metal-candle

Production-quality Rust ML library for Apple Silicon - LoRA training, text generation, and semantic embeddings

Overview

Pure Rust machine learning library optimized for Apple Silicon:

• LoRA Training: Fine-tune transformer models efficiently
• Text Generation: Streaming, multiple sampling strategies, repetition penalty
• Semantic Embeddings: E5, MiniLM, MPNet models for RAG and search
• Metal Acceleration: Native GPU acceleration on M-series chips

Why metal-candle? 25.9x faster than MLX for embeddings, single binary deployment, type-safe ML, production-ready (407 tests, 81.6% coverage)

Performance

metal-candle demonstrates exceptional performance on Apple Silicon:

Near constant-time performance: Batch 1→100 only increases by 13% (3.9ms → 4.4ms)

See BENCHMARKS.md for detailed performance analysis and methodology.

Installation

[dependencies]
metal-candle = "1.2"  # or latest from crates.io

Requirements: Rust 1.75+, Apple Silicon (M1/M2/M3/M4), macOS 12.0+

Quick Start

Text Generation

use metal_candle::inference::{Generator, GeneratorConfig, SamplingStrategy};
use metal_candle::models::Qwen;

// Load model
let model = Qwen::new(&config, vb)?;

// Configure generation
let gen_config = GeneratorConfig {
    max_tokens: 128,
    sampling: SamplingStrategy::TopP { p: 0.95 },
    temperature: 0.7,
    repetition_penalty: 1.1,  // Reduce repetition
    stop_on_eos: true,
    eos_token_id: Some(151643),  // Qwen EOS token
    ..Default::default()
};

// Generate tokens
let mut generator = Generator::new(Box::new(model), gen_config)?;
let output_ids = generator.generate(&input_ids)?;

// Or use streaming for real-time generation (v1.3.0+)
generator.generate_stream(&input_ids, |token| {
    println!("Token {}: prob={:.2}%", token.token_id, token.probability * 100.0);
    true // Continue generation
})?;

// Async streaming (requires 'streaming' feature)
#[cfg(feature = "streaming")]
{
    use futures::stream::StreamExt;
    use futures::pin_mut;
    
    let stream = generator.generate_stream_async(&input_ids);
    pin_mut!(stream);
    
    while let Some(result) = stream.next().await {
        let token = result?;
        println!("Token: {}", token.token_id);
    }
}

Semantic Embeddings (RAG & Search)

use metal_candle::embeddings::{EmbeddingModel, EmbeddingModelType};
use metal_candle::Device;

// Load embedding model with Metal acceleration (25.9x faster than MLX!)
let device = Device::new_metal(0)?;
let model = EmbeddingModel::from_pretrained(
    EmbeddingModelType::E5SmallV2,
    device,
)?;

// Generate embeddings for semantic search
let texts = vec![
    "Rust is a systems programming language",
    "Python is a high-level language",
];
let embeddings = model.encode(&texts)?;  // [batch, 384] in 3.9ms

// Batch processing: 100 docs in 4.4ms (22,831 docs/sec throughput)
let large_corpus = load_documents()?;
let batch_embeddings = model.encode(&large_corpus)?;

LoRA Training

use metal_candle::training::{
    LoRAAdapter, LoRAAdapterConfig, TargetModule,
    Trainer, TrainingConfig, LRScheduler
};

// Create LoRA adapter
let lora_config = LoRAAdapterConfig {
    rank: 8,
    alpha: 16.0,
    dropout: 0.0,
    target_modules: vec![TargetModule::QProj, TargetModule::VProj],
};
let adapter = LoRAAdapter::new(&model, lora_config, &device)?;

// Configure and train
let training_config = TrainingConfig {
    num_epochs: 3,
    lr_scheduler: LRScheduler::warmup_cosine(100, 1000, 1e-4, 1e-6),
    ..Default::default()
};
let trainer = Trainer::new(adapter, training_config)?;
let metrics = trainer.train(&dataset)?;

LoRA Adapter Management (v1.3.0+)

use metal_candle::training::{AdapterRegistry, LoRAAdapter, LoRAAdapterConfig};

// Create registry for managing multiple adapters
let mut registry = AdapterRegistry::new();

// Load task-specific adapters
let code_adapter = LoRAAdapter::new(768, 3072, 12, &config, &device)?;
let chat_adapter = LoRAAdapter::new(768, 3072, 12, &config, &device)?;

registry.add_adapter("code-assistant".to_string(), code_adapter)?;
registry.add_adapter("chat".to_string(), chat_adapter)?;

// Switch between adapters without reloading base model
registry.activate("code-assistant")?;
// ... use model for code generation ...

registry.activate("chat")?;
// ... use model for chat ...

// Memory efficient: adapters are ~0.03% of base model size
println!("Active adapter: {:?}", registry.active_adapter());

Features

• Training: LoRA with dropout, AdamW optimizer, learning rate schedulers, checkpoint management, adapter registry (v1.3.0+)
• Inference: KV-cache, multiple sampling strategies, streaming generation (sync & async), repetition penalty, rich token metadata (v1.3.0+)
• Models: Qwen2.5-Coder, safetensors format, transformer components (RoPE, GQA, MLP)
• Embeddings: E5, MiniLM, MPNet with HuggingFace Hub integration
• Quality: 407 tests, 81.6% coverage, strict clippy linting, 100% API documentation

Architecture

Built on Candle with Metal backend:

┌─────────────────────────────────────────────────────────────┐
│                    metal-candle (Public API)                 │
├─────────────────────────────────────────────────────────────┤
│  Training          │  Inference        │  Models            │
│  • LoRAAdapter     │  • KVCache        │  • ModelLoader     │
│  • Trainer         │  • Sampling       │  • Qwen           │
│  • AdamW           │  • Generator      │  • Config          │
│  • Schedulers      │                   │                    │
│  • Checkpoint      │  Embeddings       │                    │
│                    │  • EmbeddingModel │                    │
│                    │  • E5/MiniLM/MPNet│                    │
└─────────────────────────────────────────────────────────────┘
                            │
┌─────────────────────────────────────────────────────────────┐
│                      Candle Framework                        │
│  • Tensor operations  • Metal backend  • Autograd           │
└─────────────────────────────────────────────────────────────┘
                            │
┌─────────────────────────────────────────────────────────────┐
│                      Apple Metal API                         │
│  (GPU acceleration on Apple Silicon)                        │
└─────────────────────────────────────────────────────────────┘

See ARCHITECTURE.md for detailed architecture documentation.

Documentation

• API Reference - Complete API documentation
• Architecture Guide - System design and implementation details
• Contributing Guide - Development standards and guidelines
• Benchmarks - Performance analysis and methodology
• Project Plan - Development roadmap and future plans

Examples

Run examples:

cargo run --example generate_text
cargo run --example train_lora
cargo run --example embeddings_demo --features embeddings

Development

git clone https://github.com/GarthDB/metal-candle.git
cd metal-candle
cargo build && cargo test

See CONTRIBUTING.md for full guidelines. Quality standards: zero clippy warnings (pedantic), ≥80% coverage, 100% API docs.

Roadmap

See ROADMAP.md for detailed release plans and NEXT_STEPS.md for immediate priorities.

Upcoming Releases

• v1.3.1 (Jan 2025): ApplyAdapter implementation, streaming benchmarks
• v1.4.0 (Feb 2025): GGUF format support
• v1.5.0 (Mar 2025): LLaMA/Mistral architectures
• v1.6.0 (Apr 2025): 4-bit/8-bit quantization
• v1.7.0 (May 2025): Flash Attention
• v2.0.0 (Q3 2025): Multi-GPU support

Track progress on the v1.3+ Feature Roadmap project board. Vote with 👍 on issues you'd like to see prioritized!

Contributing

Contributions welcome! See CONTRIBUTING.md for development standards and testing requirements.

License

Licensed under Apache-2.0 (LICENSE). Provides explicit patent protection for production ML.

Acknowledgments

• Built on the excellent Candle framework by Hugging Face
• Inspired by MLX and llama.cpp
• LoRA implementation based on LoRA paper

Known Advisories

Two unmaintained transitive dependencies (non-security): number_prefix, paste from trusted upstream (Candle, HF). See deny.toml for details.

Support

• Issues: GitHub Issues
• Discussions: GitHub Discussions
• Documentation: ARCHITECTURE.md | CONTRIBUTING.md

Maintained by: @GarthDB