v0.2.2

Highlights

• fix GPU rotate, ModDown, parameter, and Rescale bugs in the OpenFHE GPU path
• restore passing large-ring BSGS / EncryptedLinear and DenseMNIST full-GPU inference
• fix GroupNorm conversion API mismatch and resnet benchmark harness issues
• refresh benchmark documentation and benchmark reference page

Benchmark snapshot

• mlp: 84.16 ms encrypted, MAE 0.075594
• cnn: 2820.15 ms encrypted, MAE 0.048294
• resnet: 11444.55 ms encrypted, MAE 0.098026
• transformer: 54.27 ms encrypted, MAE 0.087978

Commits in this release

• fix(gpu): correct OpenFHE GPU rotate and modulus conversion internals
• fix(backend): align CuKKS bridge and tensor flow with GPU semantics
• fix(groupnorm): accept converter kwargs and add regression coverage
• fix(bench): make benchmark models and context setup runnable
• docs: refresh README benchmark results
• docs: add current benchmark results reference
• release: bump package versions to v0.2.2