v1.8.0

Summary

Concrete ML v1.8 improves the functionality for fine-tuning LLMs on encrypted data. This release adds a new, GPU-enabled high-performance cryptographic back-end specialized for this task, improving the speed and usability of hybrid model compilation and testing. Furthermore, the developer experience was improved by using the GPU to speed up LLM hybrid model compilation and testing. Additionally, Concrete ML v1.8 adds support for Python 3.12, aligning with the latest development environment.

What's Changed

New features

• Support Python 3.12 (#847) (0edf038)
• Integrate a faster GPU backend for hybrid model (#868) ([3be0d42](3be0d42d49565df04abff8e332fbad37>
• Speed up hybrid model compilation with GPU (#963)

Fix

• Faster rounding test for weekly builds (#957) ([5b9466c](5b9466ccbe782a9a1bbb0466f6440be>
• Remove custom QAT test (#956) ([1ccaceb](1ccaceb67f420b2fcafd24d2f66e396b53e70cc>
• Add Python 3.12 test in CI pipeline (#950) (96a94c4)
• Propagate GLWE configuration on hybrid MLP unit test (#935) ([664cbd4](664cbd472d044b43>

Resources

Demo & Examples

• Add LORA fine tuning demo for LLAMA 3.2 (#958) ([ef602d9](ef602d9d00a600fcc9edd5a9b3f>