforcement.
• Runtime backend labeling and GPU activity diagnostics now follow effective user selection, preventing false "VULKAN active" status when CPU mode is selected.

Context size auto mode:

• Restored support for Context Size: Auto by preserving 0 in persisted settings and passing auto behavior through to session context-limit resolution.

Tool-call parsing fixes (Hermes):

• Introduced staged double-brace recovery: parse as-is first, unwrap one outer {{...}} layer second, and only fall back to full _normalizeDoubleBraces when all braces are consistently doubled.
• Added a consistency gate to _normalizeDoubleBraces that bails out on mixed single/double brace payloads to prevent corruption of valid nested JSON.

Tool-call parsing fixes (Magistral):

• Broadened whitespace skipping in _extractJsonObject to handle \n, \r, and \t between [ARGS] and the JSON body.

Example app (basic_app):

• Replaced toList() buffering with await for streaming for real-time token yield.
• Added tools parameter to every follow-up create() call and bounded tool-execution loop with _maxToolRounds = 10.

Test coverage:

• Added chat app regression tests for backend switching behavior and context-size auto persistence.
• Added regression tests for Hermes wrapped+nested double-brace payloads and Magistral [ARGS] with newline/nested arguments.

Example rename (server):

• Renamed example/api_server to example/llamadart_server.
• Renamed the example package/bin entrypoint to llamadart_server.
• Updated llama.cpp tool-call parity defaults/docs to target example/llamadart_server.

GLM 4.5 template parity:

• Added XML tool-call grammar generation for <tool_call> payloads with <arg_key>/<arg_value> pairs.
• Added GLM-specific preserved tokens and <|user|> stop handling for tool-call flows.
• Updated parser extraction to handle GLM XML tool calls from assistant content and reasoning blocks.

Template/native runtime fixes:

• Typed-content template rendering now activates only when messages actually include media parts.
• Native context reset now clears llama memory in-place instead of reinitializing the context.

• Added minP to GenerationParams with a default value of 0.0 and copyWith support.

• Added optional llama.cpp min_p sampler initialization in LlamaCppService when minP > 0.

• Added unit coverage for GenerationParams.minP default and copyWith behavior.

• Expanded llama.cpp parity across additional format handlers, including grammar construction, lazy-grammar triggers, preserved tokens, and parser behavior for tool-call payload extraction.
• Added shared ToolCallGrammarUtils helpers for wrapped object/array tool-call grammar generation and root-rule wrapping.

• Fixed malformed GBNF escaping in Hermes/Command-R string rules that could cause runtime llama_grammar_init_impl parse failures during tool-calling generations.

• Added and expanded handler-level parity tests (Apertus, LFM2, Nemotron V2, Magistral, Seed-OSS, Xiaomi MiMo, DeepSeek R1/V3, Hermes) and mirrored unit tests for new grammar utilities.

• Updated README internal links to absolute GitHub URLs so they resolve reliably on pub.dev.
• Updated release/migration wording after 0.5.0 publication and refreshed installation/version snippets.
• Corrected iOS simulator architecture notes and contributor prerequisites/build target docs.

• Expanded .pubignore to exclude local build outputs, large model/test artifacts, and checked-out third_party sources from package uploads.

[BREAKING] Public API Changes:

• Root exports were tightened; previously exposed internals such as ToolRegistry, LlamaTokenizer, and ChatTemplateProcessor are no longer part of the public package API.
• ChatSession now centers on create(...) streaming LlamaCompletionChunk; legacy chat(...) / chatText(...) style usage must migrate.
• LlamaChatMessage constructor names were standardized (.fromText, .withContent) in place of older named constructors.
• Default maxTokens in GenerationParams increased from 512 to 4096.
• LlamaChatMessage.toJson() no longer includes name on tool role messages.
• ModelParams.logLevel was removed; logging control now lives on LlamaEngine via setDartLogLevel(...) and setNativeLogLevel(...).
• LlamaBackend interface changed for custom backend implementers (notably getVramInfo and updated applyChatTemplate).
• Model reload behavior is stricter: loadModel(...) now requires unloading first.
• Migration details are documented in MIGRATION.md.

Template/Parser Parity Expansion:

• Added llama.cpp-aligned format detection and handlers for additional templates including FireFunction v2, Functionary v3.2, Functionary v3.1 (Llama 3.1), GPT-OSS, Seed-OSS, Nemotron V2, Apertus, Solar Open, EXAONE MoE, Xiaomi MiMo, and TranslateGemma.
• Improved parser parity for format-specific tool-calling and reasoning extraction, including <|python_tag|> parsing for Llama 3 flows.
• Narrowed generic grammar auto-application to generic/content-only routing to avoid interfering with format-specific tool schemas.

Template Extensibility APIs:

• Added global custom handler registration and template override APIs in ChatTemplateEngine.
• Added per-call customTemplate and customHandlerId routing support and threaded handler identity into parse paths.
• Added cookbook examples and regression tests for registration precedence and fallback behavior.

Logging Controls:

• Added split logging controls in LlamaEngine: setDartLogLevel and setNativeLogLevel, while keeping setLogLevel as a convenience method.
• Fixed native none log level suppression so llama.cpp/ggml logs are fully muted when requested.

Chat App Improvements:

• Added model capability badges and per-model generation presets.
• Added template-aware tool enablement guardrails and separate Dart/native log level settings in the UI.

Test Suite Overhaul:

• Expanded template parity coverage (detection, handlers, grammar, workarounds, registry precedence, and integration scenarios).
• Added additional unit tests for exceptions, logging, and core model definitions.

• Refactored LlamaBackend for strict Web isolation using "Native-First" conditional exports, ensuring native performance and full web safety.
• Standardized backend instantiation via a unified LlamaBackend() factory across all examples and scripts.

• Resolved "Max Tokens is 0" on Web by implementing getLoadedContextInfo() and robust GGUF metadata fallback in LlamaEngine.
• Improved numeric metadata extraction on Web for better compatibility with varied GGUF exporters.

• Resolved "Unexpected empty grammar stack" crash by reordering the sampler chain (filtering tokens via GBNF before performing probability-based sampling).

• Pivoted from mock-based unit tests to real-world integration tests using the actual llama.cpp native backend.
• Ensured full verification of model loading, tokenization, text generation, and grammar constraints against physical models.
• Multi-Platform Configuration: Introduced dart_test.yaml and @TestOn tags to enable seamless execution of all tests across VM and Chrome with a single dart test command.

• Implemented FD-level redirection (dup2 to /dev/null) for LlamaLogLevel.none on native platforms.
• This provides a crash-free alternative to FFI-based log callbacks, which were unstable during low-level native initialization (e.g., Metal).

• Achieved 100% clean dart analyze across the core library and all example applications.
• Replaced legacy stubs in the chat application with a clean, interface-based ModelService architecture.

• Implemented robust resumable downloads for large models using HTTP Range requests.
• Added persistent .meta files to track download progress across app restarts.

• Refined the ModelCard with a visual Pause/Resume toggle.
• Added a Trash icon in the card header for full cancellation and data discard of active or partial downloads.
• Improved progress feedback with clear "Paused" and "Downloading" states.

• Added loadMultimodalProjector to LlamaEngine.
• Introduced LlamaChatMessage.withContent and LlamaContentPart (Text, Image, Audio).
• Fix: Resolved missing multimodal symbols in native builds by properly linking the mtmd module.

• Implemented a specialized Question: / Answer: chat template fallback for Moondream models.
• Added dynamic BOS token handling: Automatically disables BOS injection for models where BOS == EOS (like Moondream) to prevent immediate "End of Generation".

• Moved high-level chat() and chatWithTools() logic from LlamaEngine to ChatSession.
• LlamaEngine is now a dedicated low-level orchestrator for model loading, tokenization, and raw inference.

• Optional Tool Calls: Tools are no longer forced by default. The model now decides when to use a tool vs. responding directly based on context.
• Final Response Generation: After a tool returns a result, the model now generates a natural language response (without grammar constraints) to interpret the result for the user.
• forceToolCall: Added a session-level flag to re-enable strict grammar-constrained tool calls for smaller models (e.g., 0.5B - 1B).

• Fixed a crash in the Flutter chat app during close/restart by implementing and using an idempotent dispose() in ChatService.
• Added Qwen 2.5 3B and 7B models to the download list with clear RAM/VRAM requirements for testing complex instruction following and tool use.

• Improved export management for MSVC to ensure symbol visibility.
• Added Sccache support for Windows builds to significantly improve CI performance.

• Implemented GitHub Actions to automate llama.cpp updates, regression testing, and release artifact generation.

• LlamaChatMessage.role now returns a LlamaChatRole enum instead of a String. All manual role string comparisons should be updated to use the enum.

• Default LlamaChatMessage constructor (string-based) is now deprecated; use .fromText() or .withContent() instead.
• LlamaChatMessage.roleString is deprecated and will be removed in v1.0.

• Native Platforms: Simplified to an on/off toggle to ensure stability. LlamaLogLevel.none suppresses all output; other levels enable default stderr logging.
• Web: Supports full granular filtering (Debug, Info, Warn, Error).

• Text generation with llama.cpp backend.
• GGUF model support.
• Hardware acceleration (Metal, Vulkan).
• Flutter Chat Example.
• CLI Basic Example.