Mullama

Comprehensive Rust bindings for llama.cpp with advanced integration features

Mullama provides memory-safe Rust bindings for llama.cpp with production-ready features including async/await support, real-time streaming, multimodal processing, and web framework integration.

Why Mullama?

Most llama.cpp Rust bindings expose low-level C APIs directly. Mullama provides an idiomatic Rust experience:

// Other wrappers: manual memory management, raw pointers, verbose setup
let params = llama_context_default_params();
let ctx = unsafe { llama_new_context_with_model(model, params) };
let tokens = unsafe { llama_tokenize(model, text.as_ptr(), ...) };
// Don't forget to free everything...

// Mullama: builder patterns, async/await, automatic resource management
let model = ModelBuilder::new()
    .path("model.gguf")
    .gpu_layers(35)
    .build().await?;

let response = model.generate("Hello", 100).await?;

Developer experience improvements:

Key Features

• Async/Await Native - Full Tokio integration for non-blocking operations
• Real-time Streaming - Token-by-token generation with backpressure handling
• Multimodal Processing - Text, image, and audio in a single pipeline
• Web Framework Ready - Direct Axum integration with REST APIs
• WebSocket Support - Real-time bidirectional communication
• Parallel Processing - Work-stealing parallelism for batch operations
• GPU Acceleration - CUDA, Metal, ROCm, and OpenCL support
• Memory Safe - Zero unsafe operations in public API

Quick Start

Installation

[dependencies]
mullama = "0.1.0"

# With all features
mullama = { version = "0.1.0", features = ["full"] }

Prerequisites

Linux (Ubuntu/Debian):

sudo apt install -y build-essential cmake pkg-config libasound2-dev libpulse-dev

macOS:

brew install cmake pkg-config portaudio

Windows: Install Visual Studio Build Tools and CMake.

See Platform Setup Guide for detailed instructions.

Basic Example

use mullama::prelude::*;

#[tokio::main]
async fn main() -> Result<(), MullamaError> {
    let model = ModelBuilder::new()
        .path("model.gguf")
        .context_size(4096)
        .build().await?;

    let response = model.generate("The future of AI is", 100).await?;
    println!("{}", response);

    Ok(())
}

Feature Flags

[dependencies.mullama]
version = "0.1.0"
features = [
    "async",              # Async/await support
    "streaming",          # Token streaming
    "web",                # Axum web framework
    "websockets",         # WebSocket support
    "multimodal",         # Image and audio processing
    "streaming-audio",    # Real-time audio capture
    "format-conversion",  # Audio/image format conversion
    "parallel",           # Rayon parallel processing
    "daemon",             # Daemon mode with TUI client
    "full"                # All features
]

Common Combinations

# Web applications
features = ["web", "websockets", "async", "streaming"]

# Multimodal AI
features = ["multimodal", "streaming-audio", "format-conversion"]

# High-performance batch processing
features = ["parallel", "async"]

# Daemon with TUI chat interface
features = ["daemon"]

Daemon Mode

Mullama includes a multi-model daemon with OpenAI-compatible HTTP API and TUI client:

# Build the CLI
cargo build --release --features daemon

# Start daemon with local model
mullama serve --model llama:./llama.gguf

# Start with HuggingFace model (auto-downloads and caches)
mullama serve --model hf:TheBloke/Llama-2-7B-GGUF

# Multiple models with custom aliases
mullama serve \
  --model llama:hf:TheBloke/Llama-2-7B-GGUF:llama-2-7b.Q4_K_M.gguf \
  --model mistral:hf:TheBloke/Mistral-7B-v0.1-GGUF

# Interactive TUI chat
mullama chat

# One-shot generation
mullama run "What is the meaning of life?"

# Model management
mullama models            # List loaded models
mullama load phi:./phi.gguf  # Load a model
mullama unload phi        # Unload a model
mullama default llama     # Set default model

# Search for models on HuggingFace
mullama search "llama 7b"          # Search GGUF models
mullama search "mistral" --files   # Show available files
mullama search "phi" --all         # Include non-GGUF models
mullama info TheBloke/Llama-2-7B-GGUF  # Show repo details

# Cache management
mullama pull hf:TheBloke/Llama-2-7B-GGUF  # Pre-download model
mullama cache list        # List cached models
mullama cache size        # Show cache size
mullama cache clear       # Clear cache

# Use OpenAI-compatible API
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "llama", "messages": [{"role": "user", "content": "Hello!"}]}'

HuggingFace Model Format

hf:<owner>/<repo>:<filename>   # Specific file
hf:<owner>/<repo>              # Auto-detect best GGUF
<alias>:hf:<owner>/<repo>      # With custom alias

Environment Variables

Cache Locations (Cross-Platform)

Architecture:

                                   ┌──────────────────────────────────┐
                                   │           Daemon                 │
┌─────────────┐                    │  ┌────────────────────────────┐  │
│  TUI Client │◄── nng (IPC) ─────►│  │     Model Manager          │  │
└─────────────┘                    │  │  ┌───────┐  ┌───────┐      │  │
                                   │  │  │Model 1│  │Model 2│ ...  │  │
┌─────────────┐                    │  │  └───────┘  └───────┘      │  │
│   curl/app  │◄── HTTP/REST ─────►│  └────────────────────────────┘  │
└─────────────┘   (OpenAI API)     │                                  │
                                   │  Endpoints:                      │
┌─────────────┐                    │  • /v1/chat/completions          │
│ Other Client│◄── nng (IPC) ─────►│  • /v1/completions               │
└─────────────┘                    │  • /v1/models                    │
                                   │  • /v1/embeddings                │
                                   └──────────────────────────────────┘

Programmatic usage:

use mullama::daemon::{DaemonClient, DaemonBuilder};

// Connect as client
let client = DaemonClient::connect_default()?;
let result = client.chat("Hello, AI!", None, 100, 0.7)?;
println!("{} ({:.1} tok/s)", result.text, result.tokens_per_second());

// List models
for model in client.list_models()? {
    println!("{}: {}M params", model.alias, model.info.parameters / 1_000_000);
}

GPU Acceleration

# NVIDIA CUDA
export LLAMA_CUDA=1

# Apple Metal (macOS)
export LLAMA_METAL=1

# AMD ROCm (Linux)
export LLAMA_HIPBLAS=1

# Intel OpenCL
export LLAMA_CLBLAST=1

Documentation

Examples

# Basic text generation
cargo run --example simple --features async

# Streaming responses
cargo run --example streaming_generation --features "async,streaming"

# Web service
cargo run --example web_service --features "web,websockets"

# Audio processing
cargo run --example streaming_audio_demo --features "streaming-audio,multimodal"

Contributing

We welcome contributions! See CONTRIBUTING.md for guidelines.

git clone --recurse-submodules https://github.com/neul-labs/mullama.git
cd mullama
cargo test --all-features

License

MIT License - see LICENSE for details.

Acknowledgments

• llama.cpp - The underlying inference engine
• ggml - Tensor operations library