v1.7.0

Summary

Concrete ML 1.7 adds functionality to fine-tune LLMs and neural networks on encrypted data using low rank approximation parameter efficient fine-tuning. This allows users to securely outsource large weight matrix computations to remote servers while keeping a small set of private fine-tuned parameters locally. Additionally, this release includes GPU support, providing a 1-2x speed-up for large neural networks on server-grade GPUs like the NVIDIA H100. Concrete ML now also supports Python 3.11 and PyTorch 2.

What's Changed

New features

• Lora fine-tuning in FHE with MLP tutorial and gpt2 use case example (#823) (4d2f2e6)
• Add GPU support (#849) (945aead)
• Add support for Python3.11 (#701) (819dca7)
• Support for embedding layers (#778) (296bc8c)
• Support for encrypted multiplication and division (#690) (a1bd9b8)
• Upgrade PyTorch to 2.3.1, and Brevitas to 0.10 (#788) (c3d7c81)

Improvements

• Relax Python version restrictions for deployment (#853) (040c308)
• Remove Protobuf 2GB limit when checking ONNX (#811) (c8908fa)
• Always using evaluation key compression (#726) (b4e1060)

Fixes

• Fix dtype check in quantizer dequant (77ced60)
• Dynamic import of transformers in hybrid model (829b68b)
• Use correct torch version for Intel Mac (#798) (a8eab89)

Resources

• Documentation:

• Add encrypted LORA fine-tuning documentation (#887) (3f65ec2)
• Add GPU usage guide (#786) (1988936)

• Demo & Examples:
• Encrypted DNA Ancestry : https://huggingface.co/spaces/zama-fhe/encrypted_dna