fhelium.engine.hybrid_keyswitch

CKKS key-switching helpers.

The central operation in this file is applying a key-switch key to one ciphertext component. In standard CKKS/RNS terminology the hot path is:

1. Split the source component into hybrid-RNS decomposition digits.
2. Convert each composite digit's residues to mixed-radix components.
3. Apply ModUp/basis extension to each digit on the active $Q_{\ell }\cup P$ basis.
4. NTT the extended digit, multiply it by the matching key digit, and accumulate the two switched output components in the $Q_{\ell }\cup P$ basis.
5. INTT the accumulators and ModDown/divide-by-$P$ back to the active $Q_{\ell }$ basis.
6. Add the first correction component to the original ciphertext component.

Hybrid decomposition digits and mixed-radix components are named separately so the code maps directly to the standard CKKS key-switch flow.

PreparedRotationKeySwitch

class View source

PreparedRotationKeySwitch(level: int, batch_shape: tuple[int, ...], ntt_digits_qp: torch.Tensor)

Reusable NTT-domain QP digits for rotations of one component.

Decomposition, ModUp, and forward NTT are independent of the rotation step and are therefore materialized once. Coefficient-domain and mixed-radix intermediates are deliberately not retained: consumers need only the final NTT digits, while tests can construct debug references explicitly without expanding the production artifact representation.

ntt_digits_qp is an integral tensor on the engine device with shape [digit, *batch, limb, ntt_index]. Limb order is the exact active rns_layout.prime_ids(level, include_p=True) order; every residue is Montgomery/lazy. digit is local active order and callers resolve each stable key_digit_index separately before indexing key storage.

Attributes

Name	Type	Default/value
level	int
batch_shape	tuple[int, ...]
ntt_digits_qp	torch.Tensor

HybridKeySwitcher

class View source

HybridKeySwitcher(*, config: CkksConfig, rns_runtime: RnsRuntime, moddown_p_drop_inverses_montgomery_by_level: list[torch.Tensor], direct_digit_consumer: DirectKeySwitchDigitConsumer, galois_generator: int=3)

Execute CKKS hybrid-RNS direct and prepared key switching.

apply_key_switch

method

def apply_key_switch(component: torch.Tensor, key_switch_key: KeySwitchKey, level: int) -> tuple[torch.Tensor, torch.Tensor]: ...

prepare_rotation_digits

method

def prepare_rotation_digits(component: torch.Tensor, level: int) -> PreparedRotationKeySwitch: ...

Materialize only reusable NTT-domain QP digits for rotate-many.