Concrete ML v1.1.0: Faster inference and a first demo of FHE LLMs

Concrete ML 1.1.0 offers significant advancements in optimizing Fully Homomorphic Encryption (FHE) for neural network models, boosting inference speed by up to 20 times while maintaining user privacy and model owner IP protection. This update includes enhanced support for both neural networks and classical models and provides resources to guide users in optimizing and deploying FHE-based machine learning models. The quantization process reduces the precision of intermediary values to as low as 4 bits, with Zama's FHE libraries efficiently rounding off the least significant bits of encrypted integers, contributing to the speed-up. A new use case illustrates how to integrate encrypted layers into large language models using the Hugging Face transformers library, demonstrating FHE’s compatibility with such models. Deployment of FHE ML models is now streamlined, as evidenced by tutorials and examples such as the Health Diagnosis notebook and a credit scoring demonstration, showcasing the ease of transitioning from development to deployment in hours. Users are encouraged to support and contribute to the FHE space through the Concrete ML GitHub repository, documentation, and the Zama Bounty Program.