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

v1.6.1

Summary

Minor fixes.

Links

Docker Image: zamafhe/concrete-ml:v1.6.1
Docker Hub: https://hub.docker.com/r/zamafhe/concrete-ml/tags
pip: https://pypi.org/project/concrete-ml/1.6.1
Documentation: https://docs.zama.ai/concrete-ml

v1.6.1

Fix

• Dynamic import of transformers in hybrid model (270efbe)

v1.6.0

Summary

Concrete ML 1.6 includes the following enhancements:

• Latency improvements on large neural networks
• Support for pre-trained tree-based models such as those trained using Federated Learning
• Enhanced collaborative computation
  • Introduction of DataFrame schemas
  • Deployment of logistic regression training

What's Changed

New features

• Enable non-interactive encrypted training for logistic regression (#660) (ec58bca)
• Support pre-trained tree-based models using from_sklearn (5ca282b)
• Add FHE training deployment (#665) (b718629)
• Support approximate rounding to speed up neural networks (9ef890e)
• Allow users to define a schema for dataframe encryption (‘ccd6641’)

Fixes

• Fix fhe-training classes behavior (a88d704)
• Update qgpt2_class.py to fix typo (d376d85)
• Fix post-processing shape mismatches for linear models (#585) (b097022)
• Disable overflow protection in rounding (4db0157)
• Make skorch import fail without error (81de55c)

Improvements

• Replace python release install with setup-python (899b9f1)
• Add support to AvgPool's missing parameters (15a8340)

Resources

• Documentation:

  • Add schema example for encrypted data-frames (#715) (b174509)
  • Add a license FAQ in README (#711) (f937c9f)
  • Update documentation on client / server API (#663) (f864407)
  • Document n_bits for compile torch functions (0306c65)
  • Add examples for the impact of feature scaling on linear models (9252f57)

• Demo & Examples:

  • Add NN-20 and NN-50 deep MLPs for MNIST classification (1b5ce84)

v1.6.0-rc3

Summary

1.6.0 - Release Candidate - 3

v1.6.0-rc0

Summary

1.6.0 - Release Candidate - 0

v1.5.0

Summary

Concrete ML v1.5 introduces several significant enhancements in this release, including a DataFrame API that enables working with encrypted stored data, a new option that speeds up neural networks by 2-3 times, an improved FHE simulation mode to quickly evaluate the impact of the speed-up on neural network accuracy.

What's Changed

Features:

• Add encrypted dataframe API (d2d6250)
• Add an option to allow approximate rounding to speed up NNs (9ef890e)
• Support ONNX conv1d operator (09ad7a6)
• Implement quantized unfold operation (fa3ef88)

Improvements:

• More accurate FHE simulation
• Encrypted aggregation of the outputs of tree ensembles
• Allow different quantization bits for tree model leaves and inputs

Fix

• Import skorch without errors due to bad docstrings (81de55c)
• Add support to AvgPool's missing parameters (15a8340)

Resources

• Documentation:
  • New structure and landing page (85cb962)
  • Add links to credit card approval space in use case examples (df81aca)
  • Improve contributing section (1696799)
  • Document n_bits for compile torch functions (0306c65)
  • Add explanation of encrypted training and federated learning (57dbdff)
  • Add documentation about scaling (9252f57)
  • Add dataframe documentation (#576) (d3bf5ac)

Links

Docker Image: zamafhe/concrete-ml:v1.5.0
Docker Hub: https://hub.docker.com/r/zamafhe/concrete-ml/tags
pip: https://pypi.org/project/concrete-ml/1.5.0
Documentation: https://docs.zama.ai/concrete-ml

v1.5.0-rc1

Summary

Add support for encrypted data-frames and approximate rounding. Fix AvgPool's count_include_pad missing parameter error.

Links

Docker Image: zamafhe/concrete-ml:v1.5.0-rc1
Docker Hub: https://hub.docker.com/r/zamafhe/concrete-ml/tags
pip: https://pypi.org/project/concrete-ml/1.5.0-rc1
Documentation: https://docs.zama.ai/concrete-ml

v1.5.0-rc1

Feature

• Add new approx rounding (9ef890e)
• Encrypted data-frame (d2d6250)
• Support conv1d operator (09ad7a6)
• Implement quantized unfold (fa3ef88)

Fix

• Fix survey link (e5661c1)
• Reinstate apidoc generated tags (2878a07)
• Make skorch import fail without error (81de55c)
• Replace python release install with setup-python (899b9f1)
• Fix concurrency issue in release process (1295ea9)
• Add support to AvgPool's missing parameters (15a8340)
• Update README.md (ba1fdad)

Documentation

• Add dataframe documentation (#576) (d3bf5ac)
• Update main landing pages (cfb862e)
• New structure and landing page (85cb962)
• Update operator list in torch support's documentation section (b617740)
• Add links to credit card approval space in use case examples (df81aca)
• Improve contributing section (1696799)
• Document n_bits for compile torch functions (0306c65)
• Add explanation of encrypted training and federated learning (57dbdff)
• Add documentation about scaling (9252f57)

v1.5.0-rc0

Summary

Add support to torch's Conv1d and Unfold operators

Links

Docker Image: zamafhe/concrete-ml:v1.5.0-rc0
Docker Hub: https://hub.docker.com/r/zamafhe/concrete-ml/tags
pip: https://pypi.org/project/concrete-ml/1.5.0-rc0
Documentation: https://docs.zama.ai/concrete-ml

v1.5.0-rc0

Feature

• Support conv1d operator (09ad7a6)
• Implement quantized unfold (fa3ef88)

Fix

• Make skorch import fail without error (81de55c)
• Replace python release install with setup-python (899b9f1)
• Fix concurrency issue in release process (1295ea9)
• Add support to AvgPool's missing parameters (15a8340)
• Update README.md (ba1fdad)

Documentation

• Update operator list in torch support's documentation section (b617740)
• Add links to credit card approval space in use case examples (df81aca)
• Improve contributing section (1696799)
• Document n_bits for compile torch functions (0306c65)
• Add explanation of encrypted training and federated learning (57dbdff)
• Add documentation about scaling (9252f57)

v1.4.1

Summary

Update Concrete-Python to 2.5.1 and fixes AvgPool's missing parameters.

Links

Docker Image: zamafhe/concrete-ml:v1.4.1
Docker Hub: https://hub.docker.com/r/zamafhe/concrete-ml/tags
pip: https://pypi.org/project/concrete-ml/1.4.1
Documentation: https://docs.zama.ai/concrete-ml

v1.4.1

Fix

• Make skorch import fail without error (5863a4b)
• Replace python release install with setup-python (f41c65c)
• Update README.md (8bef8e5)
• Add support to AvgPool's missing parameters (559d99c)

Documentation

• Document n_bits for compile torch functions (065df89)
• Add explanation of encrypted training and federated learning (8002ed8)
• Add documentation about scaling (1bad5cc)

v1.4.0

Summary

This release adds training a model on encrypted data and introduces latency optimization for the inference of tree-based models such as XGBoost, random forest, and decision trees. This optimization offers 2-3x speed-ups in typical quantization settings and allows even more accurate, high bit-width tree-based models to run with good latency.

Links

Docker Image: zamafhe/concrete-ml:v1.4.0
Docker Hub: https://hub.docker.com/r/zamafhe/concrete-ml/tags
pip: https://pypi.org/project/concrete-ml/1.4.0
Documentation: https://docs.zama.ai/concrete-ml

v1.4.0

Feature

• SGDClassifier training in FHE (0893718)
• Support Expand Equal ONNX op (cf3ce49)
• Add rounding feature on cml trees (064eb82)
• Add multi-output support (fef23a9)
• Allow QuantizedAdd produces_output_graph (0b57c71)
• Encrypted gemm support - 3d inputs - better rounding control - sgd training test (111c7e3)

Fix

• Add --no-warnings flag to linkchecker (1dc547e)
• Fix wrong assumption in ReduceSum operator's axis parameter (1a592d7)
• Mark flaky tests due to issue in simulation (4f67883)
• Update learning rate default value for XGB models (e4984d6)

Documentation

• Improve XGBClassifier notebook (b1906a5)
• Update api doc (d8e9e64)
• Update Apple Silicon install information (4c0c02f)