fhelium.engine.ckks_rescale

CKKS rescaling and the bootstrap structural-base transition.

CkksRescaler

class View source

CkksRescaler(*, engine_id: str, device: torch.device, public_level_count: int, rns_runtime: RnsRuntime, montgomery_parameters: MontgomeryParameters, dropped_q_inverses_montgomery_by_level: list[torch.Tensor], assert_engine_ciphertext: Callable[[Ciphertext], None], ciphertext_from_components: Callable[..., Ciphertext], rescale_to_next_output_scale: Callable[..., float])

Execute one-prime CKKS rescale transitions for an engine.

Each transition divides by the leading active Q prime $q_{\mathrm{drop}}$, rounds the quotient, and removes that residue row. The class serves two kinds of level transition:

• rescale_to_next_level and rescale_to_next_level_ operate only on public CKKS levels;
• _rescale_final_public_level_to_structural_base performs the one additional private transition needed immediately before bootstrap modulus raising.

rescale_to_next_level

method

def rescale_to_next_level(ct: Ciphertext, rounding: Literal['nearest', 'floor']='nearest') -> Ciphertext: ...

Drop the current scale prime at an ordinary public CKKS level.

Rescaling divides by the leading active Q prime and removes that RNS row. For each component,

$$c^{\prime }=\mathrm{Round}\left(\frac{c}{q_{\mathrm{drop}}}\right)(\mathrm{mod}{B}_{\ell +1}),\mathrm{\Delta }(c^{\prime })=\frac{\mathrm{\Delta }(c)}{q_{\mathrm{drop}}}.$$

Q input produces $Q_{\ell +1}$; QP input produces $Q_{\ell +1}P$ and retains all P rows.

Public rescaling stops before the one-prime structural basis. This method returns new ciphertext storage; rescale_to_next_level_ mutates its ciphertext argument.

Parameters

• ct: Coefficient-domain, standard-residue ciphertext over the engine's full active Q or QP layout.
• rounding: "nearest" for nearest-integer division or "floor" for the least-nonnegative-residue quotient.

Returns

A new coefficient-domain standard ciphertext at ct.level + 1 with unchanged component count, batch shape, Q/QP basis, engine integral dtype/device, exact prime_ids=ct.prime_ids[1:], and canonical residues in $[0,q_i)$. ct is unchanged and output storage is independent.

Raises

• MaximumLevelError: If no further public rescale level exists.
• InvalidScaleError: If the output scale is invalid.
• ValueError: If the ciphertext state is incompatible with rescaling.

rescale_to_next_level_

method

def rescale_to_next_level_(ct: Ciphertext, rounding: Literal['nearest', 'floor']='nearest') -> Ciphertext: ...

In-place form of rescale_to_next_level.

Native kernels update the remaining RNS rows through views into ct.data. This method then narrows the dense tensor by one row and updates level, scale, and prime_ids. The returned object is ct itself; aliases must therefore be treated as mutated.

The quotient and actual-scale equations are identical to rescale_to_next_level. Aliases observe updated surviving rows, narrowed ct.data, level, scale, and exact prime_ids. The narrowed tensor remains a view of the original allocation.