GPU accelerated encrypted tensor computation

EasyFHE

A PyTorch-derived runtime shaped around Fully Homomorphic Encryption: tensor-style Python, native CKKS material generation, CUDA FHE kernels, rotations, rescaling, and bootstrapping.

Get install command Explore runtime Read API docs

Install

Choose a CUDA wheel channel

EasyFHE keeps one Python package name and separates CUDA variants by wheel channel. The v0.1.1 CUDA 12.9 and 13.x wheels target Linux x86_64, Python 3.12, and Ubuntu 22.04/24.04 systems, with native GPU code from A100 (sm_80) through RTX 5090 (sm_120).

CUDA runtime

Python: 3.12
Platform: Linux x86_64
Ubuntu: Ubuntu 22.04, 24.04
Status: alpha

pip command

Runtime

Tensor infrastructure, narrowed for FHE

FHE frontend

`easyfhe.fhe` exposes context specs, ciphertext state, plaintext preparation, arithmetic, rotation, rescale, fused ops, and bootstrapping entry points.

Native material

CKKS contexts, ciphertext material, and rotation keys can be generated through the native EasyFHE path, including fast development modes for iteration.

CUDA kernels

Native kernels cover encoding, NTT, automorphism, key switching, modulus movement, ciphertext arithmetic, and bootstrapping.

Python easyfhe.fhe

Runtime Context, keys, layout

Native ATen FHE ops

Device CUDA and NCCL path

Reference application

ResNet-20 AESPA lives in its own repository

Runtime and packaging stay here. Larger applications can evolve in focused repositories with their own data, scripts, and notes.

easyfhe-resnet20-aespa github.com/jizhuoran/easyfhe-resnet20-aespa

Publish

Static pages, release assets, simple wheel indexes

Wheels are published as GitHub Release assets. The simple wheel indexes under this page let pip resolve the correct release asset without putting large wheel files in Git.

Source: Deploy from a branch
Branch: docs
Folder: /

python -m pip install "easyfhe==0.1.1+cu132" \
  --find-links https://jizhuoran.github.io/EasyFHE/whl/cu132