v1.9.0

Summary

Concrete ML v1.9.0 adds support for the TFHE-rs ciphertext format to tree-based classifiers, enabling interoperability in both developer mode and in client/server deployment. This release also expands the LoRA LLM fine-tuning examples and shows a fully functional encrypted fine-tuning pipeline on GPU.

What's Changed

New Features

• Add client/server for tfhers interoperable models (#1014) (b2a7c89)
• Add tfhe-rs compatibility for tree classifier model inputs /outputs (#997) (0de76b1)
• Add logs + evaluation llama lora (#962) (18547bb)

Fixes

• Mark test as flaky (#1067) (2c5f1ab)
• Release ci push step (#1066) (f3aabf0)
• Cml extensions version (1b74b2a)
• Set executor during compile (#1060) (39f281b)
• Update for release 1.9.0 (#1062) (9f9a1df)

Resources

• Documentation:

  • Concrete ML v1.9 documentation

• Blog Post:

  • Release blog post

• Demo & Examples:
  • Post-processing CML circuit with TFHE-rs rust code
  • LoRA Fine-tuning of LLAMA1B on Math Dataset