local_ai_llama_cpp 0.0.3

llama.cpp adapter for LocalAI Kit: runs any GGUF model as a LocalLlm (streaming, GBNF-constrained structured output) and as a LocalEmbedding, with all FFI work isolated in worker isolates.

local_ai_llama_cpp

llama.cpp adapter for LocalAI Kit: runs any GGUF model behind the core LocalLlm and LocalEmbedding interfaces, registered through the normal AdapterRegistry under the provider key llama-cpp.

import 'package:local_ai_kit/local_ai_kit.dart';
import 'package:local_ai_llama_cpp/local_ai_llama_cpp.dart';

final ai = await LocalAI.initialize(
  const LocalAIConfig(
    llm: LlmConfig(modelId: 'qwen-2.5-0.5b-instruct-gguf'),
    embedding: EmbeddingConfig(modelId: 'nomic-embed-text-v1.5-gguf'),
  ),
  plugins: const [LlamaCppAdapterPlugin()],
);

final answer = await ai.generate('Hello, on-device world!');
final vector = await ai.embed('semantic search query');

What it does

• Persistent KV cache — one llama_context stays alive for the whole load()…unload() lifetime, so an unchanged history prefix is not re-evaluated on every turn (PromptPlanner decides when that is safe).
• Real grammar-constrained structured output — generateStructured compiles the requested JsonSchema into a GBNF grammar (JsonSchemaToGbnf) and hands it to llama.cpp's sampler, so invalid JSON is unreachable rather than merely discouraged. The parse-and-retry loop stays as a defensive fallback.
• First real LocalEmbedding — LlamaCppEmbeddingAdapter loads an embedding-mode GGUF (nomic-embed, bge, e5, …) with mean pooling, and supports Matryoshka truncation through EmbeddingConfig.dimensions.
• Native sampling — topK/topP/temperature and repetition penalty are llama.cpp's own samplers, not a Dart-side heuristic.
• Truthful telemetry — real promptTokens/completionTokens and a finishReason that distinguishes stop, length and cancelled.
• Bring your own GGUF — ai.models.registerExternalModel(manifest, localFilePath: …) links a file you already have on disk into the standard install location; no download, no verification.

All llama.cpp calls happen in a dedicated worker isolate, so the UI isolate never blocks, and no llama_cpp_dart type appears in this package's public API.

The native library

llama_cpp_dart ships Dart FFI bindings only — it is not a Flutter plugin, so no binary reaches your app through it. Build and bundle llama.cpp yourself with the scripts in native/, then either use the default library name for your platform or point the adapter at yours:

LlamaCppRuntime.useLibrary('/absolute/path/to/libmtmd.so');

Known limits

• Multi-turn cache reuse is disabled for tokenizers that prepend special tokens (Llama 3, Gemma, …). llama_cpp_dart's setPrompt always tokenizes with add_special = true, which would inject a second BOS token mid-conversation; rather than corrupt the history, those models re-evaluate the full prompt each turn. Check LlamaCppLlmAdapter.reusesContextAcrossTurns.
• Changing sampling mid-session rebuilds the context. llama.cpp builds its sampler chain with the context and llama_cpp_dart cannot swap it, so a different temperature/topP — or a structured-output grammar — costs a rebuild and drops the KV cache. Weights come back from the OS page cache, so it is far cheaper than a cold load, but it is not free: prefer batching structured-output calls together rather than interleaving them with chat.
• Chat templates are inferred from the model id / file name, not read from the GGUF metadata — llama_cpp_dart does not expose llama_model_chat_template. ChatTemplate.detect covers ChatML, Gemma, Llama 3, Mistral and Phi; anything unrecognised falls back to a plain User:/Assistant: format.
• No vision. The native build includes mtmd, but multimodal GGUF input is not wired to any core interface yet.
• The native build scripts are not exercised in CI; see native/README.md.

Documentation

See docs/adapters.md for how this adapter fits the kit's layering, and the design spec in docs/superpowers/specs/2026-08-25-llama-cpp-adapter-design.md.

License

Apache-2.0.