llamadart
v0.6.16
pub.dev
Documentation for v0.6.16, an older release. Read v0.9.0, the latest release

Help & Reference

Recent Releases

Review recent llamadart release highlights and jump to the canonical changelog for full release notes.

On this page

For canonical full release notes, use:

0.6.16#

  • Fixed native getVramInfo() so llama.cpp GPU-class backend devices can report free/total VRAM when available, with Windows split-bundle registry fallback handling for backend-device symbols.
  • 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.
  • Compatibility note: no public API breaking changes in 0.6.16; existing 0.6.15 callers remain compatible.

0.6.15#

  • 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.
  • Compatibility note: no public API breaking changes in 0.6.15; existing 0.6.14 callers remain compatible. The chat-template changes fix multimodal serialization behavior for affected templates, and the local E2E runner is additive.

0.6.14#

  • 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.
  • Compatibility note: no public API breaking changes in 0.6.14; the WebGPU bridge asset update and ModelDownloadController are additive, and existing 0.6.13 callers remain compatible.

0.6.13#

  • Added package-managed model source downloads and cache management: ModelSource, ModelLoadOptions, ModelCachePolicy, resolver targets, download/cache metadata, progress callbacks, cache inspection, removal, clearing, and age/size pruning.
  • Added native/file-backed DefaultModelDownloadManager support for streaming HTTP downloads, .part files with atomic promotion, authenticated bearer and custom headers, cooperative cancellation, retry, HTTP Range resume, cache hit/refresh/cache-only/no-cache policies, SHA-256 verification, and persisted redacted metadata for signed URLs.
  • Improved Hugging Face hf:// ergonomics with ?revision=... parsing for branch/ref names containing slashes, plus docs for private/gated bearer-token usage, separate mmproj assets, sharded-GGUF limitations, and redaction guarantees.
  • Hardened download/cache correctness by serializing concurrent same-entry downloads, recovering missing or malformed cache metadata sidecars, treating mismatched byte-count/SHA-256 metadata as cache misses, and rejecting remote-only options for local ModelSource.path(...) inputs.
  • Added LlamaEngine.loadModelSource(...) so local path sources keep using the existing native loader, remote HTTP(S)/Hugging Face sources download through the package-managed native cache before local loading, and URL-capable web backends keep using direct URL loading for simple unauthenticated requests.
  • Added KV-cache state persistence APIs: LlamaEngine.supportsStatePersistence, stateSaveFile(...), stateLoadFile(...), backend support diagnostics, and WebGPU bridge forwarding for bridge assets v0.1.15+.
  • Compatibility note: no public API breaking changes in 0.6.13; existing loadModel(...) callers are unchanged.

0.6.12#

  • Synced default WebGPU bridge asset pinning to leehack/llama-web-bridge-assets@v0.1.14 (llama.cpp b9016) to match the native runtime pin.
  • Picked up 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.
  • Forwarded native-compatible ModelParams load tuning knobs through the WebGPU bridge path, including sequence slots, flash attention, KV cache type, RoPE overrides, split mode, and main GPU.
  • Matched native batch defaults on WebGPU so unset batchSize and microBatchSize use n_batch = n_ctx and n_ubatch = n_batch, avoiding first-embedding aborts for BERT-class/non-causal encoder models while preserving model-specific Qwen3.5-0.8B WebGPU safety tuning.
  • 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, while unknown runtime libraries stay bundled for forward compatibility.
  • Compatibility note: no public API breaking changes in 0.6.12.

0.6.11#

  • Synced native hook pinning and regenerated bindings through leehack/llamadart-native@b8955.
  • Fixed Gemma 4 streaming so <|channel>thought ... <channel|> output is emitted as thinking deltas instead of content text, including when channel markers are split across streamed chunks.
  • Tracked the chat app lockfile for stable generated Flutter plugin metadata in CI and release validation.
  • Compatibility note: no public API breaking changes in 0.6.11.

0.6.10#

  • Synced native hook pinning and regenerated bindings through leehack/llamadart-native@b8638.
  • Hardened multimodal prompt overflow handling so native failures surface as Dart exceptions, and reduced staged chat-app image size to a 384px max edge to lower multimodal context pressure.
  • Added built-in Gemma 4 template detection/render/parse support, including thinking and tool-call handling.
  • Added runtime projector capability gating so multimodal flows and the chat app respect actual supportsVision / supportsAudio results instead of model-family assumptions.
  • Compatibility note: no public API breaking changes in 0.6.10.

0.6.9#

  • 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 Android asset-based backend fallback so unsupported 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.
  • Compatibility note: no public API breaking changes in 0.6.9.

0.6.8#

  • Synced native hook pinning and regenerated bindings to leehack/llamadart-native@b8480.
  • Refreshed generated low-level FFI bindings to match the synced upstream headers.
  • Compatibility note: no public API breaking changes in 0.6.8.

0.6.7#

  • Synced native hook pinning and regenerated bindings to leehack/llamadart-native@b8373.
  • Hardened Linux bundle loading for packaged apps and improved versioned libllamadart dependency resolution.
  • Fixed Hermes tool-call parsing when whitespace appears between <tool_call> and the JSON payload.
  • Compatibility note: no public API breaking changes in 0.6.7.

0.6.6#

  • Synced native hook pin 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 example presets to Unsloth Q4_K_M GGUFs.
  • Added native perf diagnostics chips in the chat app (p_eval, eval, sample, reuse) and Android-specific Qwen tuning guidance.
  • Restored a targeted Android Vulkan fast path for local Qwen3.5 0.8B / 2B / 4B models while keeping CPU as the recommended Android preset for 0.8B / 2B.
  • Fixed local web chat app bridge/runtime handling for Qwen prompt streaming and multimodal fallback behavior.
  • Compatibility note: no public API breaking changes in 0.6.6.

0.6.5#

  • Added embedding APIs: LlamaEngine.embed(...) and LlamaEngine.embedBatch(...).
  • Added backend embedding capability interfaces for custom backend implementations.
  • Added multi-sequence embedding batching support via ModelParams.maxParallelSequences (n_seq_max).
  • Added native embedding benchmark tooling: tool/testing/native_embedding_benchmark.dart and tool/testing/native_embedding_sweep.dart.
  • Added website docs for embeddings and updated basic-app docs with embedding examples.
  • Added a Basic App SQLite vector retrieval example using bin/llamadart_sqlite_vector_example.dart.
  • Updated default WebGPU bridge asset pinning to leehack/llama-web-bridge-assets@v0.1.8.
  • Improved WebGPU runtime stability/tuning in chat app flows (backend switching, streaming smoothness, and multimodal regression gating).
  • Added GPU-path multimodal image-size capping to reduce memory/runtime pressure on larger image inputs.
  • Compatibility note: no public API breaking changes in 0.6.5.

0.6.4#

  • Aligned multimodal projector offload with effective model-load settings, including CPU-only configurations.
  • Added safer backend selection/discovery APIs and improved runtime backend status plus GPU-layer diagnostics accuracy.
  • Improved web large-model handling with cache-prefetch download UX, bridge worker fallback paths, memory-pressure retries, and wasm64-core fallback wiring.
  • Synced native hook tag to b8157 and added Android arm64 CPU-profile and variant policy support with loader hardening.

0.6.3#

  • Synced native runtime to llama.cpp b8138 and picked up Android arm64 crash/compatibility hardening.
  • Example app performance/UX polish and web model handling improvements.
  • Added example/tui_coding_agent, a terminal coding agent example with default stable text-protocol tool mode.
  • Added persisted settings log-level fallback handling with regression tests.

0.6.2#

  • Native inference performance improvements (request overhead, stream batching, and prompt-prefix reuse with parity-safe fallback).
  • Added native benchmark and prompt-reuse parity tooling, plus CI parity coverage.

0.6.1#

  • Publishing compatibility fix for hook backend-config code paths.
  • Continued parity hardening around template/parser behavior.

0.6.x line highlights#

  • Expanded llama.cpp template and parser parity.
  • Stronger handling for tool payload fidelity.
  • More deterministic behavior around template routing and fallback removal.

0.5.x line highlights#

  • Public API tightening and migration cleanup.
  • Split Dart/native log controls.
  • Example/runtime reliability improvements.

Release usage guidance#

  • For upgrade planning, combine this page with Upgrade Checklist.
  • For breaking changes, always validate against the exact release tag notes.
Edit this page

Searches the latest release. Esc to close.