llama-cpp Capacitor Plugin

A native Capacitor plugin that embeds llama.cpp directly into mobile apps, enabling offline AI inference with comprehensive support for text generation, multimodal processing, TTS, LoRA adapters, and more.

llama.cpp: Inference of LLaMA model in pure C/C++

🚀 Features

• Offline AI Inference: Run large language models completely offline on mobile devices
• Text Generation: Complete text completion with streaming support
• Chat Conversations: Multi-turn conversations with context management
• Multimodal Support: Process images and audio alongside text
• Text-to-Speech (TTS): Generate speech from text using vocoder models
• LoRA Adapters: Fine-tune models with LoRA adapters
• Embeddings: Generate vector embeddings for semantic search
• Reranking: Rank documents by relevance to queries
• Session Management: Save and load conversation states
• Benchmarking: Performance testing and optimization tools
• Structured Output: Generate JSON with schema validation
• Cross-Platform: iOS and Android support with native optimizations

✅ Complete Implementation Status

This plugin is now FULLY IMPLEMENTED with complete native integration of llama.cpp for both iOS and Android platforms. The implementation includes:

Completed Features

• Complete C++ Integration: Full llama.cpp library integration with all core components
• Native Build System: CMake-based build system for both iOS and Android
• Platform Support: iOS (arm64, x86_64) and Android (arm64-v8a, armeabi-v7a, x86, x86_64)
• TypeScript API: Complete TypeScript interface matching llama.rn functionality
• Native Methods: All 30+ native methods implemented with proper error handling
• Event System: Capacitor event system for progress and token streaming
• Documentation: Comprehensive README and API documentation

Technical Implementation

• C++ Core: Complete llama.cpp library with GGML, GGUF, and all supporting components
• iOS Framework: Native iOS framework with Metal acceleration support
• Android JNI: Complete JNI implementation with multi-architecture support
• Build Scripts: Automated build system for both platforms
• Error Handling: Robust error handling and result types

Project Structure

llama-cpp/
├── cpp/                    # Complete llama.cpp C++ library
│   ├── ggml.c             # GGML core
│   ├── gguf.cpp           # GGUF format support
│   ├── llama.cpp          # Main llama.cpp implementation
│   ├── rn-llama.cpp       # React Native wrapper (adapted)
│   ├── rn-completion.cpp  # Completion handling
│   ├── rn-tts.cpp         # Text-to-speech
│   └── tools/mtmd/        # Multimodal support
├── ios/
│   ├── CMakeLists.txt     # iOS build configuration
│   └── Sources/           # Swift implementation
├── android/
│   ├── src/main/
│   │   ├── CMakeLists.txt # Android build configuration
│   │   ├── jni.cpp        # JNI implementation
│   │   └── jni-utils.h    # JNI utilities
│   └── build.gradle       # Android build config
├── src/
│   ├── definitions.ts     # Complete TypeScript interfaces
│   ├── index.ts           # Main plugin implementation
│   └── web.ts             # Web fallback
└── build-native.sh        # Automated build script

📦 Installation

npm install llama-cpp-capacitor

🔨 Building the Native Library

The plugin includes a complete native implementation of llama.cpp. To build the native libraries:

Prerequisites

• CMake (3.16+ for iOS, 3.10+ for Android)
• Xcode (for iOS builds, macOS only)
• Android Studio with NDK (for Android builds)
• Make or Ninja build system

Automated Build

# Build for all platforms
npm run build:native

# Build for specific platforms
npm run build:ios      # iOS only
npm run build:android  # Android only

# Clean native builds
npm run clean:native

Manual Build

iOS Build

cd ios
cmake -B build -S .
cmake --build build --config Release

Android Build

cd android
./gradlew assembleRelease

Build Output

• iOS: ios/build/LlamaCpp.framework/
• Android: android/src/main/jniLibs/{arch}/libllama-cpp-{arch}.so

Updating the native source (e.g. for vision model support)

The native cpp/ layer is based on llama.cpp. To pull in a newer upstream version (e.g. for vision model support) without overwriting the Capacitor adapter code, use the bootstrap script (included in this repo):

./scripts/bootstrap.sh [branch-or-tag-or-commit]
# Example: ./scripts/bootstrap.sh master

This syncs upstream into cpp/ and keeps project-specific files (cap-*.cpp/h, tools/mtmd/, etc.) intact. After running it, reconcile any API changes in the adapter code, then rebuild with npm run build:native or ./build-native.sh. See cpp/README.md and docs/IOS_IMPLEMENTATION_GUIDE.md.

iOS implementation (step-by-step)

For a step-by-step guide on how methods are implemented on the iOS side (Swift bridge → native framework, adding/updating C symbols, and updating the native layer for vision), see docs/IOS_IMPLEMENTATION_GUIDE.md.

iOS Setup

1. Install the plugin:

npm install llama-cpp

2. Add to your iOS project:

npx cap add ios
npx cap sync ios

3. Open the project in Xcode:

npx cap open ios

Android Setup

1. Install the plugin:

npm install llama-cpp

2. Add to your Android project:

npx cap add android
npx cap sync android

3. Open the project in Android Studio:

npx cap open android

🎯 Quick Start

Basic Text Completion

import { initLlama } from 'llama-cpp';

// Initialize a model
const context = await initLlama({
  model: '/path/to/your/model.gguf',
  n_ctx: 2048,
  n_threads: 4,
  n_gpu_layers: 0,
});

// Generate text
const result = await context.completion({
  prompt: "Hello, how are you today?",
  n_predict: 50,
  temperature: 0.8,
});

console.log('Generated text:', result.text);

Chat-Style Conversations

const result = await context.completion({
  messages: [
    { role: "system", content: "You are a helpful AI assistant." },
    { role: "user", content: "What is the capital of France?" },
    { role: "assistant", content: "The capital of France is Paris." },
    { role: "user", content: "Tell me more about it." }
  ],
  n_predict: 100,
  temperature: 0.7,
});

console.log('Chat response:', result.content);

Streaming Completion

let fullText = '';
const result = await context.completion({
  prompt: "Write a short story about a robot learning to paint:",
  n_predict: 150,
  temperature: 0.8,
}, (tokenData) => {
  // Called for each token as it's generated
  fullText += tokenData.token;
  console.log('Token:', tokenData.token);
});

console.log('Final result:', result.text);

🚀 Mobile-Optimized Speculative Decoding

Achieve 2-8x faster inference with significantly reduced battery consumption!

Speculative decoding uses a smaller "draft" model to predict multiple tokens ahead, which are then verified by the main model. This results in dramatic speedups with identical output quality.

Basic Usage

import { initLlama } from 'llama-cpp-capacitor';

// Initialize with speculative decoding
const context = await initLlama({
  model: '/path/to/your/main-model.gguf',         // Main model (e.g., 7B)
  draft_model: '/path/to/your/draft-model.gguf', // Draft model (e.g., 1.5B)
  
  // Speculative decoding parameters
  speculative_samples: 3,      // Number of tokens to predict speculatively
  mobile_speculative: true,    // Enable mobile optimizations
  
  // Standard parameters
  n_ctx: 2048,
  n_threads: 4,
});

// Use normally - speculative decoding is automatic
const result = await context.completion({
  prompt: "Write a story about AI:",
  n_predict: 200,
  temperature: 0.7,
});

console.log('🚀 Generated with speculative decoding:', result.text);

Mobile-Optimized Configuration

// Recommended mobile setup for best performance/battery balance
const mobileContext = await initLlama({
  // Quantized models for mobile efficiency
  model: '/models/llama-2-7b-chat.q4_0.gguf',
  draft_model: '/models/tinyllama-1.1b-chat.q4_0.gguf',
  
  // Conservative mobile settings
  n_ctx: 1024,                 // Smaller context for mobile
  n_threads: 3,                // Conservative threading
  n_batch: 64,                 // Smaller batch size
  n_gpu_layers: 24,            // Utilize mobile GPU
  
  // Optimized speculative decoding
  speculative_samples: 3,      // 2-3 tokens ideal for mobile
  mobile_speculative: true,    // Enables mobile-specific optimizations
  
  // Memory optimizations
  use_mmap: true,              // Memory mapping for efficiency
  use_mlock: false,            // Don't lock memory on mobile
});

Performance Benefits

• 2-8x faster inference - Dramatically reduced time to generate text
• 50-80% battery savings - Less time computing = longer battery life
• Identical output quality - Same text quality as regular decoding
• Automatic fallback - Falls back to regular decoding if draft model fails
• Mobile optimized - Specifically tuned for mobile device constraints

Model Recommendations

Error Handling & Fallback

// Robust setup with automatic fallback
try {
  const context = await initLlama({
    model: '/models/main-model.gguf',
    draft_model: '/models/draft-model.gguf',
    speculative_samples: 3,
    mobile_speculative: true,
  });
  console.log('✅ Speculative decoding enabled');
} catch (error) {
  console.warn('⚠️ Falling back to regular decoding');
  const context = await initLlama({
    model: '/models/main-model.gguf',
    // No draft_model = regular decoding
  });
}

📚 API Reference

Core Functions

initLlama(params: ContextParams, onProgress?: (progress: number) => void): Promise<LlamaContext>

Initialize a new llama.cpp context with a model.

Parameters:

• params: Context initialization parameters
• onProgress: Optional progress callback (0-100)

Returns: Promise resolving to a LlamaContext instance

releaseAllLlama(): Promise<void>

Release all contexts and free memory.

toggleNativeLog(enabled: boolean): Promise<void>

Enable or disable native logging.

addNativeLogListener(listener: (level: string, text: string) => void): { remove: () => void }

Add a listener for native log messages.

LlamaContext Class

completion(params: CompletionParams, callback?: (data: TokenData) => void): Promise<NativeCompletionResult>

Generate text completion.

Parameters:

• params: Completion parameters including prompt or messages
• callback: Optional callback for token-by-token streaming

tokenize(text: string, options?: { media_paths?: string[] }): Promise<NativeTokenizeResult>

Tokenize text or text with images.

detokenize(tokens: number[]): Promise<string>

Convert tokens back to text.

embedding(text: string, params?: EmbeddingParams): Promise<NativeEmbeddingResult>

Generate embeddings for text.

rerank(query: string, documents: string[], params?: RerankParams): Promise<RerankResult[]>

Rank documents by relevance to a query.

bench(pp: number, tg: number, pl: number, nr: number): Promise<BenchResult>

Benchmark model performance.

Multimodal Support

initMultimodal(params: { path: string; use_gpu?: boolean }): Promise<boolean>

Initialize multimodal support with a projector file.

isMultimodalEnabled(): Promise<boolean>

Check if multimodal support is enabled.

getMultimodalSupport(): Promise<{ vision: boolean; audio: boolean }>

Get multimodal capabilities.

releaseMultimodal(): Promise<void>

Release multimodal resources.

TTS (Text-to-Speech)

initVocoder(params: { path: string; n_batch?: number }): Promise<boolean>

Initialize TTS with a vocoder model.

isVocoderEnabled(): Promise<boolean>

Check if TTS is enabled.

getFormattedAudioCompletion(speaker: object | null, textToSpeak: string): Promise<{ prompt: string; grammar?: string }>

Get formatted audio completion prompt.

getAudioCompletionGuideTokens(textToSpeak: string): Promise<Array<number>>

Get guide tokens for audio completion.

decodeAudioTokens(tokens: number[]): Promise<Array<number>>

Decode audio tokens to audio data.

releaseVocoder(): Promise<void>

Release TTS resources.

LoRA Adapters

applyLoraAdapters(loraList: Array<{ path: string; scaled?: number }>): Promise<void>

Apply LoRA adapters to the model.

removeLoraAdapters(): Promise<void>

Remove all LoRA adapters.

getLoadedLoraAdapters(): Promise<Array<{ path: string; scaled?: number }>>

Get list of loaded LoRA adapters.

Session Management

saveSession(filepath: string, options?: { tokenSize: number }): Promise<number>

Save current session to a file.

loadSession(filepath: string): Promise<NativeSessionLoadResult>

Load session from a file.

🔧 Configuration

Context Parameters

interface ContextParams {
  model: string;                    // Path to GGUF model file
  n_ctx?: number;                   // Context size (default: 512)
  n_threads?: number;               // Number of threads (default: 4)
  n_gpu_layers?: number;            // GPU layers (iOS only)
  use_mlock?: boolean;              // Lock memory (default: false)
  use_mmap?: boolean;               // Use memory mapping (default: true)
  embedding?: boolean;              // Embedding mode (default: false)
  cache_type_k?: string;            // KV cache type for K
  cache_type_v?: string;            // KV cache type for V
  pooling_type?: string;            // Pooling type
  // ... more parameters
}

Completion Parameters

interface CompletionParams {
  prompt?: string;                  // Text prompt
  messages?: Message[];             // Chat messages
  n_predict?: number;               // Max tokens to generate
  temperature?: number;             // Sampling temperature
  top_p?: number;                   // Top-p sampling
  top_k?: number;                   // Top-k sampling
  stop?: string[];                  // Stop sequences
  // ... more parameters
}

📱 Platform Support

🎨 Advanced Examples

Multimodal Processing

// Initialize multimodal support
await context.initMultimodal({
  path: '/path/to/mmproj.gguf',
  use_gpu: true,
});

// Process image with text
const result = await context.completion({
  messages: [
    { 
      role: "user", 
      content: [
        { type: "text", text: "What do you see in this image?" },
        { type: "image_url", image_url: { url: "file:///path/to/image.jpg" } }
      ]
    }
  ],
  n_predict: 100,
});

console.log('Image analysis:', result.content);

Text-to-Speech

// Initialize TTS
await context.initVocoder({
  path: '/path/to/vocoder.gguf',
  n_batch: 512,
});

// Generate audio
const audioCompletion = await context.getFormattedAudioCompletion(
  null, // Speaker configuration
  "Hello, this is a test of text-to-speech functionality."
);

const guideTokens = await context.getAudioCompletionGuideTokens(
  "Hello, this is a test of text-to-speech functionality."
);

const audioResult = await context.completion({
  prompt: audioCompletion.prompt,
  grammar: audioCompletion.grammar,
  guide_tokens: guideTokens,
  n_predict: 1000,
});

const audioData = await context.decodeAudioTokens(audioResult.audio_tokens);

LoRA Adapters

// Apply LoRA adapters
await context.applyLoraAdapters([
  { path: '/path/to/adapter1.gguf', scaled: 1.0 },
  { path: '/path/to/adapter2.gguf', scaled: 0.5 }
]);

// Check loaded adapters
const adapters = await context.getLoadedLoraAdapters();
console.log('Loaded adapters:', adapters);

// Generate with adapters
const result = await context.completion({
  prompt: "Test prompt with LoRA adapters:",
  n_predict: 50,
});

// Remove adapters
await context.removeLoraAdapters();

Structured Output

JSON Schema (Auto-converted to GBNF)

const result = await context.completion({
  prompt: "Generate a JSON object with a person's name, age, and favorite color:",
  n_predict: 100,
  response_format: {
    type: 'json_schema',
    json_schema: {
      strict: true,
      schema: {
        type: 'object',
        properties: {
          name: { type: 'string' },
          age: { type: 'number' },
          favorite_color: { type: 'string' }
        },
        required: ['name', 'age', 'favorite_color']
      }
    }
  }
});

console.log('Structured output:', result.content);

Direct GBNF Grammar

// Define GBNF grammar directly for maximum control
const grammar = `
root ::= "{" ws name_field "," ws age_field "," ws color_field "}"
name_field ::= "\\"name\\"" ws ":" ws string_value
age_field ::= "\\"age\\"" ws ":" ws number_value  
color_field ::= "\\"favorite_color\\"" ws ":" ws string_value
string_value ::= "\\"" [a-zA-Z ]+ "\\""
number_value ::= [0-9]+
ws ::= [ \\t\\n]*
`;

const result = await context.completion({
  prompt: "Generate a person's profile:",
  grammar: grammar,
  n_predict: 100
});

console.log('Grammar-constrained output:', result.text);

Manual JSON Schema to GBNF Conversion

import { convertJsonSchemaToGrammar } from 'llama-cpp-capacitor';

const schema = {
  type: 'object',
  properties: {
    name: { type: 'string' },
    age: { type: 'number' }
  },
  required: ['name', 'age']
};

// Convert schema to GBNF grammar
const grammar = await convertJsonSchemaToGrammar(schema);
console.log('Generated grammar:', grammar);

const result = await context.completion({
  prompt: "Generate a person:",
  grammar: grammar,
  n_predict: 100
});

🔍 Model Compatibility

This plugin supports GGUF format models, which are compatible with llama.cpp. You can find GGUF models on Hugging Face by searching for the "GGUF" tag.

Recommended Models

• Llama 2: Meta's latest language model
• Mistral: High-performance open model
• Code Llama: Specialized for code generation
• Phi-2: Microsoft's efficient model
• Gemma: Google's open model

Model Quantization

For mobile devices, consider using quantized models (Q4_K_M, Q5_K_M, etc.) to reduce memory usage and improve performance.

⚡ Performance Considerations

Memory Management

• Use quantized models for better memory efficiency
• Adjust n_ctx based on your use case
• Monitor memory usage with use_mlock: false

GPU Acceleration

• iOS: Set n_gpu_layers to use Metal GPU acceleration
• Android: GPU acceleration is automatically enabled when available

Threading

• Adjust n_threads based on device capabilities
• More threads may improve performance but increase memory usage

🐛 Troubleshooting

Common Issues

1. Model not found: Ensure the model path is correct and the file exists
2. Out of memory: Try using a quantized model or reducing n_ctx
3. Slow performance: Enable GPU acceleration or increase n_threads
4. Multimodal not working: Ensure the mmproj file is compatible with your model

Debugging

Enable native logging to see detailed information:

import { toggleNativeLog, addNativeLogListener } from 'llama-cpp';

await toggleNativeLog(true);

const logListener = addNativeLogListener((level, text) => {
  console.log(`[${level}] ${text}`);
});

📦 Publishing

To publish the package to npm:

1. Build (runs automatically on npm publish via prepublishOnly): npm run build — produces dist/ (plugin bundles, ESM, docs).
2. Optional — include native libs in the tarball: npm run build:all (requires macOS/NDK) — builds iOS framework and Android .so into ios/Frameworks and android/src/main/jniLibs.
3. Verify pack: npm run pack (JS only) or npm run pack:full (JS + native) — lists files that would be published.
4. Publish: npm publish.

See NPM_PUBLISH_GUIDE.md for 2FA/token setup and troubleshooting.

🤝 Contributing

We welcome contributions! Please see our Contributing Guide for details.

📄 License

This project is licensed under the MIT License - see the LICENSE file for details.

🙏 Acknowledgments

• llama.cpp - The core inference engine
• Capacitor - The cross-platform runtime
• llama.rn - Inspiration for the React Native implementation

📞 Support

• 📧 Email: support@arusatech.com
• 🐛 Issues: GitHub Issues
• 📖 Documentation: GitHub Wiki