liquid-edge 0.2.3

Minimal edge inference runtime for LLMs
Documentation

Liquid Edge - High-Performance Edge Inference Runtime

Liquid Edge is a production-ready inference runtime designed specifically for edge computing environments. It provides high-performance LLM inference with multiple backend support, comprehensive tokenization capabilities, and optimized memory management.

🚀 Features

  • 🎯 Multiple Backends: ONNX Runtime support with more backends planned
  • ⚡ High Performance: Optimized for edge devices with minimal latency
  • 🔄 Async Support: Full async/await support for non-blocking inference
  • 💬 Chat Templates: Jinja2 template support for conversational AI
  • 📊 Comprehensive Monitoring: Built-in metrics and performance tracking
  • 🛡️ Production Ready: Robust error handling and logging
  • 🔧 Flexible Configuration: Feature flags for customized builds
pipx install "optimum[onnxruntime]"
optimum-cli export onnx \                                                                                                                                               ✔ │ 7s   │ rust-onnx-chat 
      --model squeeze-ai-lab/TinyAgent-1.1B \
      --task text-generation ./models/tinyagen

Model Directory Structure

models/my-model/
├── model.onnx              # ONNX model file
├── tokenizer.json          # HuggingFace tokenizer
├── config.json             # Model configuration
├── tokenizer_config.json   # Tokenizer configuration
├── special_tokens_map.json # Special tokens mapping
└── chat_template.jinja     # Chat template (optional)